A scoping review of transfer learning research on medical image analysis using ImageNet

Multi-perspective region-based CNNs for vertebrae labeling in intraoperative long-length images

PURPOSE

Effective aggregation of intraoperative x-ray images that capture the patient anatomy from multiple view-angles has the potential to enable and improve automated image analysis that can be readily performed during surgery. We present multi-perspective region-based neural networks that leverage knowledge of the imaging geometry for automatic vertebrae labeling in Long-Film images - a novel tomographic imaging modality with an extended field-of-view for spine imaging.

METHOD

A multi-perspective network architecture was designed to exploit small view-angle disparities produced by a multi-slot collimator and consolidate information from overlapping image regions. A second network incorporates large view-angle disparities to jointly perform labeling on images from multiple views (viz., AP and lateral). A recurrent module incorporates contextual information and enforce anatomical order for the detected vertebrae. The three modules are combined to form the multi-view multi-slot (MVMS) network for labeling vertebrae using images from all available perspectives. The network was trained on images synthesized from 297 CT images and tested on 50 AP and 50 lateral Long-Film images acquired from 13 cadaveric specimens. Labeling performance of the multi-perspective networks was evaluated with respect to the number of vertebrae appearances and presence of surgical instrumentation.

RESULTS

The MVMS network achieved an F1 score of >96% and an average vertebral localization error of 3.3 mm, with 88.3% labeling accuracy on both AP and lateral images - (15.5% and 35.0% higher than conventional Faster R-CNN on AP and lateral views, respectively). Aggregation of multiple appearances of the same vertebra using the multi-slot network significantly improved the labeling accuracy (p < 0.05). Using the multi-view network, labeling accuracy on the more challenging lateral views was improved to the same level as that of the AP views. The approach demonstrated robustness to the presence of surgical instrumentation, commonly encountered in intraoperative images, and achieved comparable performance in images with and without instrumentation (88.9% vs. 91.2% labeling accuracy).

CONCLUSION

The MVMS network demonstrated effective multi-perspective aggregation, providing means for accurate, automated vertebrae labeling during spine surgery. The algorithms may be generalized to other imaging tasks and modalities that involve multiple views with view-angle disparities (e.g., bi-plane radiography). Predicted labels can help avoid adverse events during surgery (e.g., wrong-level surgery), establish correspondence with labels in preoperative modalities to facilitate image registration, and enable automated measurement of spinal alignment metrics for intraoperative assessment of spinal curvature.

A novel heterogeneous transfer learning method based on data stitching for the sequential coding brain computer interface

OBJECTIVE

For the brain computer interface (BCI), it is necessary to collect enough electroencephalography (EEG) signals to train the classification model. When the operation dimension of BCI is large, it will bring great burden to data acquisition. Fortunately, this problem can be solved by our proposed transfer learning method.

METHOD

For the sequential coding experimental paradigm, the multi-band data stitching with label alignment and tangent space mapping (MDSLATSM) algorithm is proposed as a novel heterogeneous transfer learning method. After filtering by multi-band filtering, the artificial signals can be obtained by data stitching from the source domain, which build a bridge between the source domain and target domain. To make the distribution of two domains closer, their covariance matrices are aligned by label alignment. After mapping to the tangent space, the features are extracted from the Riemannian manifold. Finally, the classification results are obtained with feature selection and classification.

RESULTS

Our data set includes the EEG signals from 16 subjects. For the heterogeneous transfer learning of cross-label, the average classification accuracy is 78.28%. MDSLATSM is also tested for cross-subject, and the average classification accuracy is 64.01%, which is better than existing methods.

SIGNIFICANCE

Combining multi-band filtering, data stitching, label alignment and tangent space mapping, a novel heterogeneous transfer learning method can be achieved with superior performance, which promotes the practical application of the BCI systems.

Cardiac MRI segmentation of the atria based on UU-NET

Background and objective

In today's society, people's work pressure, coupled with irregular diet, lack of exercise and other bad lifestyle, resulting in frequent cardiovascular diseases. Medical imaging has made great progress in modern society, among which the role of MRI in cardiovascular field is self-evident. Based on this research background, how to process cardiac MRI quickly and accurately by computer has been extensively discussed. By comparing and analyzing several traditional image segmentation and deep learning image segmentation, this paper proposes the left and right atria segmentation algorithm of cardiac MRI based on UU-NET network.

Methods

In this paper, an atrial segmentation algorithm for cardiac MRI images in UU-NET network is proposed. Firstly, U-shaped upper and lower sampling modules are constructed by using residual theory, which are used as encoders and decoders of the model. Then, the modules are interconnected to form multiple paths from input to output to increase the information transmission capacity of the model.

Results

The segmentation method based on UU-NET network has achieved good results proposed in this paper, compared with the current mainstream image segmentation algorithm results have been improved to a certain extent. Through the analysis of the experimental results, the image segmentation algorithm based on UU-NET network on the data set, its performance in the verification set and online set is higher than other grid models. The DSC in the verification set is 96.7%, and the DSC in the online set is 96.7%, which is nearly one percentage point higher than the deconvolution neural network model. The hausdorff distance (HD) is 1.2 mm. Compared with other deep learning models, it is significantly improved (about 3 mm error is reduced), and the time is 0.4 min.

Conclusion

The segmentation algorithm based on UU-NET improves the segmentation accuracy obviously compared with other segmentation models. Our technique will be able to help diagnose and treat cardiac complications.

Smartphone video nystagmography using convolutional neural networks: ConVNG

Background

Eye movement abnormalities are commonplace in neurological disorders. However, unaided eye movement assessments lack granularity. Although videooculography (VOG) improves diagnostic accuracy, resource intensiveness precludes its broad use. To bridge this care gap, we here validate a framework for smartphone video-based nystagmography capitalizing on recent computer vision advances.

Methods

A convolutional neural network was fine-tuned for pupil tracking using > 550 annotated frames: ConVNG. In a cross-sectional approach, slow-phase velocity of optokinetic nystagmus was calculated in 10 subjects using ConVNG and VOG. Equivalence of accuracy and precision was assessed using the “two one-sample t-test” (TOST) and Bayesian interval-null approaches. ConVNG was systematically compared to OpenFace and MediaPipe as computer vision (CV) benchmarks for gaze estimation.

Results

ConVNG tracking accuracy reached 9–15% of an average pupil diameter. In a fully independent clinical video dataset, ConVNG robustly detected pupil keypoints (median prediction confidence 0.85). SPV measurement accuracy was equivalent to VOG (TOST p < 0.017; Bayes factors (BF) > 24). ConVNG, but not MediaPipe, achieved equivalence to VOG in all SPV calculations. Median precision was 0.30°/s for ConVNG, 0.7°/s for MediaPipe and 0.12°/s for VOG. ConVNG precision was significantly higher than MediaPipe in vertical planes, but both algorithms’ precision was inferior to VOG.

Conclusions

ConVNG enables offline smartphone video nystagmography with an accuracy comparable to VOG and significantly higher precision than MediaPipe, a benchmark computer vision application for gaze estimation. This serves as a blueprint for highly accessible tools with potential to accelerate progress toward precise and personalized Medicine.

Rapid Convolutional Neural Networks for Gram-Stained Image Classification at Inference Time on Mobile Devices: Empirical Study from Transfer Learning to Optimization

Despite the emergence of mobile health and the success of deep learning (DL), deploying production-ready DL models to resource-limited devices remains challenging. Especially, during inference time, the speed of DL models becomes relevant. We aimed to accelerate inference time for Gram-stained analysis, which is a tedious and manual task involving microorganism detection on whole slide images. Three DL models were optimized in three steps: transfer learning, pruning and quantization and then evaluated on two Android smartphones. Most convolutional layers (≥80%) had to be retrained for adaptation to the Gram-stained classification task. The combination of pruning and quantization demonstrated its utility to reduce the model size and inference time without compromising model quality. Pruning mainly contributed to model size reduction by 15×, while quantization reduced inference time by 3× and decreased model size by 4×. The combination of two reduced the baseline model by an overall factor of 46×. Optimized models were smaller than 6 MB and were able to process one image in <0.6 s on a Galaxy S10. Our findings demonstrate that methods for model compression are highly relevant for the successful deployment of DL solutions to resource-limited devices.

Covid-MANet: Multi-task attention network for explainable diagnosis and severity assessment of COVID-19 from CXR images

The devastating outbreak of Coronavirus Disease (COVID-19) cases in early 2020 led the world to face health crises. Subsequently, the exponential reproduction rate of COVID-19 disease can only be reduced by early diagnosis of COVID-19 infection cases correctly. The initial research findings reported that radiological examinations using CT and CXR modality have successfully reduced false negatives by RT-PCR test. This research study aims to develop an explainable diagnosis system for the detection and infection region quantification of COVID-19 disease. The existing research studies successfully explored deep learning approaches with higher performance measures but lacked generalization and interpretability for COVID-19 diagnosis. In this study, we address these issues by the Covid-MANet network, an automated end-to-end multi-task attention network that works for 5 classes in three stages for COVID-19 infection screening. The first stage of the Covid-MANet network localizes attention of the model to the relevant lungs region for disease recognition. The second stage of the Covid-MANet network differentiates COVID-19 cases from bacterial pneumonia, viral pneumonia, normal and tuberculosis cases, respectively. To improve the interpretation and explainability, three experiments have been conducted in exploration of the most coherent and appropriate classification approach. Moreover, the multi-scale attention model MA-DenseNet201 proposed for the classification of COVID-19 cases. The final stage of the Covid-MANet network quantifies the proportion of infection and severity of COVID-19 in the lungs. The COVID-19 cases are graded into more specific severity levels such as mild, moderate, severe, and critical as per the score assigned by the RALE scoring system. The MA-DenseNet201 classification model outperforms eight state-of-the-art CNN models, in terms of sensitivity and interpretation with lung localization network. The COVID-19 infection segmentation by UNet with DenseNet121 encoder achieves dice score of 86.15% outperforming UNet, UNet++, AttentionUNet, R2UNet, with VGG16, ResNet50 and DenseNet201 encoder. The proposed network not only classifies images based on the predicted label but also highlights the infection by segmentation/localization of model-focused regions to support explainable decisions. MA-DenseNet201 model with a segmentation-based cropping approach achieves maximum interpretation of 96% with COVID-19 sensitivity of 97.75%. Finally, based on class-varied sensitivity analysis Covid-MANet ensemble network of MA-DenseNet201, ResNet50 and MobileNet achieve 95.05% accuracy and 98.75% COVID-19 sensitivity. The proposed model is externally validated on an unseen dataset, yields 98.17% COVID-19 sensitivity.

Classification of scanning electron microscope images of pharmaceutical excipients using deep convolutional neural networks with transfer learning

Convolutional Neural Networks (CNNs) are image analysis techniques that have been applied to image classification in various fields. In this study, we applied a CNN to classify scanning electron microscopy (SEM) images of pharmaceutical raw material powders to determine if a CNN can evaluate particle morphology. We tested 10 pharmaceutical excipients with widely different particle morphologies. SEM images for each excipient were acquired and divided into training, validation, and test sets. Classification models were constructed by applying transfer learning to pretrained CNN models such as VGG16 and ResNet50. The results of a 5-fold cross-validation showed that the classification accuracy of the CNN model was sufficiently high using either pretrained model and that the type of excipient could be classified with high accuracy. The results suggest that the CNN model can detect differences in particle morphology, such as particle size, shape, and surface condition. By applying Grad-CAM to the constructed CNN model, we succeeded in finding particularly important regions in the particle image of the excipients. CNNs have been found to have the potential to be applied to the identification and characterization of raw material powders for pharmaceutical development.

Artificial Intelligence for Radiation Oncology Applications Using Public Datasets

Artificial intelligence (AI) has exceptional potential to positively impact the field of radiation oncology. However, large curated datasets - often involving imaging data and corresponding annotations - are required to develop radiation oncology AI models. Importantly, the recent establishment of Findable, Accessible, Interoperable, Reusable (FAIR) principles for scientific data management have enabled an increasing number of radiation oncology related datasets to be disseminated through data repositories, thereby acting as a rich source of data for AI model building. This manuscript reviews the current and future state of radiation oncology data dissemination, with a particular emphasis on published imaging datasets, AI data challenges, and associated infrastructure. Moreover, we provide historical context of FAIR data dissemination protocols, difficulties in the current distribution of radiation oncology data, and recommendations regarding data dissemination for eventual utilization in AI models. Through FAIR principles and standardized approaches to data dissemination, radiation oncology AI research has nothing to lose and everything to gain.

COVID-19 Detection on Chest X-ray and CT Scan: A Review of the Top-100 Most Cited Papers

Since the beginning of the COVID-19 pandemic, many works have been published proposing solutions to the problems that arose in this scenario. In this vein, one of the topics that attracted the most attention is the development of computer-based strategies to detect COVID-19 from thoracic medical imaging, such as chest X-ray (CXR) and computerized tomography scan (CT scan). By searching for works already published on this theme, we can easily find thousands of them. This is partly explained by the fact that the most severe worldwide pandemic emerged amid the technological advances recently achieved, and also considering the technical facilities to deal with the large amount of data produced in this context. Even though several of these works describe important advances, we cannot overlook the fact that others only use well-known methods and techniques without a more relevant and critical contribution. Hence, differentiating the works with the most relevant contributions is not a trivial task. The number of citations obtained by a paper is probably the most straightforward and intuitive way to verify its impact on the research community. Aiming to help researchers in this scenario, we present a review of the top-100 most cited papers in this field of investigation according to the Google Scholar search engine. We evaluate the distribution of the top-100 papers taking into account some important aspects, such as the type of medical imaging explored, learning settings, segmentation strategy, explainable artificial intelligence (XAI), and finally, the dataset and code availability.

Advances in non-invasive biosensing measures to monitor wound healing progression

Impaired wound healing is a significant financial and medical burden. The synthesis and deposition of extracellular matrix (ECM) in a new wound is a dynamic process that is constantly changing and adapting to the biochemical and biomechanical signaling from the extracellular microenvironments of the wound. This drives either a regenerative or fibrotic and scar-forming healing outcome. Disruptions in ECM deposition, structure, and composition lead to impaired healing in diseased states, such as in diabetes. Valid measures of the principal determinants of successful ECM deposition and wound healing include lack of bacterial contamination, good tissue perfusion, and reduced mechanical injury and strain. These measures are used by wound-care providers to intervene upon the healing wound to steer healing toward a more functional phenotype with improved structural integrity and healing outcomes and to prevent adverse wound developments. In this review, we discuss bioengineering advances in 1) non-invasive detection of biologic and physiologic factors of the healing wound, 2) visualizing and modeling the ECM, and 3) computational tools that efficiently evaluate the complex data acquired from the wounds based on basic science, preclinical, translational and clinical studies, that would allow us to prognosticate healing outcomes and intervene effectively. We focus on bioelectronics and biologic interfaces of the sensors and actuators for real time biosensing and actuation of the tissues. We also discuss high-resolution, advanced imaging techniques, which go beyond traditional confocal and fluorescence microscopy to visualize microscopic details of the composition of the wound matrix, linearity of collagen, and live tracking of components within the wound microenvironment. Computational modeling of the wound matrix, including partial differential equation datasets as well as machine learning models that can serve as powerful tools for physicians to guide their decision-making process are discussed.

An extensive review of state-of-the-art transfer learning techniques used in medical imaging: Open issues and challenges

Deep learning techniques, which use a massive technology known as convolutional neural networks, have shown excellent results in a variety of areas, including image processing and interpretation. However, as the depth of these networks grows, so does the demand for a large amount of labeled data required to train these networks. In particular, the medical field suffers from a lack of images because the procedure for obtaining labeled medical images in the healthcare field is difficult, expensive, and requires specialized expertise to add labels to images. Moreover, the process may be prone to errors and time-consuming. Current research has revealed transfer learning as a viable solution to this problem. Transfer learning allows us to transfer knowledge gained from a previous process to improve and tackle a new problem. This study aims to conduct a comprehensive survey of recent studies that dealt with solving this problem and the most important metrics used to evaluate these methods. In addition, this study identifies problems in transfer learning techniques and highlights the problems of the medical dataset and potential problems that can be addressed in future research. According to our review, many researchers use pre-trained models on the Imagenet dataset (VGG16, ResNet, Inception v3) in many applications such as skin cancer, breast cancer, and diabetic retinopathy classification tasks. These techniques require further investigation of these models, due to training them on natural, non-medical images. In addition, many researchers use data augmentation techniques to expand their dataset and avoid overfitting. However, not enough studies have shown the effect of performance with or without data augmentation. Accuracy, recall, precision, F1 score, receiver operator characteristic curve, and area under the curve (AUC) were the most widely used measures in these studies. Furthermore, we identified problems in the datasets for melanoma and breast cancer and suggested corresponding solutions.

Using Artificial Intelligence as a Diagnostic Decision Support Tool in Skin Disease: Protocol for an Observational Prospective Cohort Study

BACKGROUND

Dermatological conditions are a relevant health problem. Each person has an average of 1.6 skin diseases per year, and consultations for skin pathology represent 20% of the total annual visits to primary care and around 35% are referred to a dermatology specialist. Machine learning (ML) models can be a good tool to help primary care professionals, as it can analyze and optimize complex sets of data. In addition, ML models are increasingly being applied to dermatology as a diagnostic decision support tool using image analysis, especially for skin cancer detection and classification.

OBJECTIVE

This study aims to perform a prospective validation of an image analysis ML model as a diagnostic decision support tool for the diagnosis of dermatological conditions.

METHODS

In this prospective study, 100 consecutive patients who visit a participant general practitioner (GP) with a skin problem in central Catalonia were recruited. Data collection was planned to last 7 months. Anonymized pictures of skin diseases were taken and introduced to the ML model interface (capable of screening for 44 different skin diseases), which returned the top 5 diagnoses by probability. The same image was also sent as a teledermatology consultation following the current stablished workflow. The GP, ML model, and dermatologist's assessments will be compared to calculate the precision, sensitivity, specificity, and accuracy of the ML model. The results will be represented globally and individually for each skin disease class using a confusion matrix and one-versus-all methodology. The time taken to make the diagnosis will also be taken into consideration.

RESULTS

Patient recruitment began in June 2021 and lasted for 5 months. Currently, all patients have been recruited and the images have been shown to the GPs and dermatologists. The analysis of the results has already started.

CONCLUSIONS

This study will provide information about ML models' effectiveness and limitations. External testing is essential for regulating these diagnostic systems to deploy ML models in a primary care practice setting.

Using Artificial Intelligence as a Diagnostic Decision Support Tool in Skin Disease: Protocol for an Observational Prospective Cohort Study

BACKGROUND

Dermatological conditions are a relevant health problem. Each person has an average of 1.6 skin diseases per year, and consultations for skin pathology represent 20% of the total annual visits to primary care and around 35% are referred to a dermatology specialist. Machine learning (ML) models can be a good tool to help primary care professionals, as it can analyze and optimize complex sets of data. In addition, ML models are increasingly being applied to dermatology as a diagnostic decision support tool using image analysis, especially for skin cancer detection and classification.

OBJECTIVE

This study aims to perform a prospective validation of an image analysis ML model as a diagnostic decision support tool for the diagnosis of dermatological conditions.

METHODS

In this prospective study, 100 consecutive patients who visit a participant general practitioner (GP) with a skin problem in central Catalonia were recruited. Data collection was planned to last 7 months. Anonymized pictures of skin diseases were taken and introduced to the ML model interface (capable of screening for 44 different skin diseases), which returned the top 5 diagnoses by probability. The same image was also sent as a teledermatology consultation following the current stablished workflow. The GP, ML model, and dermatologist's assessments will be compared to calculate the precision, sensitivity, specificity, and accuracy of the ML model. The results will be represented globally and individually for each skin disease class using a confusion matrix and one-versus-all methodology. The time taken to make the diagnosis will also be taken into consideration.

RESULTS

Patient recruitment began in June 2021 and lasted for 5 months. Currently, all patients have been recruited and the images have been shown to the GPs and dermatologists. The analysis of the results has already started.

CONCLUSIONS

This study will provide information about ML models' effectiveness and limitations. External testing is essential for regulating these diagnostic systems to deploy ML models in a primary care practice setting.

Terrestrial health applications of visual assessment technology and machine learning in spaceflight associated neuro-ocular syndrome

The neuro-ocular effects of long-duration spaceflight have been termed Spaceflight Associated Neuro-Ocular Syndrome (SANS) and are a potential challenge for future, human space exploration. The underlying pathogenesis of SANS remains ill-defined, but several emerging translational applications of terrestrial head-mounted, visual assessment technology and machine learning frameworks are being studied for potential use in SANS. To develop such technology requires close consideration of the spaceflight environment which is limited in medical resources and imaging modalities. This austere environment necessitates the utilization of low mass, low footprint technology to build a visual assessment system that is comprehensive, accessible, and efficient. In this paper, we discuss the unique considerations for developing this technology for SANS and translational applications on Earth. Several key limitations observed in the austere spaceflight environment share similarities to barriers to care for underserved areas on Earth. We discuss common terrestrial ophthalmic diseases and how machine learning and visual assessment technology for SANS can help increase screening for early intervention. The foundational developments with this novel system may help protect the visual health of both astronauts and individuals on Earth.

Comparison of Different Convolutional Neural Network Activation Functions and Methods for Building Ensembles for Small to Midsize Medical Data Sets

CNNs and other deep learners are now state-of-the-art in medical imaging research. However, the small sample size of many medical data sets dampens performance and results in overfitting. In some medical areas, it is simply too labor-intensive and expensive to amass images numbering in the hundreds of thousands. Building Deep CNN ensembles of pre-trained CNNs is one powerful method for overcoming this problem. Ensembles combine the outputs of multiple classifiers to improve performance. This method relies on the introduction of diversity, which can be introduced on many levels in the classification workflow. A recent ensembling method that has shown promise is to vary the activation functions in a set of CNNs or within different layers of a single CNN. This study aims to examine the performance of both methods using a large set of twenty activations functions, six of which are presented here for the first time: 2D Mexican ReLU, TanELU, MeLU + GaLU, Symmetric MeLU, Symmetric GaLU, and Flexible MeLU. The proposed method was tested on fifteen medical data sets representing various classification tasks. The best performing ensemble combined two well-known CNNs (VGG16 and ResNet50) whose standard ReLU activation layers were randomly replaced with another. Results demonstrate the superiority in performance of this approach.

A Fair Performance Comparison between Complex-Valued and Real-Valued Neural Networks for Disease Detection

Our aim is to contribute to the classification of anomalous patterns in biosignals using this novel approach. We specifically focus on melanoma and heart murmurs. We use a comparative study of two convolution networks in the Complex and Real numerical domains. The idea is to obtain a powerful approach for building portable systems for early disease detection. Two similar algorithmic structures were chosen so that there is no bias determined by the number of parameters to train. Three clinical data sets, ISIC2017, PH2, and Pascal, were used to carry out the experiments. Mean comparison hypothesis tests were performed to ensure statistical objectivity in the conclusions. In all cases, complex-valued networks presented a superior performance for the Precision, Recall, F1 Score, Accuracy, and Specificity metrics in the detection of associated anomalies. The best complex number-based classifier obtained in the Receiving Operating Characteristic (ROC) space presents a Euclidean distance of 0.26127 with respect to the ideal classifier, as opposed to the best real number-based classifier, whose Euclidean distance to the ideal is 0.36022 for the same task of melanoma detection. The 27.46% superiority in this metric, as in the others reported in this work, suggests that complex-valued networks have a greater ability to extract features for more efficient discrimination in the dataset.

Predicting Systemic Health Features from Retinal Fundus Images Using Transfer-Learning-Based Artificial Intelligence Models

While color fundus photos are used in routine clinical practice to diagnose ophthalmic conditions, evidence suggests that ocular imaging contains valuable information regarding the systemic health features of patients. These features can be identified through computer vision techniques including deep learning (DL) artificial intelligence (AI) models. We aim to construct a DL model that can predict systemic features from fundus images and to determine the optimal method of model construction for this task. Data were collected from a cohort of patients undergoing diabetic retinopathy screening between March 2020 and March 2021. Two models were created for each of 12 systemic health features based on the DenseNet201 architecture: one utilizing transfer learning with images from ImageNet and another from 35,126 fundus images. Here, 1277 fundus images were used to train the AI models. Area under the receiver operating characteristics curve (AUROC) scores were used to compare the model performance. Models utilizing the ImageNet transfer learning data were superior to those using retinal images for transfer learning (mean AUROC 0.78 vs. 0.65, p-value < 0.001). Models using ImageNet pretraining were able to predict systemic features including ethnicity (AUROC 0.93), age > 70 (AUROC 0.90), gender (AUROC 0.85), ACE inhibitor (AUROC 0.82), and ARB medication use (AUROC 0.78). We conclude that fundus images contain valuable information about the systemic characteristics of a patient. To optimize DL model performance, we recommend that even domain specific models consider using transfer learning from more generalized image sets to improve accuracy.

Pediatric chest radiograph interpretation: how far has artificial intelligence come? A systematic literature review

Most artificial intelligence (AI) studies have focused primarily on adult imaging, with less attention to the unique aspects of pediatric imaging. The objectives of this study were to (1) identify all publicly available pediatric datasets and determine their potential utility and limitations for pediatric AI studies and (2) systematically review the literature to assess the current state of AI in pediatric chest radiograph interpretation. We searched PubMed, Web of Science and Embase to retrieve all studies from 1990 to 2021 that assessed AI for pediatric chest radiograph interpretation and abstracted the datasets used to train and test AI algorithms, approaches and performance metrics. Of 29 publicly available chest radiograph datasets, 2 datasets included solely pediatric chest radiographs, and 7 datasets included pediatric and adult patients. We identified 55 articles that implemented an AI model to interpret pediatric chest radiographs or pediatric and adult chest radiographs. Classification of chest radiographs as pneumonia was the most common application of AI, evaluated in 65% of the studies. Although many studies report high diagnostic accuracy, most algorithms were not validated on external datasets. Most AI studies for pediatric chest radiograph interpretation have focused on a limited number of diseases, and progress is hindered by a lack of large-scale pediatric chest radiograph datasets.

A Hybrid Matching Network for Fault Diagnosis under Different Working Conditions with Limited Data

Intelligent fault diagnosis methods based on deep learning have achieved much progress in recent years. However, there are two major factors causing serious degradation of the performance of these algorithms in real industrial applications, i.e., limited labeled training data and complex working conditions. To solve these problems, this study proposed a domain generalization-based hybrid matching network utilizing a matching network to diagnose the faults using features encoded by an autoencoder. The main idea was to regularize the feature extractor of the network with an autoencoder in order to reduce the risk of overfitting with limited training samples. In addition, a training strategy using dropout with random changing rates on inputs was implemented to enhance the model’s generalization on unseen domains. The proposed method was validated on two different datasets containing artificial and real faults. The results showed that considerable performance was achieved by the proposed method under cross-domain tasks with limited training samples.

A fully-convolutional residual encoder-decoder neural network to localize breast cancer on histopathology images

Cancer detection in its early stages may allow patients to receive the proper treatment and save lives along with recovering the routine lifestyles. Breast cancer is of the top leading causes of mortality among women all around the globe. A source to find these cancerous nuclei is through analyzing histopathology images. These images, however, are very complex and large. Thus, locating the cancerous nuclei in them is very challenging. Hence, if an expert fails to diagnose their patients via these images, the situation may be exacerbated. Therefore, this study aims to introduce a method to mask as many cancer nuclei on histopathology images as possible with a high visual aesthetic to make them distinguishable by experts easily. A tailored residual fully convolutional encoder-decoder neural network based on end-to-end learning is proposed to issue the matter. The proposed method is evaluated quantitatively and qualitatively on ER + BCa H&E-stained dataset. The average detection accuracy achieved by the method is 98.61%, which is much better than that of competitors.

Automatic Classification for Sagittal Craniofacial Patterns Based on Different Convolutional Neural Networks

(1) Background: The present study aims to evaluate and compare the model performances of different convolutional neural networks (CNNs) used for classifying sagittal skeletal patterns. (2) Methods: A total of 2432 lateral cephalometric radiographs were collected. They were labeled as Class I, Class II, and Class III patterns, according to their ANB angles and Wits values. The radiographs were randomly divided into the training, validation, and test sets in the ratio of 70%:15%:15%. Four different CNNs, namely VGG16, GoogLeNet, ResNet152, and DenseNet161, were trained, and their model performances were compared. (3) Results: The accuracy of the four CNNs was ranked as follows: DenseNet161 > ResNet152 > VGG16 > GoogLeNet. DenseNet161 had the highest accuracy, while GoogLeNet possessed the smallest model size and fastest inference speed. The CNNs showed better capabilities for identifying Class III patterns, followed by Classes II and I. Most of the samples that were misclassified by the CNNs were boundary cases. The activation area confirmed the CNNs without overfitting and indicated that artificial intelligence could recognize the compensatory dental features in the anterior region of the jaws and lips. (4) Conclusions: CNNs can quickly and effectively assist orthodontists in the diagnosis of sagittal skeletal classification patterns.

Medical Image Classification Utilizing Ensemble Learning and Levy Flight-Based Honey Badger Algorithm on 6G-Enabled Internet of Things

Recently, the 6G-enabled Internet of Medical Things (IoMT) has played a key role in the development of functional health systems due to the massive data generated daily from the hospitals. Therefore, the automatic detection and prediction of future risks such as pneumonia and retinal diseases are still under research and study. However, traditional approaches did not yield good results for accurate diagnosis. In this paper, a robust 6G-enabled IoMT framework is proposed for medical image classification with an ensemble learning (EL)-based model. EL is achieved using MobileNet and DenseNet architecture as a feature extraction backbone. In addition, the developed framework uses a modified honey badger algorithm (HBA) based on Levy flight (LFHBA) as a feature selection method that aims to remove the irrelevant features from those extracted features using the EL model. For evaluation of the performance of the proposed framework, the chest X-ray (CXR) dataset and the optical coherence tomography (OCT) dataset were employed. The accuracy of our technique was 87.10% on the CXR dataset and 94.32% on OCT dataset—both very good results. Compared to other current methods, the proposed method is more accurate and efficient than other well-known and popular algorithms.

Automated knee cartilage segmentation for heterogeneous clinical MRI using generative adversarial networks with transfer learning

BACKGROUND

This study aimed to build a deep learning model to automatically segment heterogeneous clinical MRI scans by optimizing a pre-trained model built from a homogeneous research dataset with transfer learning.

METHODS

Conditional generative adversarial networks pretrained on the Osteoarthritis Initiative MR images was transferred to 30 sets of heterogenous MR images collected from clinical routines. Two trained radiologists manually segmented the 30 sets of clinical MR images for model training, validation and test. The model performance was compared to models trained from scratch with different datasets, as well as two radiologists. A 5-fold cross validation was performed.

RESULTS

The transfer learning model obtained an overall averaged Dice coefficient of 0.819, an averaged 95 percentile Hausdorff distance of 1.463 mm, and an averaged average symmetric surface distance of 0.350 mm on the 5 random holdout test sets. A 5-fold cross validation had a mean Dice coefficient of 0.801, mean 95 percentile Hausdorff distance of 1.746 mm, and mean average symmetric surface distance of 0.364 mm. It outperformed other models and performed similarly as the radiologists.

CONCLUSIONS

A transfer learning model was able to automatically segment knee cartilage, with performance comparable to human, using heterogeneous clinical MR images with a small training data size. In addition, the model proved robust when tested through cross validation and on images from a different vendor. We found it feasible to perform fully automated cartilage segmentation of clinical knee MR images, which would facilitate the clinical application of quantitative MRI techniques and other prediction models for improved patient treatment planning.

Transfer learning for medical image classification: a literature review

BACKGROUND

Transfer learning (TL) with convolutional neural networks aims to improve performances on a new task by leveraging the knowledge of similar tasks learned in advance. It has made a major contribution to medical image analysis as it overcomes the data scarcity problem as well as it saves time and hardware resources. However, transfer learning has been arbitrarily configured in the majority of studies. This review paper attempts to provide guidance for selecting a model and TL approaches for the medical image classification task.

METHODS

425 peer-reviewed articles were retrieved from two databases, PubMed and Web of Science, published in English, up until December 31, 2020. Articles were assessed by two independent reviewers, with the aid of a third reviewer in the case of discrepancies. We followed the PRISMA guidelines for the paper selection and 121 studies were regarded as eligible for the scope of this review. We investigated articles focused on selecting backbone models and TL approaches including feature extractor, feature extractor hybrid, fine-tuning and fine-tuning from scratch.

RESULTS

The majority of studies (n = 57) empirically evaluated multiple models followed by deep models (n = 33) and shallow (n = 24) models. Inception, one of the deep models, was the most employed in literature (n = 26). With respect to the TL, the majority of studies (n = 46) empirically benchmarked multiple approaches to identify the optimal configuration. The rest of the studies applied only a single approach for which feature extractor (n = 38) and fine-tuning from scratch (n = 27) were the two most favored approaches. Only a few studies applied feature extractor hybrid (n = 7) and fine-tuning (n = 3) with pretrained models.

CONCLUSION

The investigated studies demonstrated the efficacy of transfer learning despite the data scarcity. We encourage data scientists and practitioners to use deep models (e.g. ResNet or Inception) as feature extractors, which can save computational costs and time without degrading the predictive power.

Multimodal image encoding pre-training for diabetic retinopathy grading

Diabetic retinopathy is an increasingly prevalent eye disorder that can lead to severe vision impairment. The severity grading of the disease using retinal images is key to provide an adequate treatment. However, in order to learn the diverse patterns and complex relations that are required for the grading, deep neural networks require very large annotated datasets that are not always available. This has been typically addressed by reusing networks that were pre-trained for natural image classification, hence relying on additional annotated data from a different domain. In contrast, we propose a novel pre-training approach that takes advantage of unlabeled multimodal visual data commonly available in ophthalmology. The use of multimodal visual data for pre-training purposes has been previously explored by training a network in the prediction of one image modality from another. However, that approach does not ensure a broad understanding of the retinal images, given that the network may exclusively focus on the similarities between modalities while ignoring the differences. Thus, we propose a novel self-supervised pre-training that explicitly teaches the networks to learn the common characteristics between modalities as well as the characteristics that are exclusive to the input modality. This provides a complete comprehension of the input domain and facilitates the training of downstream tasks that require a broad understanding of the retinal images, such as the grading of diabetic retinopathy. To validate and analyze the proposed approach, we performed an exhaustive experimentation on different public datasets. The transfer learning performance for the grading of diabetic retinopathy is evaluated under different settings while also comparing against previous state-of-the-art pre-training approaches. Additionally, a comparison against relevant state-of-the-art works for the detection and grading of diabetic retinopathy is also provided. The results show a satisfactory performance of the proposed approach, which outperforms previous pre-training alternatives in the grading of diabetic retinopathy.

Analysis of high-resolution reconstruction of medical images based on deep convolutional neural networks in lung cancer diagnostics

BACKGROUND AND OBJECTIVE

To study the diagnostic effect of 64-slice spiral CT and MRI high-resolution images based on deep convolutional neural networks(CNN) in lung cancer.

METHODS

In this paper, we Select 74 patients with highly suspected lung cancer who were treated in our hospital from January 2017 to January 2021 as the research objects. The enhanced 64-slice spiral CT and MRI were used to detect and diagnose respectively, and the images and accuracy of CT diagnosis and MRI diagnosis were retrospectively analyzed.

RESULTS

The accuracy of CT diagnosis is 94.6% (70/74), and the accuracy of MRI diagnosis is 89.2% (66/74). CT examination has the advantages of non-invasive, convenient operation and fast examination. MRI is showing there are advantages in the relationship between the chest wall and the mediastinum, and the relationship between the lesion and the large blood vessels.

CONCLUSION

Enhanced CT and MRI examinations based on convolutional neural networks(CNN) to improve image clarity have high application value in the diagnosis of lung cancer patients, but the focus of performance is different.

Phase Attention Model for Prediction of Early Recurrence of Hepatocellular Carcinoma With Multi-Phase CT Images and Clinical Data

Hepatocellular carcinoma (HCC) is a primary liver cancer that produces a high mortality rate. It is one of the most common malignancies worldwide, especially in Asia, Africa, and southern Europe. Although surgical resection is an effective treatment, patients with HCC are at risk of recurrence after surgery. Preoperative early recurrence prediction for patients with liver cancer can help physicians develop treatment plans and will enable physicians to guide patients in postoperative follow-up. However, the conventional clinical data based methods ignore the imaging information of patients. Certain studies have used radiomic models for early recurrence prediction in HCC patients with good results, and the medical images of patients have been shown to be effective in predicting the recurrence of HCC. In recent years, deep learning models have demonstrated the potential to outperform the radiomics-based models. In this paper, we propose a prediction model based on deep learning that contains intra-phase attention and inter-phase attention. Intra-phase attention focuses on important information of different channels and space in the same phase, whereas inter-phase attention focuses on important information between different phases. We also propose a fusion model to combine the image features with clinical data. Our experiment results prove that our fusion model has superior performance over the models that use clinical data only or the CT image only. Our model achieved a prediction accuracy of 81.2%, and the area under the curve was 0.869.

Quantum transfer learning for breast cancer detection

One of the areas with the potential to be explored in quantum computing (QC) is machine learning (ML), giving rise to quantum machine learning (QML). In an era when there is so much data, ML may benefit from either speed, complexity or smaller amounts of storage. In this work, we explore a quantum approach to a machine learning problem. Based on the work of Mari et al., we train a set of hybrid classical-quantum neural networks using transfer learning (TL). Our task was to solve the problem of classifying full-image mammograms into malignant and benign, provided by BCDR. Throughout the course of our work, heatmaps were used to highlight the parts of the mammograms that were being targeted by the networks while evaluating different performance metrics. Our work shows that this method may hold benefits regarding the generalization of complex data; however, further tests are needed. We also show that, depending on the task, some architectures perform better than others. Nonetheless, our results were superior to those reported in the state-of-the-art (accuracy of 84% against 76.9%, respectively). In addition, experiments were conducted in a real quantum device, and results were compared with the classical and simulator.

A hybrid mask RCNN-based tool to localize dental cavities from real-time mixed photographic images

Nearly 3.5 billion humans have oral health issues, including dental caries, which requires dentist-patient exposure in oral examinations. The automated approaches identify and locate carious regions from dental images by localizing and processing either colored photographs or X-ray images taken via specialized dental photography cameras. The dentists' interpretation of carious regions is difficult since the detected regions are masked using solid coloring and limited to a particular dental image type. The software-based automated tools to localize caries from dental images taken via ordinary cameras requires further investigation. This research provided a mixed dataset of dental photographic (colored or X-ray) images, instantiated a deep learning approach to enhance the existing dental image carious regions' localization procedure, and implemented a full-fledged tool to present carious regions via simple dental images automatically. The instantiation mainly exploits the mixed dataset of dental images (colored photographs or X-rays) collected from multiple sources and pre-trained hybrid Mask RCNN to localize dental carious regions. The evaluations performed by the dentists showed that the correctness of annotated datasets is up to 96%, and the accuracy of the proposed system is between 78% and 92%. Moreover, the system achieved the overall satisfaction level of dentists above 80%.

High-Frequency Ultrasound Dataset for Deep Learning-Based Image Quality Assessment

This study aims at high-frequency ultrasound image quality assessment for computer-aided diagnosis of skin. In recent decades, high-frequency ultrasound imaging opened up new opportunities in dermatology, utilizing the most recent deep learning-based algorithms for automated image analysis. An individual dermatological examination contains either a single image, a couple of pictures, or an image series acquired during the probe movement. The estimated skin parameters might depend on the probe position, orientation, or acquisition setup. Consequently, the more images analyzed, the more precise the obtained measurements. Therefore, for the automated measurements, the best choice is to acquire the image series and then analyze its parameters statistically. However, besides the correctly received images, the resulting series contains plenty of non-informative data: Images with different artifacts, noise, or the images acquired for the time stamp when the ultrasound probe has no contact with the patient skin. All of them influence further analysis, leading to misclassification or incorrect image segmentation. Therefore, an automated image selection step is crucial. To meet this need, we collected and shared 17,425 high-frequency images of the facial skin from 516 measurements of 44 patients. Two experts annotated each image as correct or not. The proposed framework utilizes a deep convolutional neural network followed by a fuzzy reasoning system to assess the acquired data’s quality automatically. Different approaches to binary and multi-class image analysis, based on the VGG-16 model, were developed and compared. The best classification results reach 91.7% accuracy for the first, and 82.3% for the second analysis, respectively.

Automated selection of mid-height intervertebral disc slice in traverse lumbar spine MRI using a combination of deep learning feature and machine learning classifier

Abnormalities and defects that can cause lumbar spinal stenosis often occur in the Intervertebral Disc (IVD) of the patient’s lumbar spine. Their automatic detection and classification require an application of an image analysis algorithm on suitable images, such as mid-sagittal images or traverse mid-height intervertebral disc slices, as inputs. Hence the process of selecting and separating these images from other medical images in the patient’s set of scans is necessary. However, the technological progress in making this process automated is still lagging behind other areas in medical image classification research. In this paper, we report the result of our investigation on the suitability and performance of different approaches of machine learning to automatically select the best traverse plane that cuts closest to the half-height of an IVD from a database of lumbar spine MRI images. This study considers images features extracted using eleven different pre-trained Deep Convolution Neural Network (DCNN) models. We investigate the effectiveness of three dimensionality-reduction techniques and three feature-selection techniques on the classification performance. We also investigate the performance of five different Machine Learning (ML) algorithms and three Fully Connected (FC) neural network learning optimizers which are used to train an image classifier with hyperparameter optimization using a wide range of hyperparameter options and values. The different combinations of methods are tested on a publicly available lumbar spine MRI dataset consisting of MRI studies of 515 patients with symptomatic back pain. Our experiment shows that applying the Support Vector Machine algorithm with a short Gaussian kernel on full-length image features extracted using a pre-trained DenseNet201 model is the best approach to use. This approach gives the minimum per-class classification performance of around 0.88 when measured using the precision and recall metrics. The median performance measured using the precision metric ranges from 0.95 to 0.99 whereas that using the recall metric ranges from 0.93 to 1.0. When only considering the L3/L4, L4/L5, and L5/S1 classes, the minimum F1-Scores range between 0.93 to 0.95, whereas the median F1-Scores range between 0.97 to 0.99.

Hybrid Deep Feature Generation for Appropriate Face Mask Use Detection

Mask usage is one of the most important precautions to limit the spread of COVID-19. Therefore, hygiene rules enforce the correct use of face coverings. Automated mask usage classification might be used to improve compliance monitoring. This study deals with the problem of inappropriate mask use. To address that problem, 2075 face mask usage images were collected. The individual images were labeled as either mask, no masked, or improper mask. Based on these labels, the following three cases were created: Case 1: mask versus no mask versus improper mask, Case 2: mask versus no mask + improper mask, and Case 3: mask versus no mask. This data was used to train and test a hybrid deep feature-based masked face classification model. The presented method comprises of three primary stages: (i) pre-trained ResNet101 and DenseNet201 were used as feature generators; each of these generators extracted 1000 features from an image; (ii) the most discriminative features were selected using an improved RelieF selector; and (iii) the chosen features were used to train and test a support vector machine classifier. That resulting model attained 95.95%, 97.49%, and 100.0% classification accuracy rates on Case 1, Case 2, and Case 3, respectively. Having achieved these high accuracy values indicates that the proposed model is fit for a practical trial to detect appropriate face mask use in real time.

Transfer learning for non-image data in clinical research: A scoping review

BACKGROUND

Transfer learning is a form of machine learning where a pre-trained model trained on a specific task is reused as a starting point and tailored to another task in a different dataset. While transfer learning has garnered considerable attention in medical image analysis, its use for clinical non-image data is not well studied. Therefore, the objective of this scoping review was to explore the use of transfer learning for non-image data in the clinical literature.

METHODS AND FINDINGS

We systematically searched medical databases (PubMed, EMBASE, CINAHL) for peer-reviewed clinical studies that used transfer learning on human non-image data. We included 83 studies in the review. More than half of the studies (63%) were published within 12 months of the search. Transfer learning was most often applied to time series data (61%), followed by tabular data (18%), audio (12%) and text (8%). Thirty-three (40%) studies applied an image-based model to non-image data after transforming data into images (e.g. spectrograms). Twenty-nine (35%) studies did not have any authors with a health-related affiliation. Many studies used publicly available datasets (66%) and models (49%), but fewer shared their code (27%).

CONCLUSIONS

In this scoping review, we have described current trends in the use of transfer learning for non-image data in the clinical literature. We found that the use of transfer learning has grown rapidly within the last few years. We have identified studies and demonstrated the potential of transfer learning in clinical research in a wide range of medical specialties. More interdisciplinary collaborations and the wider adaption of reproducible research principles are needed to increase the impact of transfer learning in clinical research.

Neural network classifiers for images of genetic conditions with cutaneous manifestations

Neural networks have shown strong potential in research and in healthcare. Mainly due to the need for large datasets, these applications have focused on common medical conditions, where more data are typically available. Leveraging publicly available data, we trained a neural network classifier on images of rare genetic conditions with skin findings. We used approximately 100 images per condition to classify 6 different genetic conditions. We analyzed both preprocessed images that were cropped to show only the skin lesions as well as more complex images showing features such as the entire body segment, the person, and/or the background. The classifier construction process included attribution methods to visualize which pixels were most important for computer-based classification. Our classifier was significantly more accurate than pediatricians or medical geneticists for both types of images and suggests steps for further research involving clinical scenarios and other applications.

Breast Tumour Classification Using Ultrasound Elastography with Machine Learning: A Systematic Scoping Review

Ultrasound elastography can quantify stiffness distribution of tissue lesions and complements conventional B-mode ultrasound for breast cancer screening. Recently, the development of computer-aided diagnosis has improved the reliability of the system, whilst the inception of machine learning, such as deep learning, has further extended its power by facilitating automated segmentation and tumour classification. The objective of this review was to summarize application of the machine learning model to ultrasound elastography systems for breast tumour classification. Review databases included PubMed, Web of Science, CINAHL, and EMBASE. Thirteen (n = 13) articles were eligible for review. Shear-wave elastography was investigated in six articles, whereas seven studies focused on strain elastography (5 freehand and 2 Acoustic Radiation Force). Traditional computer vision workflow was common in strain elastography with separated image segmentation, feature extraction, and classifier functions using different algorithm-based methods, neural networks or support vector machines (SVM). Shear-wave elastography often adopts the deep learning model, convolutional neural network (CNN), that integrates functional tasks. All of the reviewed articles achieved sensitivity ³ 80%, while only half of them attained acceptable specificity ³ 95%. Deep learning models did not necessarily perform better than traditional computer vision workflow. Nevertheless, there were inconsistencies and insufficiencies in reporting and calculation, such as the testing dataset, cross-validation, and methods to avoid overfitting. Most of the studies did not report loss or hyperparameters. Future studies may consider using the deep network with an attention layer to locate the targeted object automatically and online training to facilitate efficient re-training for sequential data.

A Novel Multistage Transfer Learning for Ultrasound Breast Cancer Image Classification

Breast cancer diagnosis is one of the many areas that has taken advantage of artificial intelligence to achieve better performance, despite the fact that the availability of a large medical image dataset remains a challenge. Transfer learning (TL) is a phenomenon that enables deep learning algorithms to overcome the issue of shortage of training data in constructing an efficient model by transferring knowledge from a given source task to a target task. However, in most cases, ImageNet (natural images) pre-trained models that do not include medical images, are utilized for transfer learning to medical images. Considering the utilization of microscopic cancer cell line images that can be acquired in large amount, we argue that learning from both natural and medical datasets improves performance in ultrasound breast cancer image classification. The proposed multistage transfer learning (MSTL) algorithm was implemented using three pre-trained models: EfficientNetB2, InceptionV3, and ResNet50 with three optimizers: Adam, Adagrad, and stochastic gradient de-scent (SGD). Dataset sizes of 20,400 cancer cell images, 200 ultrasound images from Mendeley and 400 ultrasound images from the MT-Small-Dataset were used. ResNet50-Adagrad-based MSTL achieved a test accuracy of 99 ± 0.612% on the Mendeley dataset and 98.7 ± 1.1% on the MT-Small-Dataset, averaging over 5-fold cross validation. A p-value of 0.01191 was achieved when comparing MSTL against ImageNet based TL for the Mendeley dataset. The result is a significant improvement in the performance of artificial intelligence methods for ultrasound breast cancer classification compared to state-of-the-art methods and could remarkably improve the early diagnosis of breast cancer in young women.

CovNet: A Transfer Learning Framework for Automatic COVID-19 Detection From Crowd-Sourced Cough Sounds

Since the COronaVIrus Disease 2019 (COVID-19) outbreak, developing a digital diagnostic tool to detect COVID-19 from respiratory sounds with computer audition has become an essential topic due to its advantages of being swift, low-cost, and eco-friendly. However, prior studies mainly focused on small-scale COVID-19 datasets. To build a robust model, the large-scale multi-sound FluSense dataset is utilised to help detect COVID-19 from cough sounds in this study. Due to the gap between FluSense and the COVID-19-related datasets consisting of cough only, the transfer learning framework (namely CovNet) is proposed and applied rather than simply augmenting the training data with FluSense. The CovNet contains (i) a parameter transferring strategy and (ii) an embedding incorporation strategy. Specifically, to validate the CovNet's effectiveness, it is used to transfer knowledge from FluSense to COUGHVID, a large-scale cough sound database of COVID-19 negative and COVID-19 positive individuals. The trained model on FluSense and COUGHVID is further applied under the CovNet to another two small-scale cough datasets for COVID-19 detection, the COVID-19 cough sub-challenge (CCS) database in the INTERSPEECH Computational Paralinguistics challengE (ComParE) challenge and the DiCOVA Track-1 database. By training four simple convolutional neural networks (CNNs) in the transfer learning framework, our approach achieves an absolute improvement of 3.57% over the baseline of DiCOVA Track-1 validation of the area under the receiver operating characteristic curve (ROC AUC) and an absolute improvement of 1.73% over the baseline of ComParE CCS test unweighted average recall (UAR).

Transfer learning in a deep convolutional neural network for implant fixture classification: A pilot study

Purpose

This study aimed to evaluate the performance of transfer learning in a deep convolutional neural network for classifying implant fixtures.

Materials and Methods

Periapical radiographs of implant fixtures obtained using the Superline (Dentium Co. Ltd., Seoul, Korea), TS III (Osstem Implant Co. Ltd., Seoul, Korea), and Bone Level Implant (Institut Straumann AG, Basel, Switzerland) systems were selected from patients who underwent dental implant treatment. All 355 implant fixtures comprised the total dataset and were annotated with the name of the system. The total dataset was split into a training dataset and a test dataset at a ratio of 8 to 2, respectively. YOLOv3 (You Only Look Once version 3, available at https://pjreddie.com/darknet/yolo/), a deep convolutional neural network that has been pretrained with a large image dataset of objects, was used to train the model to classify fixtures in periapical images, in a process called transfer learning. This network was trained with the training dataset for 100, 200, and 300 epochs. Using the test dataset, the performance of the network was evaluated in terms of sensitivity, specificity, and accuracy.

Results

When YOLOv3 was trained for 200 epochs, the sensitivity, specificity, accuracy, and confidence score were the highest for all systems, with overall results of 94.4%, 97.9%, 96.7%, and 0.75, respectively. The network showed the best performance in classifying Bone Level Implant fixtures, with 100.0% sensitivity, specificity, and accuracy.

Conclusion

Through transfer learning, high performance could be achieved with YOLOv3, even using a small amount of data.

Cellpose 2.0: how to train your own model

Pretrained neural network models for biological segmentation can provide good out-of-the-box results for many image types. However, such models do not allow users to adapt the segmentation style to their specific needs and can perform suboptimally for test images that are very different from the training images. Here we introduce Cellpose 2.0, a new package that includes an ensemble of diverse pretrained models as well as a human-in-the-loop pipeline for rapid prototyping of new custom models. We show that models pretrained on the Cellpose dataset can be fine-tuned with only 500-1,000 user-annotated regions of interest (ROI) to perform nearly as well as models trained on entire datasets with up to 200,000 ROI. A human-in-the-loop approach further reduced the required user annotation to 100-200 ROI, while maintaining high-quality segmentations. We provide software tools such as an annotation graphical user interface, a model zoo and a human-in-the-loop pipeline to facilitate the adoption of Cellpose 2.0.

Artificial intelligence, machine learning, and deep learning for clinical outcome prediction

AI is a broad concept, grouping initiatives that use a computer to perform tasks that would usually require a human to complete. AI methods are well suited to predict clinical outcomes. In practice, AI methods can be thought of as functions that learn the outcomes accompanying standardized input data to produce accurate outcome predictions when trialed with new data. Current methods for cleaning, creating, accessing, extracting, augmenting, and representing data for training AI clinical prediction models are well defined. The use of AI to predict clinical outcomes is a dynamic and rapidly evolving arena, with new methods and applications emerging. Extraction or accession of electronic health care records and combining these with patient genetic data is an area of present attention, with tremendous potential for future growth. Machine learning approaches, including decision tree methods of Random Forest and XGBoost, and deep learning techniques including deep multi-layer and recurrent neural networks, afford unique capabilities to accurately create predictions from high dimensional, multimodal data. Furthermore, AI methods are increasing our ability to accurately predict clinical outcomes that previously were difficult to model, including time-dependent and multi-class outcomes. Barriers to robust AI-based clinical outcome model deployment include changing AI product development interfaces, the specificity of regulation requirements, and limitations in ensuring model interpretability, generalizability, and adaptability over time.

AI-Enabled, Ultrasound-Guided Handheld Robotic Device for Femoral Vascular Access

Hemorrhage is a leading cause of trauma death, particularly in prehospital environments when evacuation is delayed. Obtaining central vascular access to a deep artery or vein is important for administration of emergency drugs and analgesics, and rapid replacement of blood volume, as well as invasive sensing and emerging life-saving interventions. However, central access is normally performed by highly experienced critical care physicians in a hospital setting. We developed a handheld AI-enabled interventional device, AI-GUIDE (Artificial Intelligence Guided Ultrasound Interventional Device), capable of directing users with no ultrasound or interventional expertise to catheterize a deep blood vessel, with an initial focus on the femoral vein. AI-GUIDE integrates with widely available commercial portable ultrasound systems and guides a user in ultrasound probe localization, venous puncture-point localization, and needle insertion. The system performs vascular puncture robotically and incorporates a preloaded guidewire to facilitate the Seldinger technique of catheter insertion. Results from tissue-mimicking phantom and porcine studies under normotensive and hypotensive conditions provide evidence of the technique's robustness, with key performance metrics in a live porcine model including: a mean time to acquire femoral vein insertion point of 53 ± 36 s (5 users with varying experience, in 20 trials), a total time to insert catheter of 80 ± 30 s (1 user, in 6 trials), and a mean number of 1.1 (normotensive, 39 trials) and 1.3 (hypotensive, 55 trials) needle insertion attempts (1 user). These performance metrics in a porcine model are consistent with those for experienced medical providers performing central vascular access on humans in a hospital.

Transfer learning for fluence map prediction in adrenal stereotactic body radiation therapy

Objective:To design a deep transfer learning framework for modeling fluence map predictions for stereotactic body radiation therapy (SBRT) of adrenal cancer and similar sites that usually have a small number of cases.Approach:We developed a transfer learning framework for adrenal SBRT planning that leverages knowledge in a pancreas SBRT planning model. Treatment plans from the two sites had different dose prescriptions and beam settings but both prioritized gastrointestinal sparing. A base framework was first trained with 100 pancreas cases. This framework consists of two convolutional neural networks (CNN), which predict individual beam doses (BD-CNN) and fluence maps (FM-CNN) sequentially for 9-beam intensity-modulated radiation therapy (IMRT) plans. Forty-five adrenal plans were split into training/validation/test sets with the ratio of 20/10/15. The base BD-CNN was re-trained with transfer learning using 5/10/15/20 adrenal training cases to produce multiple candidate adrenal BD-CNN models. The base FM-CNN was directly used for adrenal cases. The deep learning (DL) plans were evaluated by several clinically relevant dosimetric endpoints, producing a percentage score relative to the clinical plans.Main results:Transfer learning significantly reduced the number of training cases and training time needed to train such a DL framework. The adrenal transfer learning model trained with 5/10/15/20 cases achieved validation plan scores of 85.4/91.2/90.7/89.4, suggesting that model performance saturated with 10 training cases. Meanwhile, a model using all 20 adrenal training cases without transfer learning only scored 80.5. For the final test set, the 5/10/15/20-case models achieved scores of 73.5/75.3/78.9/83.3.Significance:It is feasible to use deep transfer learning to train an IMRT fluence prediction framework. This technique could adapt to different dose prescriptions and beam configurations. This framework potentially enables DL modeling for clinical sites that have a limited dataset, either due to few cases or due to rapid technology evolution.

Automated segmentation of epidermis in high-frequency ultrasound of pathological skin using a cascade of DeepLab v3+ networks and fuzzy connectedness

This study proposes a novel, fully automated framework for epidermal layer segmentation in different skin diseases based on 75 MHz high-frequency ultrasound (HFUS) image data. A robust epidermis segmentation is a vital first step to detect changes in thickness, shape, and intensity and therefore support diagnosis and treatment monitoring in inflammatory and neoplastic skin lesions. Our framework links deep learning and fuzzy connectedness for image analysis. It consists of a cascade of two DeepLab v3+ models with a ResNet-50 backbone and a fuzzy connectedness analysis module for fine segmentation. Both deep models are pre-trained on the ImageNet dataset and subjected to transfer learning using our HFUS database of 580 images with atopic dermatitis, psoriasis and non-melanocytic skin tumors. The first deep model is used to detect the appropriate region of interest, while the second stands for the main segmentation procedure. We use the softmax layer of the latter twofold to prepare the input data for fuzzy connectedness analysis: as a reservoir of seed points and a direct contribution to the input image. In the experiments, we analyze different configurations of the framework, including region of interest detection, deep model backbones and training loss functions, or fuzzy connectedness analysis with parameter settings. We also use the Dice index and epidermis thickness to compare our results to state-of-the-art approaches. The Dice index of 0.919 yielded by our model over the entire dataset (and exceeding 0.93 in inflammatory diseases) proves its superiority over the other methods.

Factors determining generalization in deep learning models for scoring COVID-CT images

The COVID-19 pandemic has inspired unprecedented data collection and computer vision modelling efforts worldwide, focused on the diagnosis of COVID-19 from medical images. However, these models have found limited, if any, clinical application due in part to unproven generalization to data sets beyond their source training corpus. This study investigates the generalizability of deep learning models using publicly available COVID-19 Computed Tomography data through cross dataset validation. The predictive ability of these models for COVID-19 severity is assessed using an independent dataset that is stratified for COVID-19 lung involvement. Each inter-dataset study is performed using histogram equalization, and contrast limited adaptive histogram equalization with and without a learning Gabor filter. We show that under certain conditions, deep learning models can generalize well to an external dataset with F1 scores up to 86%. The best performing model shows predictive accuracy of between 75% and 96% for lung involvement scoring against an external expertly stratified dataset. From these results we identify key factors promoting deep learning generalization, being primarily the uniform acquisition of training images, and secondly diversity in CT slice position.

Boosting Advanced Nasopharyngeal Carcinoma Stage Prediction Using a Two-Stage Classification Framework Based on Deep Learning

Abstract

Nasopharyngeal carcinoma (NPC) is a popular malignant tumor of the head and neck which is endemic in the world, more than 75% of the NPC patients suffer from locoregionally advanced nasopharyngeal carcinoma (LA-NPC). The survival quality of these patients depends on the reliable prediction of NPC stages III and IVa. In this paper, we propose a two-stage framework to produce the classification probabilities for predicting NPC stages III and IVa. The preprocessing of MR images enhance the quality of images for further analysis. In stage one transfer learning is used to improve the classification effectiveness and the efficiency of CNN models training with limited images. Then in stage two the output of these models are aggregates using soft voting to boost the final prediction. The experimental results show the preprocessing is quite effective, the performance of transfer learning models perform better than the basic CNN model, and our ensemble model outperforms the single model as well as traditional methods, including the TNM staging system and the Radiomics method. Finally, the prediction accuracy boosted by the framework is, respectively, 0.81, indicating that our method achieves the SOTA effectiveness for LA-NPC stage prediction. In addition, the heatmaps generated with Class Activation Map technique illustrate the interpretability of the CNN models, and show their capability of assisting clinicians in medical diagnosis and follow-up treatment by producing discriminative regions related to NPC in the MR images.

Graphic Abstract

An integrated AI model to improve diagnostic accuracy of ultrasound and output known risk features in suspicious thyroid nodules

OBJECTIVES

From the viewpoint of ultrasound (US) physicians, an ideal thyroid US computer-assisted diagnostic (CAD) system for thyroid cancer should perform well in suspicious thyroid nodules with atypical risk features and be able to output explainable results. This study aims to develop an explainable US CAD model for suspicious thyroid nodules.

METHODS

A total of 2992 solid or almost-solid thyroid nodules were analyzed retrospectively. All nodules had pathological results (1070 malignancies and 1992 benignities) confirmed by ultrasound-guided fine-needle aspiration cytology and histopathology after thyroidectomy. A deep learning model (ResNet50) and a multiple risk features learning ensemble model (XGBoost) were used to train the US images of 2794 thyroid nodules. Then, an integrated AI model was generated by combining both models. The diagnostic accuracies of the three AI models (ResNet50, XGBoost, and the integrated model) were predicted in a testing set including 198 thyroid nodules and compared to the diagnostic efficacy of five ultrasonographers.

RESULTS

The accuracy of the integrated model was 76.77%, while the mean accuracy of the ultrasonographers was 68.38%. Of the risk features, microcalcifications showed the highest contribution to the diagnosis of malignant nodules.

CONCLUSIONS

The integrated AI model in our study can improve the diagnostic accuracy of suspicious thyroid nodules and output the known risk features simultaneously, thus aiding in training young ultrasonographers by linking the explainable results to their clinical experience and advancing the acceptance of AI diagnosis for thyroid cancer in clinical practice.

KEY POINTS

• We developed an artificial intelligence (AI) diagnosis model based on both deep learning and multiple risk feature ensemble learning methods. • The AI diagnosis model showed higher diagnostic accuracy for suspicious thyroid nodules than ultrasonographers. • The AI diagnosis model showed partial explainability by outputting the known risk features, thus aiding young ultrasonic doctors in increasing the diagnostic level for thyroid cancer.

Deep Mining Generation of Lung Cancer Malignancy Models from Chest X-ray Images

Lung cancer is the leading cause of cancer death and morbidity worldwide. Many studies have shown machine learning models to be effective in detecting lung nodules from chest X-ray images. However, these techniques have yet to be embraced by the medical community due to several practical, ethical, and regulatory constraints stemming from the “black-box” nature of deep learning models. Additionally, most lung nodules visible on chest X-rays are benign; therefore, the narrow task of computer vision-based lung nodule detection cannot be equated to automated lung cancer detection. Addressing both concerns, this study introduces a novel hybrid deep learning and decision tree-based computer vision model, which presents lung cancer malignancy predictions as interpretable decision trees. The deep learning component of this process is trained using a large publicly available dataset on pathological biomarkers associated with lung cancer. These models are then used to inference biomarker scores for chest X-ray images from two independent data sets, for which malignancy metadata is available. Next, multi-variate predictive models were mined by fitting shallow decision trees to the malignancy stratified datasets and interrogating a range of metrics to determine the best model. The best decision tree model achieved sensitivity and specificity of 86.7% and 80.0%, respectively, with a positive predictive value of 92.9%. Decision trees mined using this method may be considered as a starting point for refinement into clinically useful multi-variate lung cancer malignancy models for implementation as a workflow augmentation tool to improve the efficiency of human radiologists.