Deep Learning-Based Algorithms in Screening of Diabetic Retinopathy: A Systematic Review of Diagnostic Performance

Hybrid deep learning framework for diabetic retinopathy classification with optimized attention AlexNet

Diabetic retinopathy (DR) is a chronic condition associated with diabetes that can lead to vision impairment and, if not addressed, may progress to irreversible blindness. Consequently, it is essential to detect pathological changes in the retina to assess DR severity accurately. Manual examination of retinal disorders is often complex, time consuming, and susceptible to errors due to fine retinal disorder. In recent years, Deep Learning (DL) based optimizations have shown significant promises in improving DR recognition and classification. At last, the advanced classification method using metaheuristic optimization for grading severity in fundus images. This work presents an automated DR classification using metaheuristic optimization based advanced DL model. There are four stages are involved in the suggested DR classification. At first, the pre-processing stage is performed green channel conversion, CLAHE and Gaussian filtering (GF). Then, the fundus lesions are segmented by the Fuzzy Possibilistic C Ordered Means (FPCOM). Finally, the lesions are classified by Attention AlexNet based Improved Nutcracker Optimizer (At-AlexNet-ImNO). The ImNO optimizes the At-AlexNet's weights and hyperparameters and boosts the classification performance. The experimentation is performed on two benchmark datasets like APTOS-2019 Blindness-Detection and EyePacs. Accuracy, precision and recall values achieved are 99.23 %, 98 % and 98.2 % on APTOS-2019 and accuracy, precision and recall values achieved are 99.43 %, 98.2 % and 98.65 % on EyePacs respectively.

A Deep Learning Segmentation Model for Detection of Active Proliferative Diabetic Retinopathy

INTRODUCTION

Existing deep learning (DL) algorithms lack the capability to accurately identify patients in immediate need of treatment for proliferative diabetic retinopathy (PDR). We aimed to develop a DL segmentation model to detect active PDR in six-field retinal images by the annotation of new retinal vessels and preretinal hemorrhages.

METHODS

We identified six-field retinal images classified at level 4 of the International Clinical Diabetic Retinopathy Disease Severity Scale collected at the Island of Funen from 2009 to 2019 as part of the Danish screening program for diabetic retinopathy (DR). A certified grader (grader 1) manually dichotomized the images into active or inactive PDR, and the images were then reassessed by two independent certified graders. In cases of disagreement, the final classification decision was made in collaboration between grader 1 and one of the secondary graders. Overall, 637 images were classified as active PDR. We then applied our pre-established DL segmentation model to annotate nine lesion types before training the algorithm. The segmentations of new vessels and preretinal hemorrhages were corrected for any inaccuracies before training the DL algorithm. After the classification and pre-segmentation phases the images were divided into training (70%), validation (10%), and testing (20%) datasets. We added 301 images with inactive PDR to the testing dataset.

RESULTS

We included 637 images of active PDR and 301 images of inactive PDR from 199 individuals. The training dataset had 1381 new vessel and preretinal hemorrhage lesions, while the validation dataset had 123 lesions and the testing dataset 374 lesions. The DL system demonstrated a sensitivity of 90% and a specificity of 70% for annotation-assisted classification of active PDR. The negative predictive value was 94%, while the positive predictive value was 57%.

CONCLUSIONS

Our DL segmentation model achieved excellent sensitivity and acceptable specificity in distinguishing active from inactive PDR.

Diagnostic utility of artificial intelligence software through non-mydriatic digital retinography in the screening of diabetic retinopathy: an overview of reviews

OBJECTIVE

To evaluate the capability of artificial intelligence (AI) in screening for diabetic retinopathy (DR) utilizing digital retinography captured by non-mydriatic (NM) ≥45° cameras, focusing on diagnosis accuracy, effectiveness, and clinical safety.

METHODS

We performed an overview of systematic reviews (SRs) up to May 2023 in Medline, Embase, CINAHL, and Web of Science. We used AMSTAR-2 tool to assess the reliability of each SR. We reported meta-analysis estimates or ranges of diagnostic performance figures.

RESULTS

Out of 1336 records, ten SRs were selected, most deemed low or critically low quality. Eight primary studies were included in at least five of the ten SRs and 125 in less than five SRs. No SR reported efficacy, effectiveness, or safety outcomes. The sensitivity and specificity for referable DR were 68-100% and 20-100%, respectively, with an AUROC range of 88 to 99%. For detecting DR at any stage, sensitivity was 79-100%, and specificity was 50-100%, with an AUROC range of 93 to 98%.

CONCLUSIONS

AI demonstrates strong diagnostic potential for DR screening using NM cameras, with adequate sensitivity but variable specificity. While AI is increasingly integrated into routine practice, this overview highlights significant heterogeneity in AI models and the cameras used. Additionally, our study enlightens the low quality of existing systematic reviews and the significant challenge of integrating the rapidly growing volume of emerging evidence in this field. Policymakers should carefully evaluate AI tools in specific contexts, and future research must generate updated high-quality evidence to optimize their application and improve patient outcomes.

A deep learning-assisted automatic measurement of tear meniscus height on ocular surface images and its application in myopia control

Purpose

Modalities for myopia control, such as orthokeratology, repeated low-intensity red light (RLRL) treatment, and low-concentration atropine, have become popular topics. However, the effects of these three modalities on ocular surface health remain unclear. The tear meniscus height (TMH), a crucial criterion for evaluating ocular surface health and diagnosing dry eye, is conventionally measured via manual demarcation of ocular surface images, which is inefficient and involves subjective judgment. Therefore, this study sought to establish a deep learning model for automatic TMH measurement on ocular surface images to improve the efficiency and accuracy of the initial screening of dry eye associated with myopia control modalities.

Methods

To establish a model, 1,200 ocular surface images captured with an OCULUS Keratograph 5M were collected. The tear meniscus area on the image was initially marked by one experienced ophthalmologist and verified by the other. The whole image dataset was divided into a training set (70%), a validation set (20%), a test set (10%), and an external validation set (100 ocular surface images) for model construction. The deep learning model was applied to ocular surface imaging data from previous clinical trials using orthokeratology, RLRL therapy, and 0.01% atropine for myopia control. TMHs at follow-ups were automatically measured by the deep learning model.

Results

Two hundred training iterations were performed to establish the model. At the 124th iteration, the IoU of the validation set peaked at 0.913, and the parameters of the model were saved for the testing process. The model IoU was 0.928 during testing. The AUC of the ROC curve was 0.935, and the R2 of the linear regression analysis was 0.92. The good performance and comprehensive validation of the model warrants its application to automatic TMH measurement in clinical trials of myopia control. There were no significant changes in the TMH during the follow-up period after treatment with orthokeratology, RLRL, or 0.01% atropine.

Conclusion

A deep learning model was established for automatic measurement of the TMH on Keratograph 5M-captured ocular surface images. This model demonstrated high accuracy, great consistency with manual measurements, and applicability to the initial screening of dry eye associated with myopia control modalities.

A lightweight multi-deep learning framework for accurate diabetic retinopathy detection and multi-level severity identification

Accurate and timely detection of diabetic retinopathy (DR) is crucial for managing its progression and improving patient outcomes. However, developing algorithms to analyze complex fundus images continues to be a major challenge. This work presents a lightweight deep-learning network developed for DR detection. The proposed framework consists of two stages. In the first step, the developed model is used to assess the presence of DR [i.e., healthy (no DR) or diseased (DR)]. The next step involves the use of transfer learning for further subclassification of DR severity (i.e., mild, moderate, severe DR, and proliferative DR). The designed model is reused for transfer learning, as correlated images facilitate further classification of DR severity. The online dataset is used to validate the proposed framework, and results show that the proposed model is lightweight and has comparatively low learnable parameters compared to others. The proposed two-stage framework enhances the classification performance, achieving a 99.06% classification rate for DR detection and an accuracy of 90.75% for DR severity identification for APTOS 2019 dataset.

RFMiD: Retinal Image Analysis for multi-Disease Detection challenge

In the last decades, many publicly available large fundus image datasets have been collected for diabetic retinopathy, glaucoma, and age-related macular degeneration, and a few other frequent pathologies. These publicly available datasets were used to develop a computer-aided disease diagnosis system by training deep learning models to detect these frequent pathologies. One challenge limiting the adoption of a such system by the ophthalmologist is, computer-aided disease diagnosis system ignores sight-threatening rare pathologies such as central retinal artery occlusion or anterior ischemic optic neuropathy and others that ophthalmologists currently detect. Aiming to advance the state-of-the-art in automatic ocular disease classification of frequent diseases along with the rare pathologies, a grand challenge on "Retinal Image Analysis for multi-Disease Detection" was organized in conjunction with the IEEE International Symposium on Biomedical Imaging (ISBI - 2021). This paper, reports the challenge organization, dataset, top-performing participants solutions, evaluation measures, and results based on a new "Retinal Fundus Multi-disease Image Dataset" (RFMiD). There were two principal sub-challenges: disease screening (i.e. presence versus absence of pathology - a binary classification problem) and disease/pathology classification (a 28-class multi-label classification problem). It received a positive response from the scientific community with 74 submissions by individuals/teams that effectively entered in this challenge. The top-performing methodologies utilized a blend of data-preprocessing, data augmentation, pre-trained model, and model ensembling. This multi-disease (frequent and rare pathologies) detection will enable the development of generalizable models for screening the retina, unlike the previous efforts that focused on the detection of specific diseases.

Screening for diabetic retinopathy with artificial intelligence: a real world evaluation

AIM

Periodic screening for diabetic retinopathy (DR) is effective for preventing blindness. Artificial intelligence (AI) systems could be useful for increasing the screening of DR in diabetic patients. The aim of this study was to compare the performance of the DAIRET system in detecting DR to that of ophthalmologists in a real-world setting.

METHODS

Fundus photography was performed with a nonmydriatic camera in 958 consecutive patients older than 18 years who were affected by diabetes and who were enrolled in the DR screening in the Diabetes and Endocrinology Unit and in the Eye Unit of ULSS8 Berica (Italy) between June 2022 and June 2023. All retinal images were evaluated by DAIRET, which is a machine learning algorithm based on AI. In addition, all the images obtained were analysed by an ophthalmologist who graded the images. The results obtained by DAIRET were compared with those obtained by the ophthalmologist.

RESULTS

We included 958 patients, but only 867 (90.5%) patients had retinal images sufficient for evaluation by a human grader. The sensitivity for detecting cases of moderate DR and above was 1 (100%), and the sensitivity for detecting cases of mild DR was 0.84 ± 0.03. The specificity of detecting the absence of DR was lower (0.59 ± 0.04) because of the high number of false-positives.

CONCLUSION

DAIRET showed an optimal sensitivity in detecting all cases of referable DR (moderate DR or above) compared with that of a human grader. On the other hand, the specificity of DAIRET was low because of the high number of false-positives, which limits its cost-effectiveness.

Interpretation of Clinical Retinal Images Using an Artificial Intelligence Chatbot

Purpose

To assess the performance of Chat Generative Pre-Trained Transformer-4 in providing accurate diagnoses to retina teaching cases from OCTCases.

Design

Cross-sectional study.

Subjects

Retina teaching cases from OCTCases.

Methods

We prompted a custom chatbot with 69 retina cases containing multimodal ophthalmic images, asking it to provide the most likely diagnosis. In a sensitivity analysis, we inputted increasing amounts of clinical information pertaining to each case until the chatbot achieved a correct diagnosis. We performed multivariable logistic regressions on Stata v17.0 (StataCorp LLC) to investigate associations between the amount of text-based information inputted per prompt and the odds of the chatbot achieving a correct diagnosis, adjusting for the laterality of cases, number of ophthalmic images inputted, and imaging modalities.

Main Outcome Measures

Our primary outcome was the proportion of cases for which the chatbot was able to provide a correct diagnosis. Our secondary outcome was the chatbot's performance in relation to the amount of text-based information accompanying ophthalmic images.

Results

Across 69 retina cases collectively containing 139 ophthalmic images, the chatbot was able to provide a definitive, correct diagnosis for 35 (50.7%) cases. The chatbot needed variable amounts of clinical information to achieve a correct diagnosis, where the entire patient description as presented by OCTCases was required for a majority of correctly diagnosed cases (23 of 35 cases, 65.7%). Relative to when the chatbot was only prompted with a patient's age and sex, the chatbot achieved a higher odds of a correct diagnosis when prompted with an entire patient description (odds ratio = 10.1, 95% confidence interval = 3.3-30.3, P < 0.01). Despite providing an incorrect diagnosis for 34 (49.3%) cases, the chatbot listed the correct diagnosis within its differential diagnosis for 7 (20.6%) of these incorrectly answered cases.

Conclusions

This custom chatbot was able to accurately diagnose approximately half of the retina cases requiring multimodal input, albeit relying heavily on text-based contextual information that accompanied ophthalmic images. The diagnostic ability of the chatbot in interpretation of multimodal imaging without text-based information is currently limited. The appropriate use of the chatbot in this setting is of utmost importance, given bioethical concerns.

Financial Disclosures

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Strong versus Weak Data Labeling for Artificial Intelligence Algorithms in the Measurement of Geographic Atrophy

Purpose

To gain an understanding of data labeling requirements to train deep learning models for measurement of geographic atrophy (GA) with fundus autofluorescence (FAF) images.

Design

Evaluation of artificial intelligence (AI) algorithms.

Subjects

The Age-Related Eye Disease Study 2 (AREDS2) images were used for training and cross-validation, and GA clinical trial images were used for testing.

Methods

Training data consisted of 2 sets of FAF images; 1 with area measurements only and no indication of GA location (Weakly labeled) and the second with GA segmentation masks (Strongly labeled).

Main Outcome Measures

Bland-Altman plots and scatter plots were used to compare GA area measurement between ground truth and AI measurements. The Dice coefficient was used to compare accuracy of segmentation of the Strong model.

Results

In the cross-validation AREDS2 data set (n = 601), the mean (standard deviation [SD]) area of GA measured by human grader, Weakly labeled AI model, and Strongly labeled AI model was 6.65 (6.3) mm2, 6.83 (6.29) mm2, and 6.58 (6.24) mm2, respectively. The mean difference between ground truth and AI was 0.18 mm2 (95% confidence interval, [CI], -7.57 to 7.92) for the Weakly labeled model and -0.07 mm2 (95% CI, -1.61 to 1.47) for the Strongly labeled model. With GlaxoSmithKline testing data (n = 156), the mean (SD) GA area was 9.79 (5.6) mm2, 8.82 (4.61) mm2, and 9.55 (5.66) mm2 for human grader, Strongly labeled AI model, and Weakly labeled AI model, respectively. The mean difference between ground truth and AI for the 2 models was -0.97 mm2 (95% CI, -4.36 to 2.41) and -0.24 mm2 (95% CI, -4.98 to 4.49), respectively. The Dice coefficient was 0.99 for intergrader agreement, 0.89 for the cross-validation data, and 0.92 for the testing data.

Conclusions

Deep learning models can achieve reasonable accuracy even with Weakly labeled data. Training methods that integrate large volumes of Weakly labeled images with small number of Strongly labeled images offer a promising solution to overcome the burden of cost and time for data labeling.

Financial Disclosures

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Enhancing deep learning pre-trained networks on diabetic retinopathy fundus photographs with SLIC-G

Diabetic retinopathy disease contains lesions (e.g., exudates, hemorrhages, and microaneurysms) that are minute to the naked eye. Determining the lesions at pixel level poses a challenge as each pixel does not reflect any semantic entities. Furthermore, the computational cost of inspecting each pixel is expensive because the number of pixels is high even at low resolution. In this work, we propose a hybrid image processing method. Simple Linear Iterative Clustering with Gaussian Filter (SLIC-G) for the purpose of overcoming pixel constraints. The SLIC-G image processing method is divided into two stages: (1) simple linear iterative clustering superpixel segmentation and (2) Gaussian smoothing operation. In such a way, a large number of new transformed datasets are generated and then used for model training. Finally, two performance evaluation metrics that are suitable for imbalanced diabetic retinopathy datasets were used to validate the effectiveness of the proposed SLIC-G. The results indicate that, in comparison to prior published works' results, the proposed SLIC-G shows better performance on image classification of class imbalanced diabetic retinopathy datasets. This research reveals the importance of image processing and how it influences the performance of deep learning networks. The proposed SLIC-G enhances pre-trained network performance by eliminating the local redundancy of an image, which preserves local structures, but avoids over-segmented, noisy clips. It closes the research gap by introducing the use of superpixel segmentation and Gaussian smoothing operation as image processing methods in diabetic retinopathy-related tasks.

Evaluation of an AI algorithm trained on an ethnically diverse dataset to screen a previously unseen population for diabetic retinopathy

PURPOSE

This study aimed to determine the generalizability of an artificial intelligence (AI) algorithm trained on an ethnically diverse dataset to screen for referable diabetic retinopathy (RDR) in the Armenian population unseen during AI development.

METHODS

This study comprised 550 patients with diabetes mellitus visiting the polyclinics of Armenia over 10 months requiring diabetic retinopathy (DR) screening. The Medios AI-DR algorithm was developed using a robust, diverse, ethnically balanced dataset with no inherent bias and deployed offline on a smartphone-based fundus camera. The algorithm here analyzed the retinal images captured using the target device for the presence of RDR (i.e., moderate non-proliferative diabetic retinopathy (NPDR) and/or clinically significant diabetic macular edema (CSDME) or more severe disease) and sight-threatening DR (STDR, i.e., severe NPDR and/or CSDME or more severe disease). The results compared the AI output to a consensus or majority image grading of three expert graders according to the International Clinical Diabetic Retinopathy severity scale.

RESULTS

On 478 subjects included in the analysis, the algorithm achieved a high classification sensitivity of 95.30% (95% CI: 91.9%-98.7%) and a specificity of 83.89% (95% CI: 79.9%-87.9%) for the detection of RDR. The sensitivity for STDR detection was 100%.

CONCLUSION

The study proved that Medios AI-DR algorithm yields good accuracy in screening for RDR in the Armenian population. In our literature search, this is the only smartphone-based, offline AI model validated in different populations.

Construction of Risk Prediction Model of Type 2 Diabetic Kidney Disease Based on Deep Learning

BACKGRUOUND

This study aimed to develop a diabetic kidney disease (DKD) prediction model using long short term memory (LSTM) neural network and evaluate its performance using accuracy, precision, recall, and area under the curve (AUC) of the receiver operating characteristic (ROC) curve.

METHODS

The study identified DKD risk factors through literature review and physician focus group, and collected 7 years of data from 6,040 type 2 diabetes mellitus patients based on the risk factors. Pytorch was used to build the LSTM neural network, with 70% of the data used for training and the other 30% for testing. Three models were established to examine the impact of glycosylated hemoglobin (HbA1c), systolic blood pressure (SBP), and pulse pressure (PP) variabilities on the model's performance.

RESULTS

The developed model achieved an accuracy of 83% and an AUC of 0.83. When the risk factor of HbA1c variability, SBP variability, or PP variability was removed one by one, the accuracy of each model was significantly lower than that of the optimal model, with an accuracy of 78% (P<0.001), 79% (P<0.001), and 81% (P<0.001), respectively. The AUC of ROC was also significantly lower for each model, with values of 0.72 (P<0.001), 0.75 (P<0.001), and 0.77 (P<0.05).

CONCLUSION

The developed DKD risk predictive model using LSTM neural networks demonstrated high accuracy and AUC value. When HbA1c, SBP, and PP variabilities were added to the model as featured characteristics, the model's performance was greatly improved.

A multimodal approach using fundus images and text meta-data in a machine learning classifier with embeddings to predict years with self-reported diabetes - An exploratory analysis

AIMS

Machine learning models can use image and text data to predict the number of years since diabetes diagnosis; such model can be applied to new patients to predict, approximately, how long the new patient may have lived with diabetes unknowingly. We aimed to develop a model to predict self-reported diabetes duration.

METHODS

We used the Brazilian Multilabel Ophthalmological Dataset. Unit of analysis was the fundus image and its meta-data, regardless of the patient. We included people 40 + years and fundus images without diabetic retinopathy. Fundus images and meta-data (sex, age, comorbidities and taking insulin) were passed to the MedCLIP model to extract the embedding representation. The embedding representation was passed to an Extra Tree Classifier to predict: 0-4, 5-9, 10-14 and 15 + years with self-reported diabetes.

RESULTS

There were 988 images from 563 people (mean age = 67 years; 64 % were women). Overall, the F1 score was 57 %. The group 15 + years of self-reported diabetes had the highest precision (64 %) and F1 score (63 %), while the highest recall (69 %) was observed in the group 0-4 years. The proportion of correctly classified observations was 55 % for the group 0-4 years, 51 % for 5-9 years, 58 % for 10-14 years, and 64 % for 15 + years with self-reported diabetes.

CONCLUSIONS

The machine learning model had acceptable accuracy and F1 score, and correctly classified more than half of the patients according to diabetes duration. Using large foundational models to extract image and text embeddings seems a feasible and efficient approach to predict years living with self-reported diabetes.

Feasibility and acceptance of artificial intelligence-based diabetic retinopathy screening in Rwanda

BACKGROUND

Evidence on the practical application of artificial intelligence (AI)-based diabetic retinopathy (DR) screening is needed.

METHODS

Consented participants were screened for DR using retinal imaging with AI interpretation from March 2021 to June 2021 at four diabetes clinics in Rwanda. Additionally, images were graded by a UK National Health System-certified retinal image grader. DR grades based on the International Classification of Diabetic Retinopathy with a grade of 2.0 or higher were considered referable. The AI system was designed to detect optic nerve and macular anomalies outside of DR. A vertical cup to disc ratio of 0.7 and higher and/or macular anomalies recognised at a cut-off of 60% and higher were also considered referable by AI.

RESULTS

Among 827 participants (59.6% women (n=493)) screened by AI, 33.2% (n=275) were referred for follow-up. Satisfaction with AI screening was high (99.5%, n=823), and 63.7% of participants (n=527) preferred AI over human grading. Compared with human grading, the sensitivity of the AI for referable DR was 92% (95% CI 0.863%, 0.968%), with a specificity of 85% (95% CI 0.751%, 0.882%). Of the participants referred by AI: 88 (32.0%) were for DR only, 109 (39.6%) for DR and an anomaly, 65 (23.6%) for an anomaly only and 13 (4.73%) for other reasons. Adherence to referrals was highest for those referred for DR at 53.4%.

CONCLUSION

DR screening using AI led to accurate referrals from diabetes clinics in Rwanda and high rates of participant satisfaction, suggesting AI screening for DR is practical and acceptable.

Vision transformer with masked autoencoders for referable diabetic retinopathy classification based on large-size retina image

Computer-aided diagnosis systems based on deep learning algorithms have shown potential applications in rapid diagnosis of diabetic retinopathy (DR). Due to the superior performance of Transformer over convolutional neural networks (CNN) on natural images, we attempted to develop a new model to classify referable DR based on a limited number of large-size retinal images by using Transformer. Vision Transformer (ViT) with Masked Autoencoders (MAE) was applied in this study to improve the classification performance of referable DR. We collected over 100,000 publicly fundus retinal images larger than 224×224, and then pre-trained ViT on these retinal images using MAE. The pre-trained ViT was applied to classify referable DR, the performance was also compared with that of ViT pre-trained using ImageNet. The improvement in model classification performance by pre-training with over 100,000 retinal images using MAE is superior to that pre-trained with ImageNet. The accuracy, area under curve (AUC), highest sensitivity and highest specificity of the present model are 93.42%, 0.9853, 0.973 and 0.9539, respectively. This study shows that MAE can provide more flexibility to the input image and substantially reduce the number of images required. Meanwhile, the pretraining dataset scale in this study is much smaller than ImageNet, and the pre-trained weights from ImageNet are not required also.

Present and future screening programs for diabetic retinopathy: a narrative review

Diabetes is a prevalent global concern, with an estimated 12% of the global adult population affected by 2045. Diabetic retinopathy (DR), a sight-threatening complication, has spurred diverse screening approaches worldwide due to advances in DR knowledge, rapid technological developments in retinal imaging and variations in healthcare resources.Many high income countries have fully implemented or are on the verge of completing a national Diabetic Eye Screening Programme (DESP). Although there have been some improvements in DR screening in Africa, Asia, and American countries further progress is needed. In low-income countries, only one out of 29, partially implemented a DESP, while 21 out of 50 lower-middle-income countries have started the DR policy cycle. Among upper-middle-income countries, a third of 59 nations have advanced in DR agenda-setting, with five having a comprehensive national DESP and 11 in the early stages of implementation.Many nations use 2-4 fields fundus images, proven effective with 80-98% sensitivity and 86-100% specificity compared to the traditional seven-field evaluation for DR. A cell phone based screening with a hand held retinal camera presents a potential low-cost alternative as imaging device. While this method in low-resource settings may not entirely match the sensitivity and specificity of seven-field stereoscopic photography, positive outcomes are observed.Individualized DR screening intervals are the standard in many high-resource nations. In countries that lacks a national DESP and resources, screening are more sporadic, i.e. screening intervals are not evidence-based and often less frequently, which can lead to late recognition of treatment required DR.The rising global prevalence of DR poses an economic challenge to nationwide screening programs AI-algorithms have showed high sensitivity and specificity for detection of DR and could provide a promising solution for the future screening burden.In summary, this narrative review enlightens on the epidemiology of DR and the necessity for effective DR screening programs. Worldwide evolution in existing approaches for DR screening has showed promising results but has also revealed limitations. Technological advancements, such as handheld imaging devices, tele ophthalmology and artificial intelligence enhance cost-effectiveness, but also the accessibility of DR screening in countries with low resources or where distance to or a shortage of ophthalmologists exists.

Development of LuxIA, a Cloud-Based AI Diabetic Retinopathy Screening Tool Using a Single Color Fundus Image

Purpose

Diabetic retinopathy (DR) is the leading cause of vision impairment in working-age adults. Automated screening can increase DR detection at early stages at relatively low costs. We developed and evaluated a cloud-based screening tool that uses artificial intelligence (AI), the LuxIA algorithm, to detect DR from a single fundus image.

Methods

Color fundus images that were previously graded by expert readers were collected from the Canarian Health Service (Retisalud) and used to train LuxIA, a deep-learning-based algorithm for the detection of more than mild DR. The algorithm was deployed in the Discovery cloud platform to evaluate each test set. Sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve were computed using a bootstrapping method to evaluate the algorithm performance and compared through different publicly available datasets. A usability test was performed to assess the integration into a clinical tool.

Results

Three separate datasets, Messidor-2, APTOS, and a holdout set from Retisalud were evaluated. Mean sensitivity and specificity with 95% confidence intervals (CIs) reached for these three datasets were 0.901 (0.901-0.902) and 0.955 (0.955-0.956), 0.995 (0.995-0.995) and 0.821 (0.821-0.823), and 0.911 (0.907-0.912) and 0.880 (0.879-0.880), respectively. The usability test confirmed the successful integration of LuxIA into Discovery.

Conclusions

Clinical data were used to train the deep-learning-based algorithm LuxIA to an expert-level performance. The whole process (image uploading and analysis) was integrated into the cloud-based platform Discovery, allowing more patients to have access to expert-level screening tools.

Translational Relevance

Using the cloud-based LuxIA tool as part of a screening program may give diabetic patients greater access to specialist-level decisions, without the need for consultation.

CHILDSTAR: CHIldren Living With Diabetes See and Thrive with AI Review

Background

Artificial intelligence (AI) appears capable of detecting diabetic retinopathy (DR) with a high degree of accuracy in adults; however, there are few studies in children and young adults.

Methods

Children and young adults (3-26 years) with type 1 diabetes mellitus (T1DM) or type 2 diabetes mellitus (T2DM) were screened at the Dhaka BIRDEM-2 hospital, Bangladesh. All gradable fundus images were uploaded to Cybersight AI for interpretation. Two main outcomes were considered at a patient level: 1) Any DR, defined as mild non-proliferative diabetic retinopathy (NPDR or more severe; and 2) Referable DR, defined as moderate NPDR or more severe. Diagnostic test performance comparing Orbis International's Cybersight AI with the reference standard, a fully qualified optometrist certified in DR grading, was assessed using the Matthews correlation coefficient (MCC), area under the receiver operating characteristic curve (AUC-ROC), area under the precision-recall curve (AUC-PR), sensitivity, specificity, positive and negative predictive values.

Results

Among 1274 participants (53.1% female, mean age 16.7 years), 19.4% (n = 247) had any DR according to AI. For referable DR, 2.35% (n = 30) were detected by AI. The sensitivity and specificity of AI for any DR were 75.5% (CI 69.7-81.3%) and 91.8% (CI 90.2-93.5%) respectively, and for referable DR, these values were 84.2% (CI 67.8-100%) and 98.9% (CI 98.3%-99.5%). The MCC, AUC-ROC and the AUC-PR for referable DR were 63.4, 91.2 and 76.2% respectively. AI was most successful in accurately classifying younger children with shorter duration of diabetes.

Conclusions

Cybersight AI accurately detected any DR and referable DR among children and young adults, despite its algorithms having been trained on adults. The observed high specificity is particularly important to avoid over-referral in low-resource settings. AI may be an effective tool to reduce demands on scarce physician resources for the care of children with diabetes in low-resource settings.

Human-Centered Design to Address Biases in Artificial Intelligence

The potential of artificial intelligence (AI) to reduce health care disparities and inequities is recognized, but it can also exacerbate these issues if not implemented in an equitable manner. This perspective identifies potential biases in each stage of the AI life cycle, including data collection, annotation, machine learning model development, evaluation, deployment, operationalization, monitoring, and feedback integration. To mitigate these biases, we suggest involving a diverse group of stakeholders, using human-centered AI principles. Human-centered AI can help ensure that AI systems are designed and used in a way that benefits patients and society, which can reduce health disparities and inequities. By recognizing and addressing biases at each stage of the AI life cycle, AI can achieve its potential in health care.

Development of a Computer System for Automatically Generating a Laser Photocoagulation Plan to Improve the Retinal Coagulation Quality in the Treatment of Diabetic Retinopathy

In this article, the development of a computer system for high-tech medical uses in ophthalmology is proposed. An overview of the main methods and algorithms that formed the basis of the coagulation plan planning system is presented. The system provides the formation of a more effective plan for laser coagulation in comparison with the use of existing coagulation techniques. An analysis of monopulse- and pattern-based laser coagulation techniques in the treatment of diabetic retinopathy has shown that modern treatment methods do not provide the required efficacy of medical laser coagulation procedures, as the laser energy is nonuniformly distributed across the pigment epithelium and may exert an excessive effect on parts of the retina and anatomical elements. The analysis has shown that the efficacy of retinal laser coagulation for the treatment of diabetic retinopathy is determined by the relative position of coagulates and parameters of laser exposure. In the course of the development of the computer system proposed herein, main stages of processing diagnostic data were identified. They are as follows: the allocation of the laser exposure zone, the evaluation of laser pulse parameters that would be safe for the fundus, mapping a coagulation plan in the laser exposure zone, followed by the analysis of the generated plan for predicting the therapeutic effect. In the course of the study, it was found that the developed algorithms for placing coagulates in the area of laser exposure provide a more uniform distribution of laser energy across the pigment epithelium when compared to monopulse- and pattern-based laser coagulation techniques.

Economics of Artificial Intelligence in Healthcare: Diagnosis vs. Treatment

Motivation: The price of medical treatment continues to rise due to (i) an increasing population; (ii) an aging human growth; (iii) disease prevalence; (iv) a rise in the frequency of patients that utilize health care services; and (v) increase in the price. Objective: Artificial Intelligence (AI) is already well-known for its superiority in various healthcare applications, including the segmentation of lesions in images, speech recognition, smartphone personal assistants, navigation, ride-sharing apps, and many more. Our study is based on two hypotheses: (i) AI offers more economic solutions compared to conventional methods; (ii) AI treatment offers stronger economics compared to AI diagnosis. This novel study aims to evaluate AI technology in the context of healthcare costs, namely in the areas of diagnosis and treatment, and then compare it to the traditional or non-AI-based approaches. Methodology: PRISMA was used to select the best 200 studies for AI in healthcare with a primary focus on cost reduction, especially towards diagnosis and treatment. We defined the diagnosis and treatment architectures, investigated their characteristics, and categorized the roles that AI plays in the diagnostic and therapeutic paradigms. We experimented with various combinations of different assumptions by integrating AI and then comparing it against conventional costs. Lastly, we dwell on three powerful future concepts of AI, namely, pruning, bias, explainability, and regulatory approvals of AI systems. Conclusions: The model shows tremendous cost savings using AI tools in diagnosis and treatment. The economics of AI can be improved by incorporating pruning, reduction in AI bias, explainability, and regulatory approvals.

Digital innovations for retinal care in diabetic retinopathy

AIM

The purpose of this review is to examine the applications of novel digital technology domains for the screening and management of patients with diabetic retinopathy (DR).

METHODS

A PubMed engine search was performed, using the terms "Telemedicine", "Digital health", "Telehealth", "Telescreening", "Artificial intelligence", "Deep learning", "Smartphone", "Triage", "Screening", "Home-based", "Monitoring", "Ophthalmology", "Diabetes", "Diabetic Retinopathy", "Retinal imaging". Full-text English language studies from January 1, 2010, to February 1, 2022, and reference lists were considered for the conceptual framework of this review.

RESULTS

Diabetes mellitus and its eye complications, including DR, are particularly well suited to digital technologies, providing an ideal model for telehealth initiatives and real-world applications. The current development in the adoption of telemedicine, artificial intelligence and remote monitoring as an alternative to or in addition to traditional forms of care will be discussed.

CONCLUSIONS

Advances in digital health have created an ecosystem ripe for telemedicine in the field of DR to thrive. Stakeholders and policymakers should adopt a participatory approach to ensure sustained implementation of these technologies after the COVID-19 pandemic. This article belongs to the Topical Collection "Diabetic Eye Disease", managed by Giuseppe Querques.

Impact of Artificial Intelligence Assessment of Diabetic Retinopathy on Referral Service Uptake in a Low-Resource Setting: The RAIDERS Randomized Trial

PURPOSE

This trial was designed to determine if artificial intelligence (AI)-supported diabetic retinopathy (DR) screening improved referral uptake in Rwanda.

DESIGN

The Rwanda Artificial Intelligence for Diabetic Retinopathy Screening (RAIDERS) study was an investigator-masked, parallel-group randomized controlled trial.

PARTICIPANTS

Patients ≥ 18 years of age with known diabetes who required referral for DR based on AI interpretation.

METHODS

The RAIDERS study screened for DR using retinal imaging with AI interpretation implemented at 4 facilities from March 2021 through July 2021. Eligible participants were assigned randomly (1:1) to immediate feedback of AI grading (intervention) or communication of referral advice after human grading was completed 3 to 5 days after the initial screening (control).

MAIN OUTCOME MEASURES

Difference between study groups in the rate of presentation for referral services within 30 days of being informed of the need for a referral visit.

RESULTS

Of the 823 clinic patients who met inclusion criteria, 275 participants (33.4%) showed positive findings for referable DR based on AI screening and were randomized for inclusion in the trial. Study participants (mean age, 50.7 years; 58.2% women) were randomized to the intervention (n = 136 [49.5%]) or control (n = 139 [50.5%]) groups. No significant intergroup differences were found at baseline, and main outcome data were available for analyses for 100% of participants. Referral adherence was statistically significantly higher in the intervention group (70/136 [51.5%]) versus the control group (55/139 [39.6%]; P = 0.048), a 30.1% increase. Older age (odds ratio [OR], 1.04; 95% confidence interval [CI], 1.02-1.05; P < 0.0001), male sex (OR, 2.07; 95% CI, 1.22-3.51; P = 0.007), rural residence (OR, 1.79; 95% CI, 1.07-3.01; P = 0.027), and intervention group (OR, 1.74; 95% CI, 1.05-2.88; P = 0.031) were statistically significantly associated with acceptance of referral in multivariate analyses.

CONCLUSIONS

Immediate feedback on referral status based on AI-supported screening was associated with statistically significantly higher referral adherence compared with delayed communications of results from human graders. These results provide evidence for an important benefit of AI screening in promoting adherence to prescribed treatment for diabetic eye care in sub-Saharan Africa.

Recommendations for the development and use of imaging test sets to investigate the test performance of artificial intelligence in health screening

Rigorous evaluation of artificial intelligence (AI) systems for image classification is essential before deployment into health-care settings, such as screening programmes, so that adoption is effective and safe. A key step in the evaluation process is the external validation of diagnostic performance using a test set of images. We conducted a rapid literature review on methods to develop test sets, published from 2012 to 2020, in English. Using thematic analysis, we mapped themes and coded the principles using the Population, Intervention, and Comparator or Reference standard, Outcome, and Study design framework. A group of screening and AI experts assessed the evidence-based principles for completeness and provided further considerations. From the final 15 principles recommended here, five affect population, one intervention, two comparator, one reference standard, and one both reference standard and comparator. Finally, four are appliable to outcome and one to study design. Principles from the literature were useful to address biases from AI; however, they did not account for screening specific biases, which we now incorporate. The principles set out here should be used to support the development and use of test sets for studies that assess the accuracy of AI within screening programmes, to ensure they are fit for purpose and minimise bias.

Automated Systems for Calculating Arteriovenous Ratio in Retinographies: A Scoping Review

There is evidence of an association between hypertension and retinal arteriolar narrowing. Manual measurement of retinal vessels comes with additional variability, which can be eliminated using automated software. This scoping review aims to summarize research on automated retinal vessel analysis systems. Searches were performed on Medline, Scopus, and Cochrane to find studies examining automated systems for the diagnosis of retinal vascular alterations caused by hypertension using the following keywords: diagnosis; diagnostic screening programs; image processing, computer-assisted; artificial intelligence; electronic data processing; hypertensive retinopathy; hypertension; retinal vessels; arteriovenous ratio and retinal image analysis. The searches generated 433 articles. Of these, 25 articles published from 2010 to 2022 were included in the review. The retinographies analyzed were extracted from international databases and real scenarios. Automated systems to detect alterations in the retinal vasculature are being introduced into clinical practice for diagnosis in ophthalmology and other medical specialties due to the association of such changes with various diseases. These systems make the classification of hypertensive retinopathy and cardiovascular risk more reliable. They also make it possible for diagnosis to be performed in primary care, thus optimizing ophthalmological visits.

Robust Collaborative Learning of Patch-Level and Image-Level Annotations for Diabetic Retinopathy Grading From Fundus Image

Diabetic retinopathy (DR) grading from fundus images has attracted increasing interest in both academic and industrial communities. Most convolutional neural network-based algorithms treat DR grading as a classification task via image-level annotations. However, these algorithms have not fully explored the valuable information in the DR-related lesions. In this article, we present a robust framework, which collaboratively utilizes patch-level and image-level annotations, for DR severity grading. By an end-to-end optimization, this framework can bidirectionally exchange the fine-grained lesion and image-level grade information. As a result, it exploits more discriminative features for DR grading. The proposed framework shows better performance than the recent state-of-the-art algorithms and three clinical ophthalmologists with over nine years of experience. By testing on datasets of different distributions (such as label and camera), we prove that our algorithm is robust when facing image quality and distribution variations that commonly exist in real-world practice. We inspect the proposed framework through extensive ablation studies to indicate the effectiveness and necessity of each motivation. The code and some valuable annotations are now publicly available.

Artificial intelligence and its impact on the domains of universal health coverage, health emergencies and health promotion: An overview of systematic reviews

BACKGROUND

Artificial intelligence is fueling a new revolution in medicine and in the healthcare sector. Despite the growing evidence on the benefits of artificial intelligence there are several aspects that limit the measure of its impact in people's health. It is necessary to assess the current status on the application of AI towards the improvement of people's health in the domains defined by WHO's Thirteenth General Programme of Work (GPW13) and the European Programme of Work (EPW), to inform about trends, gaps, opportunities, and challenges.

OBJECTIVE

To perform a systematic overview of systematic reviews on the application of artificial intelligence in the people's health domains as defined in the GPW13 and provide a comprehensive and updated map on the application specialties of artificial intelligence in terms of methodologies, algorithms, data sources, outcomes, predictors, performance, and methodological quality.

METHODS

A systematic search in MEDLINE, EMBASE, Cochrane and IEEEXplore was conducted between January 2015 and June 2021 to collect systematic reviews using a combination of keywords related to the domains of universal health coverage, health emergencies protection, and better health and wellbeing as defined by the WHO's PGW13 and EPW. Eligibility criteria was based on methodological quality and the inclusion of practical implementation of artificial intelligence. Records were classified and labeled using ICD-11 categories into the domains of the GPW13. Descriptors related to the area of implementation, type of modeling, data entities, outcomes and implementation on care delivery were extracted using a structured form and methodological aspects of the included reviews studies was assessed using the AMSTAR checklist.

RESULTS

The search strategy resulted in the screening of 815 systematic reviews from which 203 were assessed for eligibility and 129 were included in the review. The most predominant domain for artificial intelligence applications was Universal Health Coverage (N = 98) followed by Health Emergencies (N = 16) and Better Health and Wellbeing (N = 15). Neoplasms area on Universal Health Coverage was the disease area featuring most of the applications (21.7 %, N = 28). The reviews featured analytics primarily over both public and private data sources (67.44 %, N = 87). The most used type of data was medical imaging (31.8 %, N = 41) and predictors based on regions of interest and clinical data. The most prominent subdomain of Artificial Intelligence was Machine Learning (43.4 %, N = 56), in which Support Vector Machine method was predominant (20.9 %, N = 27). Regarding the purpose, the application of Artificial Intelligence I is focused on the prediction of the diseases (36.4 %, N = 47). With respect to the validation, more than a half of the reviews (54.3 %, N = 70) did not report a validation procedure and, whenever available, the main performance indicator was the accuracy (28.7 %, N = 37). According to the methodological quality assessment, a third of the reviews (34.9 %, N = 45) implemented methods for analysis the risk of bias and the overall AMSTAR score below was 5 (4.01 ± 1.93) on all the included systematic reviews.

CONCLUSION

Artificial intelligence is being used for disease modelling, diagnose, classification and prediction in the three domains of GPW13. However, the evidence is often limited to laboratory and the level of adoption is largely unbalanced between ICD-11 categoriesand diseases. Data availability is a determinant factor on the developmental stage of artificial intelligence applications. Most of the reviewed studies show a poor methodological quality and are at high risk of bias, which limits the reproducibility of the results and the reliability of translating these applications to real clinical scenarios. The analyzed papers show results only in laboratory and testing scenarios and not in clinical trials nor case studies, limiting the supporting evidence to transfer artificial intelligence to actual care delivery.

Diabetic retinopathy screening in the emerging era of artificial intelligence

Diabetic retinopathy is a frequent complication in diabetes and a leading cause of visual impairment. Regular eye screening is imperative to detect sight-threatening stages of diabetic retinopathy such as proliferative diabetic retinopathy and diabetic macular oedema in order to treat these before irreversible visual loss occurs. Screening is cost-effective and has been implemented in various countries in Europe and elsewhere. Along with optimised diabetes care, this has substantially reduced the risk of visual loss. Nevertheless, the growing number of patients with diabetes poses an increasing burden on healthcare systems and automated solutions are needed to alleviate the task of screening and improve diagnostic accuracy. Deep learning by convolutional neural networks is an optimised branch of artificial intelligence that is particularly well suited to automated image analysis. Pivotal studies have demonstrated high sensitivity and specificity for classifying advanced stages of diabetic retinopathy and identifying diabetic macular oedema in optical coherence tomography scans. Based on this, different algorithms have obtained regulatory approval for clinical use and have recently been implemented to some extent in a few countries. Handheld mobile devices are another promising option for self-monitoring, but so far they have not demonstrated comparable image quality to that of fundus photography using non-portable retinal cameras, which is the gold standard for diabetic retinopathy screening. Such technology has the potential to be integrated in telemedicine-based screening programmes, enabling self-captured retinal images to be transferred virtually to reading centres for analysis and planning of further steps. While emerging technologies have shown a lot of promise, clinical implementation has been sparse. Legal obstacles and difficulties in software integration may partly explain this, but it may also indicate that existing algorithms may not necessarily integrate well with national screening initiatives, which often differ substantially between countries.

A New Deep Learning Algorithm with Activation Mapping for Diabetic Retinopathy: Backtesting after 10 Years of Tele-Ophthalmology

We report the development of a deep learning algorithm (AI) to detect signs of diabetic retinopathy (DR) from fundus images. For this, we use a ResNet-50 neural network with a double resolution, the addition of Squeeze–Excitation blocks, pre-trained in ImageNet, and trained for 50 epochs using the Adam optimizer. The AI-based algorithm not only classifies an image as pathological or not but also detects and highlights those signs that allow DR to be identified. For development, we have used a database of about half a million images classified in a real clinical environment by family doctors (FDs), ophthalmologists, or both. The AI was able to detect more than 95% of cases worse than mild DR and had 70% fewer misclassifications of healthy cases than FDs. In addition, the AI was able to detect DR signs in 1258 patients before they were detected by FDs, representing 7.9% of the total number of DR patients detected by the FDs. These results suggest that AI is at least comparable to the evaluation of FDs. We suggest that it may be useful to use signaling tools such as an aid to diagnosis rather than an AI as a stand-alone tool.

Protocol for a systematic review and meta-analysis of the diagnostic accuracy of artificial intelligence for grading of ophthalmology imaging modalities

BACKGROUND

With the rise of artificial intelligence (AI) in ophthalmology, the need to define its diagnostic accuracy is increasingly important. The review aims to elucidate the diagnostic accuracy of AI algorithms in screening for all ophthalmic conditions in patient care settings that involve digital imaging modalities, using the reference standard of human graders.

METHODS

This is a systematic review and meta-analysis. A literature search will be conducted on Ovid MEDLINE, Ovid EMBASE, and Wiley Cochrane CENTRAL from January 1, 2000, to December 20, 2021. Studies will be selected via screening the titles and abstracts, followed by full-text screening. Articles that compare the results of AI-graded ophthalmic images with results from human graders as a reference standard will be included; articles that do not will be excluded. The systematic review software DistillerSR will be used to automate part of the screening process as an adjunct to human reviewers. After the full-text screening, data will be extracted from each study via the categories of study characteristics, patient information, AI methods, intervention, and outcomes. Risk of bias will be scored using Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) by two trained independent reviewers. Disagreements at any step will be addressed by a third adjudicator. The study results will include summary receiver operating characteristic (sROC) curve plots as well as pooled sensitivity and specificity of artificial intelligence for detection of any ophthalmic conditions based on imaging modalities compared to the reference standard. Statistics will be calculated in the R statistical software.

DISCUSSION

This study will provide novel insights into the diagnostic accuracy of AI in new domains of ophthalmology that have not been previously studied. The protocol also outlines the use of an AI-based software to assist in article screening, which may serve as a reference for improving the efficiency and accuracy of future large systematic reviews.

TRIAL REGISTRATION

PROSPERO, CRD42021274441.

Artificial Intelligence-Assisted Score Analysis for Predicting the Expression of the Immunotherapy Biomarker PD-L1 in Lung Cancer

Programmed cell death ligand 1 (PD-L1) is a critical biomarker for predicting the response to immunotherapy. However, traditional quantitative evaluation of PD-L1 expression using immunohistochemistry staining remains challenging for pathologists. Here we developed a deep learning (DL)-based artificial intelligence (AI) model to automatically analyze the immunohistochemical expression of PD-L1 in lung cancer patients. A total of 1,288 patients with lung cancer were included in the study. The diagnostic ability of three different AI models (M1, M2, and M3) was assessed in both PD-L1 (22C3) and PD-L1 (SP263) assays. M2 and M3 showed improved performance in the evaluation of PD-L1 expression in the PD-L1 (22C3) assay, especially at 1% cutoff. Highly accurate performance in the PD-L1 (SP263) was also achieved, with accuracy and specificity of 96.4 and 96.8% in both M2 and M3, respectively. Moreover, the diagnostic results of these three AI-assisted models were highly consistent with those from the pathologist. Similar performances of M1, M2, and M3 in the 22C3 dataset were also obtained in lung adenocarcinoma and lung squamous cell carcinoma in both sampling methods. In conclusion, these results suggest that AI-assisted diagnostic models in PD-L1 expression are a promising tool for improving the efficiency of clinical pathologists.

A systematic review of transfer learning-based approaches for diabetic retinopathy detection

Cases of diabetes and related diabetic retinopathy (DR) have been increasing at an alarming rate in modern times. Early detection of DR is an important problem since it may cause permanent blindness in the late stages. In the last two decades, many different approaches have been applied in DR detection. Reviewing academic literature shows that deep neural networks (DNNs) have become the most preferred approach for DR detection. Among these DNN approaches, Convolutional Neural Network (CNN) models are the most used ones in the field of medical image classification. Designing a new CNN architecture is a tedious and time-consuming approach. Additionally, training an enormous number of parameters is also a difficult task. Due to this reason, instead of training CNNs from scratch, using pre-trained models has been suggested in recent years as transfer learning approach. Accordingly, the present study as a review focuses on DNN and Transfer Learning based applications of DR detection considering 43 publications between 2015 and 2021. The published papers are summarized using 3 figures and 10 tables, giving information about 29 pre-trained CNN models, 13 DR data sets and standard performance metrics.

Medical deep learning-A systematic meta-review

Deep learning has remarkably impacted several different scientific disciplines over the last few years. For example, in image processing and analysis, deep learning algorithms were able to outperform other cutting-edge methods. Additionally, deep learning has delivered state-of-the-art results in tasks like autonomous driving, outclassing previous attempts. There are even instances where deep learning outperformed humans, for example with object recognition and gaming. Deep learning is also showing vast potential in the medical domain. With the collection of large quantities of patient records and data, and a trend towards personalized treatments, there is a great need for automated and reliable processing and analysis of health information. Patient data is not only collected in clinical centers, like hospitals and private practices, but also by mobile healthcare apps or online websites. The abundance of collected patient data and the recent growth in the deep learning field has resulted in a large increase in research efforts. In Q2/2020, the search engine PubMed returned already over 11,000 results for the search term 'deep learning', and around 90% of these publications are from the last three years. However, even though PubMed represents the largest search engine in the medical field, it does not cover all medical-related publications. Hence, a complete overview of the field of 'medical deep learning' is almost impossible to obtain and acquiring a full overview of medical sub-fields is becoming increasingly more difficult. Nevertheless, several review and survey articles about medical deep learning have been published within the last few years. They focus, in general, on specific medical scenarios, like the analysis of medical images containing specific pathologies. With these surveys as a foundation, the aim of this article is to provide the first high-level, systematic meta-review of medical deep learning surveys.

Cardiovascular Risk Stratification in Diabetic Retinopathy via Atherosclerotic Pathway in COVID-19/Non-COVID-19 Frameworks Using Artificial Intelligence Paradigm: A Narrative Review

Diabetes is one of the main causes of the rising cases of blindness in adults. This microvascular complication of diabetes is termed diabetic retinopathy (DR) and is associated with an expanding risk of cardiovascular events in diabetes patients. DR, in its various forms, is seen to be a powerful indicator of atherosclerosis. Further, the macrovascular complication of diabetes leads to coronary artery disease (CAD). Thus, the timely identification of cardiovascular disease (CVD) complications in DR patients is of utmost importance. Since CAD risk assessment is expensive for low-income countries, it is important to look for surrogate biomarkers for risk stratification of CVD in DR patients. Due to the common genetic makeup between the coronary and carotid arteries, low-cost, high-resolution imaging such as carotid B-mode ultrasound (US) can be used for arterial tissue characterization and risk stratification in DR patients. The advent of artificial intelligence (AI) techniques has facilitated the handling of large cohorts in a big data framework to identify atherosclerotic plaque features in arterial ultrasound. This enables timely CVD risk assessment and risk stratification of patients with DR. Thus, this review focuses on understanding the pathophysiology of DR, retinal and CAD imaging, the role of surrogate markers for CVD, and finally, the CVD risk stratification of DR patients. The review shows a step-by-step cyclic activity of how diabetes and atherosclerotic disease cause DR, leading to the worsening of CVD. We propose a solution to how AI can help in the identification of CVD risk. Lastly, we analyze the role of DR/CVD in the COVID-19 framework.

Deep Learning to Detect OCT-derived Diabetic Macular Edema from Color Retinal Photographs: A Multicenter Validation Study

PURPOSE

To validate the generalizability of a deep learning system (DLS) that detects diabetic macular edema (DME) from 2-dimensional color fundus photographs (CFP), for which the reference standard for retinal thickness and fluid presence is derived from 3-dimensional OCT.

DESIGN

Retrospective validation of a DLS across international datasets.

PARTICIPANTS

Paired CFP and OCT of patients from diabetic retinopathy (DR) screening programs or retina clinics. The DLS was developed using data sets from Thailand, the United Kingdom, and the United States and validated using 3060 unique eyes from 1582 patients across screening populations in Australia, India, and Thailand. The DLS was separately validated in 698 eyes from 537 screened patients in the United Kingdom with mild DR and suspicion of DME based on CFP.

METHODS

The DLS was trained using DME labels from OCT. The presence of DME was based on retinal thickening or intraretinal fluid. The DLS's performance was compared with expert grades of maculopathy and to a previous proof-of-concept version of the DLS. We further simulated the integration of the current DLS into an algorithm trained to detect DR from CFP.

MAIN OUTCOME MEASURES

The superiority of specificity and noninferiority of sensitivity of the DLS for the detection of center-involving DME, using device-specific thresholds, compared with experts.

RESULTS

The primary analysis in a combined data set spanning Australia, India, and Thailand showed the DLS had 80% specificity and 81% sensitivity, compared with expert graders, who had 59% specificity and 70% sensitivity. Relative to human experts, the DLS had significantly higher specificity (P = 0.008) and noninferior sensitivity (P < 0.001). In the data set from the United Kingdom, the DLS had a specificity of 80% (P < 0.001 for specificity of >50%) and a sensitivity of 100% (P = 0.02 for sensitivity of > 90%).

CONCLUSIONS

The DLS can generalize to multiple international populations with an accuracy exceeding that of experts. The clinical value of this DLS to reduce false-positive referrals, thus decreasing the burden on specialist eye care, warrants a prospective evaluation.

Five-Year Cost-Effectiveness Modeling of Primary Care-Based, Nonmydriatic Automated Retinal Image Analysis Screening Among Low-Income Patients With Diabetes

BACKGROUND

Artificial intelligence-based technology systems offer an alternative solution for diabetic retinopathy (DR) screening compared with standard, in-office dilated eye examinations. We performed a cost-effectiveness analysis of Automated Retinal Image Analysis System (ARIAS)-based DR screening in a primary care medicine clinic that serves a low-income patient population.

METHODS

A model-based, cost-effectiveness analysis of two DR screening systems was created utilizing data from a recent study comparing adherence rates to follow-up eye care among adults ages 18 or older with a clinical diagnosis of diabetes. In the study, the patients were prescreened with an ARIAS-based, nonmydriatic (undilated), point-of-care tool in the primary care setting and were compared with patients with diabetes who were referred for dilated retinal screening without prescreening, as is the current standard of care. Using a Markov model with microsimulation resulting in a total of 600 000 simulated patient experiences, we calculated the incremental cost-utility ratio (ICUR) of the two screening approaches, with regard to five-year cost-effectiveness of DR screening and treatment of vision-threatening DR.

RESULTS

At five years, ARIAS-based screening showed similar utility as the standard of care screening systems. However, ARIAS reduced costs by 23.3%, with an ICUR of $258 721.81 comparing the current practice to ARIAS.

CONCLUSIONS

Primary care-based ARIAS DR screening is cost-effective when compared with standard of care screening methods.

Untangling Computer-Aided Diagnostic System for Screening Diabetic Retinopathy Based on Deep Learning Techniques

Diabetic Retinopathy (DR) is a predominant cause of visual impairment and loss. Approximately 285 million worldwide population is affected with diabetes, and one-third of these patients have symptoms of DR. Specifically, it tends to affect the patients with 20 years or more with diabetes, but it can be reduced by early detection and proper treatment. Diagnosis of DR by using manual methods is a time-consuming and expensive task which involves trained ophthalmologists to observe and evaluate DR using digital fundus images of the retina. This study aims to systematically find and analyze high-quality research work for the diagnosis of DR using deep learning approaches. This research comprehends the DR grading, staging protocols and also presents the DR taxonomy. Furthermore, identifies, compares, and investigates the deep learning-based algorithms, techniques, and, methods for classifying DR stages. Various publicly available dataset used for deep learning have also been analyzed and dispensed for descriptive and empirical understanding for real-time DR applications. Our in-depth study shows that in the last few years there has been an increasing inclination towards deep learning approaches. 35% of the studies have used Convolutional Neural Networks (CNNs), 26% implemented the Ensemble CNN (ECNN) and, 13% Deep Neural Networks (DNN) are amongst the most used algorithms for the DR classification. Thus using the deep learning algorithms for DR diagnostics have future research potential for DR early detection and prevention based solution.

Diagnostic accuracy of teleretinal screening for detection of diabetic retinopathy and age-related macular degeneration: a systematic review and meta-analysis

Objective

To evaluate the diagnostic accuracy of teleretinal screening compared with face-to-face examination for detection of diabetic retinopathy (DR) and age-related macular degeneration (AMD).

Methods and analysis

This study adhered to the Preferred Reporting Items for a Systematic Review and Meta-analysis of Diagnostic Test Accuracy Studies (PRISMA-DTA). A comprehensive search of OVID MEDLINE, EMBASE and Cochrane CENTRAL was performed from January 2010 to July 2021. QUADAS-2 tool was used to assess methodological quality and applicability of the studies. A bivariate random effects model was used to perform the meta-analysis. Referrable DR was defined as any disease severity equal to or worse than moderate non-proliferative DR or diabetic macular oedema (DMO).

Results

28 articles were included. Teleretinal screening achieved a sensitivity of 0.91 (95% CI: 0.82 to 0.96) and specificity of 0.88 (0.74 to 0.95) for any DR (13 studies, n=7207, Grading of Recommendations, Assessment, Development and Evaluation (GRADE) low). Accuracy for referrable DR (10 studies, n=6373, GRADE moderate) was lower with a sensitivity of 0.88 (0.81 to 0.93) and specificity of 0.86 (0.79 to 0.90). After exclusion of ungradable images, the specificity for referrable DR increased to 0.95 (0.90 to 0.98), while the sensitivity remained nearly unchanged at 0.85 (0.76 to 0.91). Teleretinal screening achieved a sensitivity of 0.71 (0.49 to 0.86) and specificity of 0.88 (0.85 to 0.90) for detection of AMD (three studies, n=697, GRADE low).

Conclusion

Teleretinal screening is highly accurate for detecting any DR and DR warranting referral. Data for AMD screening is promising but warrants further investigation.

PROSPERO registration number

CRD42020191994.

Quality assessment standards in artificial intelligence diagnostic accuracy systematic reviews: a meta-research study

Artificial intelligence (AI) centred diagnostic systems are increasingly recognised as robust solutions in healthcare delivery pathways. In turn, there has been a concurrent rise in secondary research studies regarding these technologies in order to influence key clinical and policymaking decisions. It is therefore essential that these studies accurately appraise methodological quality and risk of bias within shortlisted trials and reports. In order to assess whether this critical step is performed, we undertook a meta-research study evaluating adherence to the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) tool within AI diagnostic accuracy systematic reviews. A literature search was conducted on all studies published from 2000 to December 2020. Of 50 included reviews, 36 performed the quality assessment, of which 27 utilised the QUADAS-2 tool. Bias was reported across all four domains of QUADAS-2. Two hundred forty-three of 423 studies (57.5%) across all systematic reviews utilising QUADAS-2 reported a high or unclear risk of bias in the patient selection domain, 110 (26%) reported a high or unclear risk of bias in the index test domain, 121 (28.6%) in the reference standard domain and 157 (37.1%) in the flow and timing domain. This study demonstrates the incomplete uptake of quality assessment tools in reviews of AI-based diagnostic accuracy studies and highlights inconsistent reporting across all domains of quality assessment. Poor standards of reporting act as barriers to clinical implementation. The creation of an AI-specific extension for quality assessment tools of diagnostic accuracy AI studies may facilitate the safe translation of AI tools into clinical practice.

A deep learning approach to permanent tooth germ detection on pediatric panoramic radiographs

Purpose

The aim of this study was to assess the performance of a deep learning system for permanent tooth germ detection on pediatric panoramic radiographs.

Materials and Methods

In total, 4518 anonymized panoramic radiographs of children between 5 and 13 years of age were collected. YOLOv4, a convolutional neural network (CNN)-based object detection model, was used to automatically detect permanent tooth germs. Panoramic images of children processed in LabelImg were trained and tested in the YOLOv4 algorithm. True-positive, false-positive, and false-negative rates were calculated. A confusion matrix was used to evaluate the performance of the model.

Results

The YOLOv4 model, which detected permanent tooth germs on pediatric panoramic radiographs, provided an average precision value of 94.16% and an F1 value of 0.90, indicating a high level of significance. The average YOLOv4 inference time was 90 ms.

Conclusion

The detection of permanent tooth germs on pediatric panoramic X-rays using a deep learning-based approach may facilitate the early diagnosis of tooth deficiency or supernumerary teeth and help dental practitioners find more accurate treatment options while saving time and effort.

Cross-Camera External Validation for Artificial Intelligence Software in Diagnosis of Diabetic Retinopathy

AIMS

To investigate the applicability of deep learning image assessment software VeriSee DR to different color fundus cameras for the screening of diabetic retinopathy (DR).

METHODS

Color fundus images of diabetes patients taken with three different nonmydriatic fundus cameras, including 477 Topcon TRC-NW400, 459 Topcon TRC-NW8 series, and 471 Kowa nonmyd 8 series that were judged as "gradable" by one ophthalmologist were enrolled for validation. VeriSee DR was then used for the diagnosis of referable DR according to the International Clinical Diabetic Retinopathy Disease Severity Scale. Gradability, sensitivity, and specificity were calculated for each camera model.

RESULTS

All images (100%) from the three camera models were gradable for VeriSee DR. The sensitivity for diagnosing referable DR in the TRC-NW400, TRC-NW8, and non-myd 8 series was 89.3%, 94.6%, and 95.7%, respectively, while the specificity was 94.2%, 90.4%, and 89.3%, respectively. Neither the sensitivity nor the specificity differed significantly between these camera models and the original camera model used for VeriSee DR development (p = 0.40, p = 0.065, respectively).

CONCLUSIONS

VeriSee DR was applicable to a variety of color fundus cameras with 100% agreement with ophthalmologists in terms of gradability and good sensitivity and specificity for the diagnosis of referable DR.

ANALYSIS OF TRANSFER LEARNING FOR SELECT RETINAL DISEASE CLASSIFICATION

PURPOSE

To analyze the effect of transfer learning for classification of diabetic retinopathy (DR) by fundus photography and select retinal diseases by spectral domain optical coherence tomography (SD-OCT).

METHODS

Five widely used open-source deep neural networks and four customized simpler and smaller networks, termed the CBR family, were trained and evaluated on two tasks: 1) classification of DR using fundus photography and 2) classification of drusen, choroidal neovascularization, and diabetic macular edema using SD-OCT. For DR classification, the quadratic weighted Kappa coefficient was used to measure the level of agreement between each network and ground truth-labeled test cases. For SD-OCT-based classification, accuracy was calculated for each network. Kappa and accuracy were compared between iterations with and without use of transfer learning for each network to assess for its effect.

RESULTS

For DR classification, Kappa increased with transfer learning for all networks (range of increase 0.152-0.556). For SD-OCT-based classification, accuracy increased for four of five open-source deep neural networks (range of increase 1.8%-3.5%), slightly decreased for the remaining deep neural network (-0.6%), decreased slightly for three of four CBR networks (range of decrease 0.9%-1.8%), and decreased by 9.6% for the remaining CBR network.

CONCLUSION

Transfer learning improved performance, as measured by Kappa, for DR classification for all networks, although the effect ranged from small to substantial. Transfer learning had minimal effect on accuracy for SD-OCT-based classification for eight of the nine networks analyzed. These results imply that transfer learning may substantially increase performance for DR classification but may have minimal effect for SD-OCT-based classification.

Barriers and facilitators to diabetic retinopathy screening within Australian primary care

Background

Despite recent incentives through Medicare (Australia’s universal health insurance scheme) to increase retinal screening rates in primary care, comprehensive diabetic retinopathy (DR) screening has not been reached in Australia. The current study aimed to identify key factors affecting the delivery of diabetic retinopathy (DR) screening in Australian general practices.

Methods

A descriptive qualitative study involving in-depth interviews was carried out from November 2019 to March 2020. Using purposive snowballing sampling, 15 general practitioners (GPs) were recruited from urban and rural general practices in New South Wales and Western Australia. A semi-structured interview guide was used to collect data from participants. All interviews were conducted over the phone by one facilitator, and each interview lasted up to 45 min. The Socio-Ecological Model was used to inform the content of the interview topic guides and subsequent data analysis. Recorded data were transcribed verbatim, and thematic analysis was conducted to identify and classify recurrent themes.

Results

Of 15 GPs interviewed, 13 were male doctors, and the mean age was 54.7 ± 15.5 years. Seven participants were practising in urban areas, while eight were practising in regional or remote areas. All participants had access to a direct ophthalmoscope, but none owned retinal cameras. None of the participants reported performing DR screening. Only three participants were aware of the Medicare Benefits Schedule (MBS) items 12,325 and 12,326 that allow GPs to bill for retinal screening. Seven themes, a combination of facilitators and barriers, emerged from interviews with the GPs. Despite the strong belief in their role in managing chronic diseases, barriers such as costs of retinal cameras, time constraints, lack of skills to make DR diagnosis, and unawareness of Medicare incentives for non-mydriatic retinal photography made it difficult to conduct DR screening in general practice. However, several enabling strategies to deliver DR screening within primary care include increasing GPs’ access to continuing professional development, subsidising the cost of retinal cameras, and the need for a champion ace to take the responsibility of retinal photography.

Conclusion

This study identified essential areas at the system level that require addressing to promote the broader implementation of DR screening, in particular, a nationwide awareness campaign to maximise the use of MBS items, improve GPs’ competency, and subsidise costs of the retinal cameras for small and rural general practices.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12875-021-01586-7.

Artificial intelligence in dry eye disease

Dry eye disease (DED) has a prevalence of between 5 and 50%, depending on the diagnostic criteria used and population under study. However, it remains one of the most underdiagnosed and undertreated conditions in ophthalmology. Many tests used in the diagnosis of DED rely on an experienced observer for image interpretation, which may be considered subjective and result in variation in diagnosis. Since artificial intelligence (AI) systems are capable of advanced problem solving, use of such techniques could lead to more objective diagnosis. Although the term 'AI' is commonly used, recent success in its applications to medicine is mainly due to advancements in the sub-field of machine learning, which has been used to automatically classify images and predict medical outcomes. Powerful machine learning techniques have been harnessed to understand nuances in patient data and medical images, aiming for consistent diagnosis and stratification of disease severity. This is the first literature review on the use of AI in DED. We provide a brief introduction to AI, report its current use in DED research and its potential for application in the clinic. Our review found that AI has been employed in a wide range of DED clinical tests and research applications, primarily for interpretation of interferometry, slit-lamp and meibography images. While initial results are promising, much work is still needed on model development, clinical testing and standardisation.

Artificial intelligence for improving sickle cell retinopathy diagnosis and management

Sickle cell retinopathy is often initially asymptomatic even in proliferative stages, but can progress to cause vision loss due to vitreous haemorrhages or tractional retinal detachments. Challenges with access and adherence to screening dilated fundus examinations, particularly in medically underserved areas where the burden of sickle cell disease is highest, highlight the need for novel approaches to screening for patients with vision-threatening sickle cell retinopathy. This article reviews the existing literature on and suggests future research directions for coupling artificial intelligence with multimodal retinal imaging to expand access to automated, accurate, imaging-based screening for sickle cell retinopathy. Given the variability in retinal specialist practice patterns with regards to monitoring and treatment of sickle cell retinopathy, we also discuss recent progress toward development of machine learning models that can quantitatively track disease progression over time. These artificial intelligence-based applications have great potential for informing evidence-based and resource-efficient clinical diagnosis and management of sickle cell retinopathy.

Use of artificial intelligence for image analysis in breast cancer screening programmes: systematic review of test accuracy

OBJECTIVE

To examine the accuracy of artificial intelligence (AI) for the detection of breast cancer in mammography screening practice.

DESIGN

Systematic review of test accuracy studies.

DATA SOURCES

Medline, Embase, Web of Science, and Cochrane Database of Systematic Reviews from 1 January 2010 to 17 May 2021.

ELIGIBILITY CRITERIA

Studies reporting test accuracy of AI algorithms, alone or in combination with radiologists, to detect cancer in women's digital mammograms in screening practice, or in test sets. Reference standard was biopsy with histology or follow-up (for screen negative women). Outcomes included test accuracy and cancer type detected.

STUDY SELECTION AND SYNTHESIS

Two reviewers independently assessed articles for inclusion and assessed the methodological quality of included studies using the QUality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool. A single reviewer extracted data, which were checked by a second reviewer. Narrative data synthesis was performed.

RESULTS

Twelve studies totalling 131 822 screened women were included. No prospective studies measuring test accuracy of AI in screening practice were found. Studies were of poor methodological quality. Three retrospective studies compared AI systems with the clinical decisions of the original radiologist, including 79 910 women, of whom 1878 had screen detected cancer or interval cancer within 12 months of screening. Thirty four (94%) of 36 AI systems evaluated in these studies were less accurate than a single radiologist, and all were less accurate than consensus of two or more radiologists. Five smaller studies (1086 women, 520 cancers) at high risk of bias and low generalisability to the clinical context reported that all five evaluated AI systems (as standalone to replace radiologist or as a reader aid) were more accurate than a single radiologist reading a test set in the laboratory. In three studies, AI used for triage screened out 53%, 45%, and 50% of women at low risk but also 10%, 4%, and 0% of cancers detected by radiologists.

CONCLUSIONS

Current evidence for AI does not yet allow judgement of its accuracy in breast cancer screening programmes, and it is unclear where on the clinical pathway AI might be of most benefit. AI systems are not sufficiently specific to replace radiologist double reading in screening programmes. Promising results in smaller studies are not replicated in larger studies. Prospective studies are required to measure the effect of AI in clinical practice. Such studies will require clear stopping rules to ensure that AI does not reduce programme specificity.

STUDY REGISTRATION

Protocol registered as PROSPERO CRD42020213590.

The Impact of Artificial Intelligence and Deep Learning in Eye Diseases: A Review

Artificial intelligence (AI) is a subset of computer science dealing with the development and training of algorithms that try to replicate human intelligence. We report a clinical overview of the basic principles of AI that are fundamental to appreciating its application to ophthalmology practice. Here, we review the most common eye diseases, focusing on some of the potential challenges and limitations emerging with the development and application of this new technology into ophthalmology.

Diagnostic performance of deep-learning-based screening methods for diabetic retinopathy in primary care-A meta-analysis

Background

Diabetic retinopathy (DR) affects 10–24% of patients with diabetes mellitus type 1 or 2 in the primary care (PC) sector. As early detection is crucial for treatment, deep learning screening methods in PC setting could potentially aid in an accurate and timely diagnosis.

Purpose

The purpose of this meta-analysis was to determine the current state of knowledge regarding deep learning (DL) screening methods for DR in PC.

Data sources

A systematic literature search was conducted using Medline, Web of Science, and Scopus to identify suitable studies.

Study selection

Suitable studies were selected by two researchers independently. Studies assessing DL methods and the suitability of these screening systems (diagnostic parameters such as sensitivity and specificity, information on datasets and setting) in PC were selected. Excluded were studies focusing on lesions, applying conventional diagnostic imaging tools, conducted in secondary or tertiary care, and all publication types other than original research studies on human subjects.

Data extraction

The following data was extracted from included studies: authors, title, year of publication, objectives, participants, setting, type of intervention/method, reference standard, grading scale, outcome measures, dataset, risk of bias, and performance measures.

Data synthesis and conclusion

The summed sensitivity of all included studies was 87% and specificity was 90%. Given a prevalence of DR of 10% in patients with DM Type 2 in PC, the negative predictive value is 98% while the positive predictive value is 49%.

Limitations

Selected studies showed a high variation in sample size and quality and quantity of available data.