Convex Representations Using Deep Archetypal Analysis for Predicting Glaucoma

Application of artificial intelligence in glaucoma care: An updated review

The application of artificial intelligence (AI) in ophthalmology has been increasingly explored in the past decade. Numerous studies have shown promising results supporting the utility of AI to improve the management of ophthalmic diseases, and glaucoma is of no exception. Glaucoma is an irreversible vision condition with insidious onset, complex pathophysiology, and chronic treatment. Since there remain various challenges in the clinical management of glaucoma, the potential role of AI in facilitating glaucoma care has garnered significant attention. In this study, we reviewed the relevant literature published in recent years that investigated the application of AI in glaucoma management. The main aspects of AI applications that will be discussed include glaucoma risk prediction, glaucoma detection and diagnosis, visual field estimation and pattern analysis, glaucoma progression detection, and other applications.

An Artificial Intelligence Enabled System for Retinal Nerve Fiber Layer Thickness Damage Severity Staging

Purpose

To develop an objective glaucoma damage severity classification system based on OCT-derived retinal nerve fiber layer (RNFL) thickness measurements.

Design

Algorithm development for RNFL damage severity classification based on multicenter OCT data.

Subjects and Participants

A total of 6561 circumpapillary RNFL profiles from 2269 eyes of 1171 subjects to develop models, and 2505 RNFL profiles from 1099 eyes of 900 subjects to validate models.

Methods

We developed an unsupervised k-means model to identify clusters of eyes with similar RNFL thickness profiles. We annotated the clusters based on their respective global RNFL thickness. We computed the optimal global RNFL thickness thresholds that discriminated different severity levels based on Bayes' minimum error principle. We validated the proposed pipeline based on an independent validation dataset with 2505 RNFL profiles from 1099 eyes of 900 subjects.

Main Outcome Measures

Accuracy, area under the receiver operating characteristic curve, and confusion matrix.

Results

The k-means clustering discovered 4 clusters with 1382, 1613, 1727, and 1839 samples with mean (standard deviation) global RNFL thickness of 58.3 (8.9) μm, 78.9 (6.7) μm, 87.7 (8.2) μm, and 101.5 (7.9) μm. The Bayes' minimum error classifier identified optimal global RNFL values of > 95 μ m , 86 to 95 μ m , 70 to 85 μ m , and < 70 μ m for discriminating normal eyes and eyes at the early, moderate, and advanced stages of RNFL thickness loss, respectively. About 4% of normal eyes and 98% of eyes with advanced RNFL loss had either global, or ≥ 1 quadrant, RNFL thickness outside of normal limits provided by the OCT instrument.

Conclusions

Unsupervised machine learning discovered that the optimal RNFL thresholds for separating normal eyes and eyes with early, moderate, and advanced RNFL loss were 95 μ m , 85 μm, and 70 μ m , respectively. This RNFL loss classification system is unbiased as there was no preassumption or human expert intervention in the development process. Additionally, it is objective, easy to use, and consistent, which may augment glaucoma research and day-to-day clinical practice.

Financial Disclosures

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

Artificial intelligence in glaucoma: opportunities, challenges, and future directions

Artificial intelligence (AI) has shown excellent diagnostic performance in detecting various complex problems related to many areas of healthcare including ophthalmology. AI diagnostic systems developed from fundus images have become state-of-the-art tools in diagnosing retinal conditions and glaucoma as well as other ocular diseases. However, designing and implementing AI models using large imaging data is challenging. In this study, we review different machine learning (ML) and deep learning (DL) techniques applied to multiple modalities of retinal data, such as fundus images and visual fields for glaucoma detection, progression assessment, staging and so on. We summarize findings and provide several taxonomies to help the reader understand the evolution of conventional and emerging AI models in glaucoma. We discuss opportunities and challenges facing AI application in glaucoma and highlight some key themes from the existing literature that may help to explore future studies. Our goal in this systematic review is to help readers and researchers to understand critical aspects of AI related to glaucoma as well as determine the necessary steps and requirements for the successful development of AI models in glaucoma.

Archetypal analysis of longitudinal visual fields for idiopathic intracranial hypertension patients presenting in a clinic setting

We previously applied archetypal analysis (AA) using visual fields (VF) from the Idiopathic Intracranial Hypertension Treatment Trial (IIHTT) to derive a model, which quantified patterns (or archetypes [ATs] of VF loss), anticipated recovery, and identified residual VF deficits. We hypothesized that AA could produce similar results using IIH VFs collected in clinical practice. We applied AA to 803 VFs from 235 eyes with IIH from an outpatient neuro-ophthalmology clinic and created a clinic-derived model of ATs, with the relative weight (RW) and average total deviation (TD) for each AT. We also created a combined-derived model from an input dataset containing the clinic VFs and 2862 VFs from the IIHTT. We used both models to decompose clinic VF into ATs of varying percent weight (PW), correlated presentation AT PW with mean deviation (MD), and evaluated final visit VFs considered "normal" by MD ≥ -2.00 dB for residual abnormal ATs. The 14-AT clinic-derived and combined-derived models revealed similar patterns of VF loss previously identified in the IIHTT model. AT1 (a normal pattern) was most prevalent in both models (RW = 51.8% for clinic-derived; 35.4% for combined-derived). Presentation AT1 PW correlated with final visit MD (r = 0.82, p < 0.001 for the clinic-derived model; r = 0.59, p < 0.001 for the combined-derived model). Both models showed ATs with similar patterns of regional VF loss. The most common patterns of VF loss in "normal" final visit VFs using each model were clinic-derived AT2 (mild global depression with enlarged blind spot; 44/125 VFs; 34%) and combined-derived AT2 (near-normal; 93/149 VFs; 62%). AA provides quantitative values for IIH-related patterns of VF loss that can be used to monitor VF changes in a clinic setting. Presentation AT1 PW is associated with the degree of VF recovery. AA identifies residual VF deficits not otherwise indicated by MD.

Applications of Artificial Intelligence and Deep Learning in Glaucoma

Diagnosis and detection of progression of glaucoma remains challenging. Artificial intelligence-based tools have the potential to improve and standardize the assessment of glaucoma but development of these algorithms is difficult given the multimodal and variable nature of the diagnosis. Currently, most algorithms are focused on a single imaging modality, specifically screening and diagnosis based on fundus photos or optical coherence tomography images. Use of anterior segment optical coherence tomography and goniophotographs is limited. The majority of algorithms designed for disease progression prediction are based on visual fields. No studies in our literature search assessed the use of artificial intelligence for treatment response prediction and no studies conducted prospective testing of their algorithms. Additional challenges to the development of artificial intelligence-based tools include scarcity of data and a lack of consensus in diagnostic criteria. Although research in the use of artificial intelligence for glaucoma is promising, additional work is needed to develop clinically usable tools.

Clinical Applications of Artificial Intelligence in Glaucoma

Ophthalmology is one of the major imaging-intensive fields of medicine and thus has potential for extensive applications of artificial intelligence (AI) to advance diagnosis, drug efficacy, and other treatment-related aspects of ocular disease. AI has made impressive progress in ophthalmology within the past few years and two autonomous AI-enabled systems have received US regulatory approvals for autonomously screening for mid-level or advanced diabetic retinopathy and macular edema. While no autonomous AI-enabled system for glaucoma screening has yet received US regulatory approval, numerous assistive AI-enabled software tools are already employed in commercialized instruments for quantifying retinal images and visual fields to augment glaucoma research and clinical practice. In this literature review (non-systematic), we provide an overview of AI applications in glaucoma, and highlight some limitations and considerations for AI integration and adoption into clinical practice.

Machine-Identified Patterns of Visual Field Loss and an Association with Rapid Progression in the Ocular Hypertension Treatment Study

PURPOSE

To identify patterns of visual field (VF) loss based on unsupervised machine learning and to identify patterns that are associated with rapid progression.

DESIGN

Cross-sectional and longitudinal study.

PARTICIPANTS

A total of 2231 abnormal VFs from 205 eyes of 176 Ocular Hypertension Treatment Study (OHTS) participants followed over approximately 16 years.

METHODS

Visual fields were assessed by an unsupervised deep archetypal analysis algorithm and an OHTS-certified VF reader to identify prevalent patterns of VF loss. Machine-identified patterns of glaucoma damage were compared against those patterns previously identified (expert-identified) in the OHTS in 2003. Based on the longitudinal VFs of each eye, VF loss patterns that were strongly associated with rapid glaucoma progression were identified.

MAIN OUTCOME MEASURES

Machine-expert correspondence and type of patterns of VF loss associated with rapid progression.

RESULTS

The average VF mean deviation (MD) at conversion to glaucoma was -2.7 decibels (dB) (standard deviation [SD] = 2.4 dB), whereas the average MD of the eyes at the last visit was -5.2 dB (SD = 5.5 dB). Fifty out of 205 eyes had MD rate of -1 dB/year or worse and were considered rapid progressors. Eighteen machine-identified patterns of VF loss were compared with expert-identified patterns, in which 13 patterns of VF loss were similar. The most prevalent expert-identified patterns included partial arcuate, paracentral, and nasal step defects, and the most prevalent machine-identified patterns included temporal wedge, partial arcuate, nasal step, and paracentral VF defects. One of the machine-identified patterns of VF loss predicted future rapid VF progression after adjustment for age, sex, and initial MD.

CONCLUSIONS

An automated machine learning system can identify patterns of VF loss and could provide objective and reproducible nomenclature for characterizing early signs of visual defects and rapid progression in patients with glaucoma.

Archetypal analysis of visual fields in optic neuritis reveals functional biomarkers associated with outcome and treatment response

BACKGROUND AND OBJECTIVES

Archetypal analysis (AA), a form of unsupervised machine learning, can identify quantifiable visual field (VF) patterns seen in optic neuritis (ON), known as archetypes (ATs). We hypothesized that AT weight changes over time would reflect the course of recovery and the effects of therapy in ON. We explored whether baseline AT weights would be associated with VF status at the clinical trial outcome and if ATs would indicate residual VF defects in eyes with mean deviation (MD) ≥ -2.00 at six months.

METHODS

We used a published 16-AT model derived from 3892 Optic Neuritis Treatment Trial VFs (456 eyes) for all analyses. We measured AT weight changes over the six-month study period and used asymptotic regression to analyze the rate of change. We compared AT weights at six months between treatment groups. We evaluated associations between baseline AT weight thresholds and VF outcome or treatment effect. We calculated residual AT weights in eyes with MD ≥ -2.00 dB at six months.

RESULTS

Over six months, AT1 (a normal VF pattern) demonstrated the greatest median weight change, increasing from 0.00% (IQR 0.00-0.00%) at baseline to 60.0% (IQR 38.3-70.8%) at six months (p < 0.001). At outcome, the intravenous methylprednisolone (IVMP) group had the highest median AT1 weight (IVMP: 63.3%, IQR 51.3-72.8%; placebo: 56.2%, IQR 35.1-71.6%; prednisone 58.3%, IQR 35.1-71.6%; p = 0.019). Eyes with AT1 weight ≥ 19% at baseline had superior median MD values (-0.91 vs. -2.07 dB, p < 0.001) and AT1 weights (70.8% vs. 57.8% p < 0.001) at six months. Only eyes with AT1 weight < 19% at baseline showed a treatment benefit for IVMP, with a higher six-month median AT1 weight compared to placebo (p = 0.015) and prednisone (p = 0.016), and a higher median MD compared to placebo (p = 0.027). At six months, 182 (80.2%) VFs with MD ≥ -2.00 had at least one abnormal AT.

DISCUSSION

Changes in quantifiable, archetypal patterns of VF loss reflect recovery in ON. Machine learning analysis of the VFs in optic neuritis reveals associations with response to therapy and VF outcome, and uncovers residual deficits, not readily seen with standard evaluations.

A Machine Learning-Based Predictive Model of Epidermal Growth Factor Mutations in Lung Adenocarcinomas

Data from 758 patients with lung adenocarcinoma were retrospectively collected. All patients had undergone computed tomography imaging and EGFR gene testing. Radiomic features were extracted using the medical imaging tool 3D-Slicer and were combined with the clinical features to build a machine learning prediction model. The high-dimensional feature set was screened for optimal feature subsets using principal component analysis (PCA) and the least absolute shrinkage and selection operator (LASSO). Model prediction of EGFR mutation status in the validation group was evaluated using multiple classifiers. We showed that six clinical features and 622 radiomic features were initially collected. Thirty-one radiomic features with non-zero correlation coefficients were obtained by LASSO regression, and 24 features correlated with label values were obtained by PCA. The shared radiomic features determined by these two methods were selected and combined with the clinical features of the respective patient to form a subset of features related to EGFR mutations. The full dataset was partitioned into training and test sets at a ratio of 7:3 using 10-fold cross-validation. The area under the curve (AUC) of the four classifiers with cross-validations was: (1) K-nearest neighbor (AUCmean = 0.83, Acc = 81%); (2) random forest (AUCmean = 0.91, Acc = 83%); (3) LGBM (AUCmean = 0.94, Acc = 88%); and (4) support vector machine (AUCmean = 0.79, Acc = 83%). In summary, the subset of radiographic and clinical features selected by feature engineering effectively predicted the EGFR mutation status of this NSCLC patient cohort.

An Objective and Easy-to-Use Glaucoma Functional Severity Staging System Based on Artificial Intelligence

OBJECTIVE

The objective of this study was to develop an objective and easy-to-use glaucoma staging system based on visual fields (VFs).

SUBJECTS AND PARTICIPANTS

A total of 13,231 VFs from 8077 subjects were used to develop models and 8024 VFs from 4445 subjects were used to validate models.

METHODS

We developed an unsupervised machine learning model to identify clusters with similar VF values. We annotated the clusters based on their respective mean deviation (MD). We computed optimal MD thresholds that discriminate clusters with the highest accuracy based on Bayes minimum error principle. We evaluated the accuracy of the staging system and validated findings based on an independent validation dataset.

RESULTS

The unsupervised k -means algorithm discovered 4 clusters with 6784, 4034, 1541, and 872 VFs and average MDs of 0.0 dB (±1.4: SD), -4.8 dB (±1.9), -12.2 dB (±2.9), and -23.0 dB (±3.8), respectively. The supervised Bayes minimum error classifier identified optimal MD thresholds of -2.2, -8.0, and -17.3 dB for discriminating normal eyes and eyes at the early, moderate, and advanced stages of glaucoma. The accuracy of the glaucoma staging system was 94%, based on identified MD thresholds with respect to the initial k -means clusters.

CONCLUSIONS

We discovered that 4 severity levels based on MD thresholds of -2.2, -8.0, and -17.3 dB, provides the optimal number of severity stages based on unsupervised and supervised machine learning. This glaucoma staging system is unbiased, objective, easy-to-use, and consistent, which makes it highly suitable for use in glaucoma research and for day-to-day clinical practice.

A deep-learning system predicts glaucoma incidence and progression using retinal photographs

BackgroundDeep learning has been widely used for glaucoma diagnosis. However, there is no clinically validated algorithm for glaucoma incidence and progression prediction. This study aims to develop a clinically feasible deep-learning system for predicting and stratifying the risk of glaucoma onset and progression based on color fundus photographs (CFPs), with clinical validation of performance in external population cohorts.MethodsWe established data sets of CFPs and visual fields collected from longitudinal cohorts. The mean follow-up duration was 3 to 5 years across the data sets. Artificial intelligence (AI) models were developed to predict future glaucoma incidence and progression based on the CFPs of 17,497 eyes in 9346 patients. The area under the receiver operating characteristic (AUROC) curve, sensitivity, and specificity of the AI models were calculated with reference to the labels provided by experienced ophthalmologists. Incidence and progression of glaucoma were determined based on longitudinal CFP images or visual fields, respectively.ResultsThe AI model to predict glaucoma incidence achieved an AUROC of 0.90 (0.81-0.99) in the validation set and demonstrated good generalizability, with AUROCs of 0.89 (0.83-0.95) and 0.88 (0.79-0.97) in external test sets 1 and 2, respectively. The AI model to predict glaucoma progression achieved an AUROC of 0.91 (0.88-0.94) in the validation set, and also demonstrated outstanding predictive performance with AUROCs of 0.87 (0.81-0.92) and 0.88 (0.83-0.94) in external test sets 1 and 2, respectively.ConclusionOur study demonstrates the feasibility of deep-learning algorithms in the early detection and prediction of glaucoma progression.FUNDINGNational Natural Science Foundation of China (NSFC); the High-level Hospital Construction Project, Zhongshan Ophthalmic Center, Sun Yat-sen University; the Science and Technology Program of Guangzhou, China (2021), the Science and Technology Development Fund (FDCT) of Macau, and FDCT-NSFC.

Unsupervised Machine Learning Identifies Quantifiable Patterns of Visual Field Loss in Idiopathic Intracranial Hypertension

Purpose

Archetypal analysis, a form of unsupervised machine learning, identifies archetypal patterns within a visual field (VF) dataset such that any VF is described as a weighted sum of its archetypes (ATs) and has been used to quantify VF defects in glaucoma. We applied archetypal analysis to VFs affected by nonglaucomatous optic neuropathy caused by idiopathic intracranial hypertension (IIH).

Methods

We created an AT model from 2862 VFs prospectively collected from 330 eyes in the IIH Treatment Trial (IIHTT). We compared baseline IIH AT patterns with their descriptive VF classifications from the IIHTT.

Results

The optimum IIH AT model yielded 14 ATs resembling VF patterns reported in the IIHTT. Baseline VFs contained four or fewer meaningful ATs in 147 (89%) of study eyes. AT2 (mild general VF depression pattern) demonstrated the greatest number of study eyes with meaningful AT weight at baseline (n = 114), followed by AT1 (n = 91). Other ATs captured patterns of blind spot enlargement, hemianopia, arcuate, nasal defects, and more nonspecific patterns of general VF depression. Of all ATs, AT1 (normal pattern) had the strongest correlation with mean deviation (r = 0.69, P < 0.001). For 65 of the 93 VFs with a dominant AT, this AT matched the expert classification.

Conclusions

Archetypal analysis identifies quantifiable, archetypal VF defects that resemble those commonly seen in IIH.

Translational Relevance

Archetypal analysis provides a quantitative, objective method of measuring and monitoring disease-specific regional VF defects in IIH.