Artificial Intelligence Algorithms for Analysis of Geographic Atrophy: A Review and Evaluation

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

Purpose

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

Design

Evaluation of artificial intelligence (AI) algorithms.

Subjects

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

Methods

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

Main Outcome Measures

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

Results

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

Conclusions

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

Financial Disclosures

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

Disease activity and therapeutic response to pegcetacoplan for geographic atrophy identified by deep learning-based analysis of OCT

PURPOSE

To quantify morphological changes of the photoreceptors (PR) and retinal pigment epithelium (RPE) layers under pegcetacoplan therapy in geographic atrophy (GA) using deep learning-based analysis of optical coherence tomography (OCT) images.

DESIGN

Post-hoc longitudinal image analysis SUBJECTS: Patients with GA due to age-related macular degeneration from two prospective randomized phase III clinical trials (OAKS and DERBY) METHODS: Deep learning-based segmentation of RPE loss and PR degeneration, defined as loss of the ellipsoid zone (EZ) layer on OCT, over 24 months on SD-OCT images MAIN OUTCOME MEASURES: Change in the mean area of RPE loss and EZ loss over time in the pooled sham arms and the monthly (PM)/every other month (PEOM) treatment arms RESULTS: 897 eyes of 897 patients were included. There was a therapeutic reduction of RPE loss growth by 22%/20% in OAKS and 27%/21% in DERBY for PM/PEOM compared to sham, respectively, at 24 months. The reduction on the EZ level was significantly higher with 53%/46% in OAKS and 47%/46% in DERBY for PM/PEOM compared to sham at 24 months. The baseline EZ-RPE difference had an impact on disease activity and therapeutic response. The therapeutic benefit for RPE loss growth increased with larger EZ-RPE difference quartiles from 21.9%, 23.1%, 23.9% to 33.6% for PM vs. sham (all p<0.01) and from 13.6% (p=0.11), 23.8%, 23.8% to 20.0% for PEOM vs. sham (p<0.01) in quartiles 1,2,3 and 4, respectively, at 24 months. Regarding EZ layer maintenance, the therapeutic reduction of loss increased from 14.8% (p=0.09), 33.3%, 46.6% to 77.8% (p<0.0001) between PM and sham and from 15.9% (p=0.08), 33.8%, 52.0% to 64.9% (p<0.0001) between PEOM and sham for quartiles 1-4 at 24 months.

CONCLUSION

OCT-based AI analysis objectively identifies and quantifies PR and RPE degeneration in GA. Reductions in further PR degeneration consistent with EZ loss on OCT are even higher than the effect on RPE loss in phase 3 trials of pegcetacoplan treatment. The EZ-RPE difference has a strong impact on disease progression and therapeutic response. Identification of patients with higher EZ-RPE loss difference may become an important criterion for the management of GA secondary to AMD.

Combination of optical coherence tomography-derived shape and texture features are associated with development of sub-foveal geographic atrophy in dry AMD

Geographic atrophy (GA) is an advanced form of dry age-related macular degeneration (AMD) that leads to progressive and irreversible vision loss. Identifying patients with greatest risk of GA progression is important for targeted utilization of emerging therapies. This study aimed to comprehensively evaluate the role of shape-based fractal dimension features ( F fd ) of sub-retinal pigment epithelium (sub-RPE) compartment and texture-based radiomics features ( F t ) of Ellipsoid Zone (EZ)-RPE and sub-RPE compartments for risk stratification for subfoveal GA (sfGA) progression. This was a retrospective study of 137 dry AMD subjects with a 5-year follow-up. Based on sfGA status at year 5, eyes were categorized as Progressors and Non-progressors. A total of 15 shape-based F fd of sub-RPE surface and 494 F t from each of sub-RPE and EZ-RPE compartments were extracted from baseline spectral domain-optical coherence tomography scans. The top nine features were identified from F fd and F t feature pool separately using minimum Redundancy maximum Relevance feature selection and used to train a Random Forest (RF) classifier independently using three-fold cross validation on the training set ( S t , N = 90) to distinguish between sfGA Progressors and Non-progressors. Combined F fd and F t was also evaluated in predicting risk of sfGA progression. The RF classifier yielded AUC of 0.85, 0.79 and 0.89 on independent test set ( S v , N = 47) using F fd , F t , and their combination, respectively. Using combined F fd and F t , the improvement in AUC was statistically significant on S v with p-values of 0.032 and 0.04 compared to using only F fd and only F t , respectively. Combined F fd and F t appears to identify high-risk patients. Our results show that FD and texture features could be potentially used for predicting risk of sfGA progression and future therapeutic response.

Oral Antioxidant and Lutein/Zeaxanthin Supplements Slow Geographic Atrophy Progression to the Fovea in Age-Related Macular Degeneration

PURPOSE

The only treatments approved to slow geographic atrophy (GA) progression in age-related macular degeneration (AMD) require frequent intraocular injection and suffer from modest efficacy, important risks, and high costs. The purpose of this study was to determine whether oral supplements slow GA progression in AMD.

DESIGN

Post hoc analysis of the Age-Related Eye Diseases Study (AREDS) and AREDS2, multi-center randomized placebo-controlled trials of oral micronutrient supplementation, each with 2x2 factorial design.

PARTICIPANTS

392 eyes (318 participants) with GA in AREDS; 1210 eyes (891 participants) with GA in AREDS2.

METHODS

AREDS participants were randomly assigned to oral antioxidants (500 mg vitamin C; 400 IU vitamin E; 15 mg β-carotene); 80 mg zinc; combination; or placebo. AREDS2 participants were randomly assigned to 10 mg lutein/2 mg zeaxanthin; 350 mg docosahexaenoic acid/650 mg eicosapentaenoic acid; combination; or placebo. Consenting AREDS2 participants were also randomly assigned to alternative AREDS formulations: original; no beta-carotene; 25 mg zinc instead of 80 mg; both.

MAIN OUTCOME MEASURES

(1) Change in GA proximity to central macula over time, and (2) change in square root GA area over time, each measured from color fundus photographs at annual visits and analyzed by mixed-model regression according to randomized assignments.

RESULTS

In AREDS eyes with non-central GA (n=208), proximity-based progression towards the central macula was significantly slower with randomization to antioxidants versus none, at 50.7 μm/year (95% CI 38.0-63.4 μm/year) versus 72.9 μm/year (95% CI 61.3-84.5 μm/year; p=0.012), respectively. In AREDS2 eyes with non-central GA, in participants assigned to AREDS antioxidants without β-carotene (n=325 eyes), proximity-based progression was significantly slower with randomization to lutein/zeaxanthin versus none, at 80.1 μm/year (95% CI 60.9-99.3 μm/year) versus 114.4 μm/year (95% CI 96.2-132.7 μm/year; p=0.011), respectively. In AREDS eyes with any GA (n=392), area-based progression was not significantly different with randomization to antioxidants versus none (p=0.63). In AREDS2 eyes with any GA, in participants assigned to AREDS antioxidants without β-carotene (n=505 eyes), area-based progression was not significantly different with randomization to lutein/zeaxanthin versus none (p=0.64).

CONCLUSIONS

Oral micronutrient supplementation slowed GA progression towards the central macula, likely by augmenting the natural phenomenon of foveal sparing.

Quantifying Geographic Atrophy in Age-Related Macular Degeneration: A Comparative Analysis Across 12 Deep Learning Models

Purpose

AI algorithms have shown impressive performance in segmenting geographic atrophy (GA) from fundus autofluorescence (FAF) images. However, selection of artificial intelligence (AI) architecture is an important variable in model development. Here, we explore 12 distinct AI architecture combinations to determine the most effective approach for GA segmentation.

Methods

We investigated various AI architectures, each with distinct combinations of encoders and decoders. The architectures included three decoders-FPN (Feature Pyramid Network), UNet, and PSPNet (Pyramid Scene Parsing Network)-and serve as the foundation framework for segmentation task. Encoders including EfficientNet, ResNet (Residual Networks), VGG (Visual Geometry Group) and Mix Vision Transformer (mViT) have a role in extracting optimum latent features for accurate GA segmentation. Performance was measured through comparison of GA areas between human and AI predictions and Dice Coefficient (DC).

Results

The training dataset included 601 FAF images from AREDS2 study and validation included 156 FAF images from the GlaxoSmithKline study. The mean absolute difference between grader measured and AI predicted areas ranged from -0.08 (95% CI = -1.35, 1.19) to 0.73 mm2 (95% CI = -5.75,4.29) and DC between 0.884-0.993. The best-performing models were UNet and FPN frameworks with mViT, and the least-performing models were PSPNet framework.

Conclusions

The choice of AI architecture impacts GA segmentation performance. Vision transformers with FPN and UNet architectures demonstrate stronger suitability for this task compared to Convolutional Neural Network- and PSPNet-based models. Selecting an AI architecture must be tailored to the specific goals of the project, and developers should consider which architecture is ideal for their project.

Pegcetacoplan Treatment and Consensus Features of Geographic Atrophy Over 24 Months

Importance

Despite widespread availability and consensus on its advantages for detailed imaging of geographic atrophy (GA), spectral-domain optical coherence tomography (SD-OCT) might benefit from automated quantitative OCT analyses in GA diagnosis, monitoring, and reporting of its landmark clinical trials.

Objective

To analyze the association between pegcetacoplan and consensus GA SD-OCT end points.

Design, Setting, and Participants

This was a post hoc analysis of 11 614 SD-OCT volumes from 936 of the 1258 participants in 2 parallel phase 3 studies, the Study to Compare the Efficacy and Safety of Intravitreal APL-2 Therapy With Sham Injections in Patients With Geographic Atrophy (GA) Secondary to Age-Related Macular Degeneration (OAKS) and Study to Compare the Efficacy and Safety of Intravitreal APL-2 Therapy With Sham Injections in Patients With Geographic Atrophy (GA) Secondary to Age-Related Macular Degeneration (DERBY). OAKS and DERBY were 24-month, multicenter, randomized, double-masked, sham-controlled studies conducted from August 2018 to July 2020 among adults with GA with total area 2.5 to 17.5 mm2 on fundus autofluorescence imaging (if multifocal, at least 1 lesion ≥1.25 mm2). This analysis was conducted from September to December 2023.

Interventions

Study participants received pegcetacoplan, 15 mg per 0.1-mL intravitreal injection, monthly or every other month, or sham injection monthly or every other month.

Main Outcomes and Measures

The primary end point was the least squares mean change from baseline in area of retinal pigment epithelium and outer retinal atrophy in each of the 3 treatment arms (pegcetacoplan monthly, pegcetacoplan every other month, and pooled sham [sham monthly and sham every other month]) at 24 months. Feature-specific area analysis was conducted by Early Treatment Diabetic Retinopathy Study (ETDRS) regions of interest (ie, foveal, parafoveal, and perifoveal).

Results

Among 936 participants, the mean (SD) age was 78.5 (7.22) years, and 570 participants (60.9%) were female. Pegcetacoplan, but not sham treatment, was associated with reduced growth rates of SD-OCT biomarkers for GA for up to 24 months. Reductions vs sham in least squares mean (SE) change from baseline of retinal pigment epithelium and outer retinal atrophy area were detectable at every time point from 3 through 24 months (least squares mean difference vs pooled sham at month 24, pegcetacoplan monthly: -0.86 mm2; 95% CI, -1.15 to -0.57; P < .001; pegcetacoplan every other month: -0.69 mm2; 95% CI, -0.98 to -0.39; P < .001). This association was more pronounced with more frequent dosing (pegcetacoplan monthly vs pegcetacoplan every other month at month 24: -0.17 mm2; 95% CI, -0.43 to 0.08; P = .17). Stronger associations were observed in the parafoveal and perifoveal regions for both pegcetacoplan monthly and pegcetacoplan every other month.

Conclusions and Relevance

These findings offer additional insight into the potential effects of pegcetacoplan on the development of GA, including potential effects on the retinal pigment epithelium and photoreceptors.

Trial Registration

ClinicalTrials.gov Identifiers: NCT03525600 and NCT03525613.

Deep-learning automated quantification of longitudinal OCT scans demonstrates reduced RPE loss rate, preservation of intact macular area and predictive value of isolated photoreceptor degeneration in geographic atrophy patients receiving C3 inhibition treatment

OBJECTIVE

To evaluate the role of automated optical coherence tomography (OCT) segmentation, using a validated deep-learning model, for assessing the effect of C3 inhibition on the area of geographic atrophy (GA); the constituent features of GA on OCT (photoreceptor degeneration (PRD), retinal pigment epithelium (RPE) loss and hypertransmission); and the area of unaffected healthy macula.To identify OCT predictive biomarkers for GA growth.

METHODS

Post hoc analysis of the FILLY trial using a deep-learning model for spectral domain OCT (SD-OCT) autosegmentation. 246 patients were randomised 1:1:1 into pegcetacoplan monthly (PM), pegcetacoplan every other month (PEOM) and sham treatment (pooled) for 12 months of treatment and 6 months of therapy-free monitoring. Only participants with Heidelberg SD-OCT were included (n=197, single eye per participant).The primary efficacy endpoint was the square root transformed change in area of GA as complete RPE and outer retinal atrophy (cRORA) in each treatment arm at 12 months, with secondary endpoints including RPE loss, hypertransmission, PRD and intact macular area.

RESULTS

Eyes treated PM showed significantly slower mean change of cRORA progression at 12 and 18 months (0.151 and 0.277 mm, p=0.0039; 0.251 and 0.396 mm, p=0.039, respectively) and RPE loss (0.147 and 0.287 mm, p=0.0008; 0.242 and 0.410 mm, p=0.00809). PEOM showed significantly slower mean change of RPE loss compared with sham at 12 months (p=0.0313). Intact macular areas were preserved in PM compared with sham at 12 and 18 months (p=0.0095 and p=0.044). PRD in isolation and intact macula areas was predictive of reduced cRORA growth at 12 months (coefficient 0.0195, p=0.01 and 0.00752, p=0.02, respectively) CONCLUSION: The OCT evidence suggests that pegcetacoplan slows progression of cRORA overall and RPE loss specifically while protecting the remaining photoreceptors and slowing the progression of healthy retina to iRORA.

PREVALENCE AND PERSISTENCE OF HYPERTRANSMISSION DEFECTS OF VARIOUS SIZES IN EYES WITH INTERMEDIATE AGE-RELATED MACULAR DEGENERATION

PURPOSE

To determine the prevalence and rate of persistence over 2 years of various-sized hypertransmission defects (hyperTDs) in eyes with intermediate age-related macular degeneration.

METHODS

Retrospective analysis of optical coherence tomography data from consecutive intermediate age-related macular degeneration patients. Choroidal en face optical coherence tomography images were evaluated for the presence and number of hyperTDs of three different sizes based on greatest linear dimension (small, 63-124 µ m; medium, 125-249 µ m; large, ≥250 µ m) at baseline and at the 2-year follow-up. Interreader agreement was determined by Gwet's agreement coefficient. Disagreements between graders were resolved by the senior investigator to yield a single consensus for all cases.

RESULTS

From 273 intermediate age-related macular degeneration eyes (247 patients), 72 and 76 hyperTD lesions were independently identified by two graders at baseline and overall agreement coefficient was 0.89 (95% CI, 0.86-0.93). After adjudication by the senior grader, the final consensus yielded 78 hyperTD lesions from 46 eyes (16.8%) of 42 patients (17.0%) in this study cohort. Among eyes with follow-up optical coherence tomography, 32 of 45 hyperTD lesions (71.1%) persisted. The rates of persistence were 100.0%, 72.7%, and 53.3% in large, medium, and small hyperTD sizes, respectively.

CONCLUSION

HyperTDs were present in a significant proportion of intermediate age-related macular degeneration eyes. Acceptable interreader agreement was demonstrated in identifying hyperTD. Larger hyperTD lesions were more likely to persist over 2 years.

Explainable artificial intelligence model for the detection of geographic atrophy using colour retinal photographs

OBJECTIVE

To develop and validate an explainable artificial intelligence (AI) model for detecting geographic atrophy (GA) via colour retinal photographs.

METHODS AND ANALYSIS

We conducted a prospective study where colour fundus images were collected from healthy individuals and patients with retinal diseases using an automated imaging system. All images were categorised into three classes: healthy, GA and other retinal diseases, by two experienced retinologists. Simultaneously, an explainable learning model using class activation mapping techniques categorised each image into one of the three classes. The AI system's performance was then compared with manual evaluations.

RESULTS

A total of 540 colour retinal photographs were collected. Data was divided such that 300 images from each class trained the AI model, 120 for validation and 120 for performance testing. In distinguishing between GA and healthy eyes, the model demonstrated a sensitivity of 100%, specificity of 97.5% and an overall diagnostic accuracy of 98.4%. Performance metrics like area under the receiver operating characteristic (AUC-ROC, 0.988) and the precision-recall (AUC-PR, 0.952) curves reinforced the model's robust achievement. When differentiating GA from other retinal conditions, the model preserved a diagnostic accuracy of 96.8%, a precision of 90.9% and a recall of 100%, leading to an F1-score of 0.952. The AUC-ROC and AUC-PR scores were 0.975 and 0.909, respectively.

CONCLUSIONS

Our explainable AI model exhibits excellent performance in detecting GA using colour retinal images. With its high sensitivity, specificity and overall diagnostic accuracy, the AI model stands as a powerful tool for the automated diagnosis of GA.

Deep-GA-Net for Accurate and Explainable Detection of Geographic Atrophy on OCT Scans

Objective

To propose Deep-GA-Net, a 3-dimensional (3D) deep learning network with 3D attention layer, for the detection of geographic atrophy (GA) on spectral domain OCT (SD-OCT) scans, explain its decision making, and compare it with existing methods.

Design

Deep learning model development.

Participants

Three hundred eleven participants from the Age-Related Eye Disease Study 2 Ancillary SD-OCT Study.

Methods

A dataset of 1284 SD-OCT scans from 311 participants was used to develop Deep-GA-Net. Cross-validation was used to evaluate Deep-GA-Net, where each testing set contained no participant from the corresponding training set. En face heatmaps and important regions at the B-scan level were used to visualize the outputs of Deep-GA-Net, and 3 ophthalmologists graded the presence or absence of GA in them to assess the explainability (i.e., understandability and interpretability) of its detections.

Main Outcome Measures

Accuracy, area under receiver operating characteristic curve (AUC), area under precision-recall curve (APR).

Results

Compared with other networks, Deep-GA-Net achieved the best metrics, with accuracy of 0.93, AUC of 0.94, and APR of 0.91, and received the best gradings of 0.98 and 0.68 on the en face heatmap and B-scan grading tasks, respectively.

Conclusions

Deep-GA-Net was able to detect GA accurately from SD-OCT scans. The visualizations of Deep-GA-Net were more explainable, as suggested by 3 ophthalmologists. The code and pretrained models are publicly available at https://github.com/ncbi/Deep-GA-Net.

Financial Disclosures

The author(s) have no proprietary or commercial interest in any materials discussed in this article.

Novel Fractal-Based Sub-RPE Compartment OCT Radiomics Biomarkers Are Associated With Subfoveal Geographic Atrophy in Dry AMD

OBJECTIVE

The purpose of this study was to quantitatively characterize the shape of the sub-retinal pigment epithelium (sub-RPE, i.e., space bounded by RPE and Bruch's membrane) compartment on SD-OCT using fractal dimension (FD) features and evaluate their impact on risk of subfoveal geographic atrophy (sfGA) progression.

METHODS

This was an IRB-approved retrospective study of 137 subjects with dry age-related macular degeneration (AMD) with subfoveal GA. Based on sfGA status at year five, eyes were categorized as "Progressors" and "Non-progressors". FD analysis allows quantification of the degree of shape complexity and architectural disorder associated with a structure. To characterize the structural irregularities along the sub-RPE surface between the two groups of patients, a total of 15 shape descriptors of FD were extracted from the sub-RPE compartment of baseline OCT scans. The top four features were identified using minimum Redundancy maximum Relevance (mRmR) feature selection method and evaluated with Random Forest (RF) classifier using three-fold cross validation from the training set (N = 90). Classifier performance was subsequently validated on the independent test set (N = 47).

RESULTS

Using the top four FD features, a RF classifier yielded an AUC of 0.85 on the independent test set. Mean fractal entropy (p-value = 4.8e-05) was identified as the most significant biomarker; higher values of entropy being associated with greater shape disorder and risk for sfGA progression.

CONCLUSIONS

FD assessment holds promise for identifying high-risk eyes for GA progression.

SIGNIFICANCE

With further validation, FD features could be potentially used for clinical trial enrichment and assessments for therapeutic response in dry AMD patients.

Progression Dynamics of Early versus Later Stage Atrophic Lesions in Nonneovascular Age-Related Macular Degeneration Using Quantitative OCT Biomarker Segmentation

PURPOSE

To investigate the progression of geographic atrophy secondary to nonneovascular age-related macular degeneration in early and later stage lesions using artificial intelligence-based precision tools.

DESIGN

Retrospective analysis of an observational cohort study.

SUBJECTS

Seventy-four eyes of 49 patients with ≥ 1 complete retinal pigment epithelial and outer retinal atrophy (cRORA) lesion secondary to age-related macular degeneration were included. Patients were divided between recently developed cRORA and lesions with advanced disease status.

METHODS

Patients were prospectively imaged by spectral-domain OCT volume scans. The study period encompassed 18 months with scheduled visits every 6 months. Growth rates of recent cRORA-converted lesions were compared with lesions in an advanced disease status using mixed effect models.

MAIN OUTCOME MEASURES

The progression of retinal pigment epithelial loss (RPEL) was considered the primary end point. Secondary end points consisted of external limiting membrane disruption and ellipsoid zone loss. These pathognomonic imaging biomarkers were quantified using validated deep-learning algorithms. Further, the ellipsoid zone/RPEL ratio was analyzed in both study cohorts.

RESULTS

Mean (95% confidence interval [CI]) square root progression of recently converted lesions was 79.68 (95% CI, -77.14 to 236.49), 68.22 (95% CI, -101.21 to 237.65), and 84.825 (95% CI, -124.82 to 294.47) mm/half year for RPEL, external limiting membrane loss, and ellipsoid zone loss respectively. Mean square root progression of advanced lesions was 131.74 (95% CI, -22.57 to 286.05), 129.96 (95% CI, -36.67 to 296.59), and 116.84 (95% CI, -90.56 to 324.3) mm/half year for RPEL, external limiting membrane loss, and ellipsoid zone loss, respectively. RPEL (P = 0.038) and external limiting membrane disruption (P = 0.026) progression showed significant differences between the 2 study cohorts. Further recent converters had significantly (P < 0.001) higher ellipsoid zone/RPEL ratios at all time points compared with patients in an advanced disease status (1.71 95% CI, 1.12-2.28 vs. 1.14; 95% CI, 0.56-1.71).

CONCLUSION

Early cRORA lesions have slower growth rates in comparison to atrophic lesions in advanced disease stages. Differences in growth dynamics may play a crucial role in understanding the pathophysiology of nonneovascular age-related macular degeneration and for the interpretation of clinical trials in geographic atrophy. Individual disease monitoring using artificial intelligence-based quantification paves the way toward optimized geographic atrophy management.

FINANCIAL DISCLOSURE(S)

The author(s) have no proprietary or commercial interest in any materials discussed in this article.

Artificial intelligence in retinal image analysis: Development, advances, and challenges

Modern advances in diagnostic technologies offer the potential for unprecedented insight into ophthalmic conditions relating to the retina. We discuss the current landscape of artificial intelligence in retina with respect to screening, diagnosis, and monitoring of retinal pathologies such as diabetic retinopathy, diabetic macular edema, central serous chorioretinopathy, and age-related macular degeneration. We review the methods used in these models and evaluate their performance in both research and clinical contexts and discuss potential future directions for investigation, use of multiple imaging modalities in artificial intelligence algorithms, and challenges in the application of artificial intelligence in retinal pathologies.

Geographic Atrophy Segmentation Using Multimodal Deep Learning

Purpose

To examine deep learning (DL)-based methods for accurate segmentation of geographic atrophy (GA) lesions using fundus autofluorescence (FAF) and near-infrared (NIR) images.

Methods

This retrospective analysis utilized imaging data from study eyes of patients enrolled in Proxima A and B (NCT02479386; NCT02399072) natural history studies of GA. Two multimodal DL networks (UNet and YNet) were used to automatically segment GA lesions on FAF; segmentation accuracy was compared with annotations by experienced graders. The training data set comprised 940 image pairs (FAF and NIR) from 183 patients in Proxima B; the test data set comprised 497 image pairs from 154 patients in Proxima A. Dice coefficient scores, Bland-Altman plots, and Pearson correlation coefficient (r) were used to assess performance.

Results

On the test set, Dice scores for the DL network to grader comparison ranged from 0.89 to 0.92 for screening visit; Dice score between graders was 0.94. GA lesion area correlations (r) for YNet versus grader, UNet versus grader, and between graders were 0.981, 0.959, and 0.995, respectively. Longitudinal GA lesion area enlargement correlations (r) for screening to 12 months (n = 53) were lower (0.741, 0.622, and 0.890, respectively) compared with the cross-sectional results at screening. Longitudinal correlations (r) from screening to 6 months (n = 77) were even lower (0.294, 0.248, and 0.686, respectively).

Conclusions

Multimodal DL networks to segment GA lesions can produce accurate results comparable with expert graders.

Translational Relevance

DL-based tools may support efficient and individualized assessment of patients with GA in clinical research and practice.

Deep learning to detect macular atrophy in wet age-related macular degeneration using optical coherence tomography

Here, we have developed a deep learning method to fully automatically detect and quantify six main clinically relevant atrophic features associated with macular atrophy (MA) using optical coherence tomography (OCT) analysis of patients with wet age-related macular degeneration (AMD). The development of MA in patients with AMD results in irreversible blindness, and there is currently no effective method of early diagnosis of this condition, despite the recent development of unique treatments. Using OCT dataset of a total of 2211 B-scans from 45 volumetric scans of 8 patients, a convolutional neural network using one-against-all strategy was trained to present all six atrophic features followed by a validation to evaluate the performance of the models. The model predictive performance has achieved a mean dice similarity coefficient score of 0.706 ± 0.039, a mean Precision score of 0.834 ± 0.048, and a mean Sensitivity score of 0.615 ± 0.051. These results show the unique potential of using artificially intelligence-aided methods for early detection and identification of the progression of MA in wet AMD, which can further support and assist clinical decisions.

Detection of macular atrophy in age-related macular degeneration aided by artificial intelligence

INTRODUCTION

Age-related macular degeneration (AMD) is a leading cause of irreversible visual impairment worldwide. The endpoint of AMD, both in its dry or wet form, is macular atrophy (MA) which is characterized by the permanent loss of the RPE and overlying photoreceptors either in dry AMD or in wet AMD. A recognized unmet need in AMD is the early detection of MA development.

AREAS COVERED

Artificial Intelligence (AI) has demonstrated great impact in detection of retinal diseases, especially with its robust ability to analyze big data afforded by ophthalmic imaging modalities, such as color fundus photography (CFP), fundus autofluorescence (FAF), near-infrared reflectance (NIR), and optical coherence tomography (OCT). Among these, OCT has been shown to have great promise in identifying early MA using the new criteria in 2018.

EXPERT OPINION

There are few studies in which AI-OCT methods have been used to identify MA; however, results are very promising when compared to other imaging modalities. In this paper, we review the development and advances of ophthalmic imaging modalities and their combination with AI technology to detect MA in AMD. In addition, we emphasize the application of AI-OCT as an objective, cost-effective tool for the early detection and monitoring of the progression of MA in AMD.

Clinical validation for automated geographic atrophy monitoring on OCT under complement inhibitory treatment

Geographic atrophy (GA) represents a late stage of age-related macular degeneration, which leads to irreversible vision loss. With the first successful therapeutic approach, namely complement inhibition, huge numbers of patients will have to be monitored regularly. Given these perspectives, a strong need for automated GA segmentation has evolved. The main purpose of this study was the clinical validation of an artificial intelligence (AI)-based algorithm to segment a topographic 2D GA area on a 3D optical coherence tomography (OCT) volume, and to evaluate its potential for AI-based monitoring of GA progression under complement-targeted treatment. 100 GA patients from routine clinical care at the Medical University of Vienna for internal validation and 113 patients from the FILLY phase 2 clinical trial for external validation were included. Mean Dice Similarity Coefficient (DSC) was 0.86 ± 0.12 and 0.91 ± 0.05 for total GA area on the internal and external validation, respectively. Mean DSC for the GA growth area at month 12 on the external test set was 0.46 ± 0.16. Importantly, the automated segmentation by the algorithm corresponded to the outcome of the original FILLY trial measured manually on fundus autofluorescence. The proposed AI approach can reliably segment GA area on OCT with high accuracy. The availability of such tools represents an important step towards AI-based monitoring of GA progression under treatment on OCT for clinical management as well as regulatory trials.

Quantitative approaches in multimodal fundus imaging: State of the art and future perspectives

When it first appeared, multimodal fundus imaging revolutionized the diagnostic workup and provided extremely useful new insights into the pathogenesis of fundus diseases. The recent addition of quantitative approaches has further expanded the amount of information that can be obtained. In spite of the growing interest in advanced quantitative metrics, the scientific community has not reached a stable consensus on repeatable, standardized quantitative techniques to process and analyze the images. Furthermore, imaging artifacts may considerably affect the processing and interpretation of quantitative data, potentially affecting their reliability. The aim of this survey is to provide a comprehensive summary of the main multimodal imaging techniques, covering their limitations as well as their strengths. We also offer a thorough analysis of current quantitative imaging metrics, looking into their technical features, limitations, and interpretation. In addition, we describe the main imaging artifacts and their potential impact on imaging quality and reliability. The prospect of increasing reliance on artificial intelligence-based analyses suggests there is a need to develop more sophisticated quantitative metrics and to improve imaging technologies, incorporating clear, standardized, post-processing procedures. These measures are becoming urgent if these analyses are to cross the threshold from a research context to real-life clinical practice.

Application of Machine Learning to Ranking Predictors of Anti-VEGF Response

Age-related macular degeneration (AMD) is a heterogeneous disease affecting the macula of individuals and is a cause of irreversible vision loss. Patients with neovascular AMD (nAMD) are candidates for the anti-vascular endothelial growth factor (anti-VEGF) treatment, designed to regress the growth of abnormal blood vessels in the eye. Some patients fail to maintain vision despite treatment. This study aimed to develop a prediction model based on features weighted in order of importance with respect to their impact on visual acuity (VA). Evaluations included an assessment of clinical, lifestyle, and demographic factors from patients that were treated over a period of two years. The methods included mixed-effects and relative importance modelling, and models were tested against model selection criteria, diagnostic and assumption checks, and forecasting errors. The most important predictors of an anti-VEGF response were the baseline VA of the treated eye, the time (in weeks), treatment quantity, and the treated eye. The model also ranked the impact of other variables, such as intra-retinal fluid, haemorrhage, pigment epithelium detachment, treatment drug, baseline VA of the untreated eye, and various lifestyle and demographic factors. The results identified variables that could be targeted for further investigation in support of personalised treatments based on patient data.

Risk Classification for Progression to Subfoveal Geographic Atrophy in Dry Age-Related Macular Degeneration Using Machine Learning-Enabled Outer Retinal Feature Extraction

BACKGROUND AND OBJECTIVE

To evaluate the utility of spectral-domain optical coherence tomography biomarkers to predict the development of subfoveal geographic atrophy (sfGA).

PATIENTS AND METHODS

This was a retrospective cohort analysis including 137 individuals with dry age-related macular degeneration without sfGA with 5 years of follow-up. Multiple spectral-domain optical coherence tomography quantitative metrics were generated, including ellipsoid zone (EZ) integrity and subretinal pigment epithelium (sub-RPE) compartment features.

RESULTS

Reduced mean EZ-RPE central subfield thickness and increased sub-RPE compartment thickness were significantly different between sfGA convertors and nonconvertors at baseline in both 2-year and 5-year sfGA risk assessment. Longitudinal change assessment showed a significantly higher degradation of EZ integrity in sfGA convertors. The predictive performance of a machine learning classification model based on 5-year and 2-year risk conversion to sfGA demonstrated an area under the receiver operating characteristic curve of 0.92 ± 0.06 and 0.96 ± 0.04, respectively.

CONCLUSIONS

Quantitative outer retinal and sub-RPE feature assessment using a machine learning-enabled retinal segmentation platform provides multiple parameters that are associated with progression to sfGA. [Ophthalmic Surg Lasers Imaging. 2022;53:31-39.].

Progression of Geographic Atrophy: Epistemic Uncertainties Affecting Mathematical Models and Machine Learning

Purpose

The purpose of this study was to identify a taxonomy of epistemic uncertainties that affect results for geographic atrophy (GA) assessment and progression.

Methods

An important source of variability is called "epistemic uncertainty," which is due to incomplete system knowledge (i.e. limitations in measurement devices, artifacts, and human subjective evaluation, including annotation errors). In this study, different epistemic uncertainties affecting the analysis of GA were identified and organized into a taxonomy. The uncertainties were discussed and analyzed, and an example was provided in the case of model structure uncertainty by characterizing progression of GA by mathematical modelling and machine learning. It was hypothesized that GA growth follows a logistic (sigmoidal) function. Using case studies, the GA growth data were used to test the sigmoidal hypothesis.

Results

Epistemic uncertainties were identified, including measurement error (imperfect outcomes from measuring tools), subjective judgment (grading affected by grader's vision and experience), model input uncertainties (data corruption or entry errors), and model structure uncertainties (elucidating the right progression pattern). Using GA growth data from case studies, it was demonstrated that GA growth can be represented by a sigmoidal function, where growth eventually approaches an upper limit.

Conclusion

Epistemic uncertainties contribute to errors in study results and are reducible if identified and addressed. By prior identification of epistemic uncertainties, it is possible to (a) quantify uncertainty not accounted for by natural statistical variability, and (b) reduce the presence of these uncertainties in future studies.

Translational Relevance

Lowering epistemic uncertainty will reduce experimental error, improve consistency and reproducibility, and increase confidence in diagnostics.

Clinically relevant deep learning for detection and quantification of geographic atrophy from optical coherence tomography: a model development and external validation study

BACKGROUND

Geographic atrophy is a major vision-threatening manifestation of age-related macular degeneration, one of the leading causes of blindness globally. Geographic atrophy has no proven treatment or method for easy detection. Rapid, reliable, and objective detection and quantification of geographic atrophy from optical coherence tomography (OCT) retinal scans is necessary for disease monitoring, prognostic research, and to serve as clinical endpoints for therapy development. To this end, we aimed to develop and validate a fully automated method to detect and quantify geographic atrophy from OCT.

METHODS

We did a deep-learning model development and external validation study on OCT retinal scans at Moorfields Eye Hospital Reading Centre and Clinical AI Hub (London, UK). A modified U-Net architecture was used to develop four distinct deep-learning models for segmentation of geographic atrophy and its constituent retinal features from OCT scans acquired with Heidelberg Spectralis. A manually segmented clinical dataset for model development comprised 5049 B-scans from 984 OCT volumes selected randomly from 399 eyes of 200 patients with geographic atrophy secondary to age-related macular degeneration, enrolled in a prospective, multicentre, phase 2 clinical trial for the treatment of geographic atrophy (FILLY study). Performance was externally validated on an independently recruited dataset from patients receiving routine care at Moorfields Eye Hospital (London, UK). The primary outcome was segmentation and classification agreement between deep-learning model geographic atrophy prediction and consensus of two independent expert graders on the external validation dataset.

FINDINGS

The external validation cohort included 884 B-scans from 192 OCT volumes taken from 192 eyes of 110 patients as part of real-life clinical care at Moorfields Eye Hospital between Jan 1, 2016, and Dec, 31, 2019 (mean age 78·3 years [SD 11·1], 58 [53%] women). The resultant geographic atrophy deep-learning model produced predictions similar to consensus human specialist grading on the external validation dataset (median Dice similarity coefficient [DSC] 0·96 [IQR 0·10]; intraclass correlation coefficient [ICC] 0·93) and outperformed agreement between human graders (DSC 0·80 [0·28]; ICC 0·79). Similarly, the three independent feature-specific deep-learning models could accurately segment each of the three constituent features of geographic atrophy: retinal pigment epithelium loss (median DSC 0·95 [IQR 0·15]), overlying photoreceptor degeneration (0·96 [0·12]), and hypertransmission (0·97 [0·07]) in the external validation dataset versus consensus grading.

INTERPRETATION

We present a fully developed and validated deep-learning composite model for segmentation of geographic atrophy and its subtypes that achieves performance at a similar level to manual specialist assessment. Fully automated analysis of retinal OCT from routine clinical practice could provide a promising horizon for diagnosis and prognosis in both research and real-life patient care, following further clinical validation FUNDING: Apellis Pharmaceuticals.

Improving Interpretability in Machine Diagnosis

Purpose

Manually identifying geographic atrophy (GA) presence and location on OCT volume scans can be challenging and time consuming. This study developed a deep learning model simultaneously (1) to perform automated detection of GA presence or absence from OCT volume scans and (2) to provide interpretability by demonstrating which regions of which B-scans show GA.

Design

Med-XAI-Net, an interpretable deep learning model was developed to detect GA presence or absence from OCT volume scans using only volume scan labels, as well as to interpret the most relevant B-scans and B-scan regions.

Participants

One thousand two hundred eighty-four OCT volume scans (each containing 100 B-scans) from 311 participants, including 321 volumes with GA and 963 volumes without GA.

Methods

Med-XAI-Net simulates the human diagnostic process by using a region-attention module to locate the most relevant region in each B-scan, followed by an image-attention module to select the most relevant B-scans for classifying GA presence or absence in each OCT volume scan. Med-XAI-Net was trained and tested (80% and 20% participants, respectively) using gold standard volume scan labels from human expert graders.

Main Outcome Measures

Accuracy, area under the receiver operating characteristic (ROC) curve, F₁ score, sensitivity, and specificity.

Results

In the detection of GA presence or absence, Med-XAI-Net obtained superior performance (91.5%, 93.5%, 82.3%, 82.8%, and 94.6% on accuracy, area under the ROC curve, F₁ score, sensitivity, and specificity, respectively) to that of 2 other state-of-the-art deep learning methods. The performance of ophthalmologists grading only the 5 B-scans selected by Med-XAI-Net as most relevant (95.7%, 95.4%, 91.2%, and 100%, respectively) was almost identical to that of ophthalmologists grading all volume scans (96.0%, 95.7%, 91.8%, and 100%, respectively). Even grading only 1 region in 1 B-scan, the ophthalmologists demonstrated moderately high performance (89.0%, 87.4%, 77.6%, and 100%, respectively).

Conclusions

Despite using ground truth labels during training at the volume scan level only, Med-XAI-Net was effective in locating GA in B-scans and selecting relevant B-scans within each volume scan for GA diagnosis. These results illustrate the strengths of Med-XAI-Net in interpreting which regions and B-scans contribute to GA detection in the volume scan.

Modeling of atrophy size trajectories: variable transformation, prediction and age-of-onset estimation

Background

To model the progression of geographic atrophy (GA) in patients with age-related macular degeneration (AMD) by building a suitable statistical regression model for GA size measurements obtained from fundus autofluorescence imaging.

Methods

Based on theoretical considerations, we develop a linear mixed-effects model for GA size progression that incorporates covariable-dependent enlargement rates as well as correlations between longitudinally collected GA size measurements. To capture nonlinear progression in a flexible way, we systematically assess Box-Cox transformations with different transformation parameters λ. Model evaluation is performed on data collected for two longitudinal, prospective multi-center cohort studies on GA size progression.

Results

A transformation parameter of λ=0.45 yielded the best model fit regarding the Akaike information criterion (AIC). When hypertension and hypercholesterolemia were included as risk factors in the model, they showed an association with progression of GA size. The mean estimated age-of-onset in this model was 67.21±6.49 years.

Conclusions

We provide a comprehensive framework for modeling the course of uni- or bilateral GA size progression in longitudinal observational studies. Specifically, the model allows for age-of-onset estimation, identification of risk factors and prediction of future GA size. A square-root transformation of atrophy size is recommended before model fitting.

Supplementary Information

The online version contains supplementary material available at (10.1186/s12874-021-01356-0).

Deep Learning Applied to Automated Segmentation of Geographic Atrophy in Fundus Autofluorescence Images

Purpose

This study describes the development of a deep learning algorithm based on the U-Net architecture for automated segmentation of geographic atrophy (GA) lesions in fundus autofluorescence (FAF) images.

Methods

Image preprocessing and normalization by modified adaptive histogram equalization were used for image standardization to improve effectiveness of deep learning. A U-Net-based deep learning algorithm was developed and trained and tested by fivefold cross-validation using FAF images from clinical datasets. The following metrics were used for evaluating the performance for lesion segmentation in GA: dice similarity coefficient (DSC), DSC loss, sensitivity, specificity, mean absolute error (MAE), accuracy, recall, and precision.

Results

In total, 702 FAF images from 51 patients were analyzed. After fivefold cross-validation for lesion segmentation, the average training and validation scores were found for the most important metric, DSC (0.9874 and 0.9779), for accuracy (0.9912 and 0.9815), for sensitivity (0.9955 and 0.9928), and for specificity (0.8686 and 0.7261). Scores for testing were all similar to the validation scores. The algorithm segmented GA lesions six times more quickly than human performance.

Conclusions

The deep learning algorithm can be implemented using clinical data with a very high level of performance for lesion segmentation. Automation of diagnostics for GA assessment has the potential to provide savings with respect to patient visit duration, operational cost and measurement reliability in routine GA assessments.

Translational Relevance

A deep learning algorithm based on the U-Net architecture and image preprocessing appears to be suitable for automated segmentation of GA lesions on clinical data, producing fast and accurate results.