Steven T. Bailey

Aim: To implement a deep learning-based segmentation algorithm to quantify reticular pseudodrusen (RPD) and drusen volumes on optical coherence tomography (OCT) and investigate their association with progression to late age-related macular degeneration (AMD). Methods: A retrospective analysis included study eyes with RPD and contralateral neovascular AMD using 6×6 mm macular OCT (Solix; Visionix/Optovue, Inc). Automated segmentation quantified RPD and drusen volumes, including large drusen and drusenoid pigment epithelial detachment (PED), and late AMD development was evaluated over 2 years. Associations between baseline volumetric biomarkers and progression were evaluated using Cox proportional hazards models. Results: Fifty-one eyes (mean age 74.9±7.65 years) were included. The median (IQR) baseline RPD volume was 0.018 mm³ (0.004–0.52) and total drusen volume was 0.009 mm³ (0.001–0.064). Over 24.2±1.20 months, late AMD developed in 20 eyes (39.2%). In multivariable Cox regression models adjusted for age, each 0.01 mm³ increase in baseline RPD volume (HR: 1.082; p=0.002) and total drusen volume (HR: 1.080; p<0.001) were associated with progression to late AMD. In an exploratory drusen subtype analysis, progression to late AMD was independently associated with baseline RPD volume, large drusen volume and drusenoid PED volume. Conclusion: The deep learning-based volumetric segmentation tool allows OCT-derived automated volume quantification of different types of drusen based on OCT. In eyes with RPD and contralateral neovascular AMD, greater RPD volume and large drusen and/or drusenoid PED volume carried greater risk of developing late AMD in 2 years.

Objective: Artificial intelligence (AI)-based disease classifiers have achieved specialist-level performances in several diagnostic tasks. However, real-world adoption of these classifiers remains challenging due to the black box issue. Here, we report a novel biomarker activation map (BAM) generation framework that can provide clinically meaningful explainability to current AI-based disease classifiers. Methods: We designed the framework based on the concept of residual counterfactual explanation by generating counterfactual outputs that could reverse the decision-making of the disease classifier. The BAM was generated as the difference map between the counterfactual output and original input with postprocessing. We evaluated the BAM on four different disease classifiers, including an age-related macular degeneration classier based on fundus photography, a diabetic retinopathy classifier based on optical coherence tomography angiography, a brain tumor classifier based on magnetic resonance imaging (MRI), and a breast cancer classifier based on computerized tomography (CT) scans. Results: The highlighted regions in the BAM correlated highly with manually demarcated biomarkers of each disease. Conclusion/Significance: The BAM can improve the clinical applicability of an AI-based disease classifier by providing intuitive output clinicians can use to understand and verify the diagnostic decision.

An abstract is unavailable. This article is available as a PDF only.

Purpose To demonstrate the capabilities of single-shot widefield swept-source OCT angiography (SS-OCTA) in detecting subclinical retinal neovascularization (RNV), quantifying nonperfusion areas (NPAs), and exploring the relations between NPAs and subclinical RNV in eyes graded as nonproliferative diabetic retinopathy (NPDR). Methods Eyes clinically graded as moderate to severe NPDR underwent SS-OCTA imaging. Expert graders identified subclinical RNV, defined as vessels with a flow signal above the internal limiting membrane on OCTA that are not visible on dilated fundus examination. This identification was based on a combination of en face OCT, en face OCTA, and cross-sectional OCTA overlaid on OCT. NPA index was calculated as a percentage of automatically quantified NPA over the area in the posterior pole, the mid-periphery, and the total imaged area. Results Totally 37 eyes, including 21 had severe NPDR and 16 had moderate NPDR. Subclinical RNV was present in 14 eyes (37.8%). The eyes with RNV had significantly higher mid-peripheral and total NPA indices but not in the posterior region (mid-peripheral NPA: 31.97% ± 7.02% vs. 24.80% ± 6.60%, p = 0.041; total NPA: 27.96% ± 6.36% vs. 21.61% ± 5.65%, p = 0.046; all values are reported as mean ± standard deviation). The total NPA index showed the highest diagnostic accuracy for subclinical RNV detection (AUC: 0.761; 95% CI, 0.592–929, with a sensitivity of 64.3% and a specificity of 87% at a cutoff value of 28.84%). Conclusion Widefield SS-OCTA can detect subclinical RNV. The eyes with higher mid-peripheral NPA indices are more likely to have subclinical RNV, indicating that the NPA index may be a useful biomarker for identifying eyes at risk of RNV.

Purpose: To demonstrate the capabilities of single-shot widefield swept-source OCT angiography (SS-OCTA) in detecting subclinical retinal neovascularization (RNV), quantifying nonperfusion areas (NPAs), and exploring the relations between NPA and subclinical RNV in eyes graded as nonproliferative diabetic retinopathy (NPDR). Methods: Eyes clinically graded as moderate to severe NPDR underwent SS-OCTA imaging. Expert graders identified subclinical RNV, defined as vessels with flow signal above the internal limiting membrane on OCTA that are not visible on dilated fundus examination. This identification was based on a combination of en face OCT, en face OCTA, and cross-sectional OCTA overlaid on OCT. NPA index was calculated as a percentage of automatically quantified NPA over area in the posterior pole, the mid-periphery, and the total imaged area. Results: Totally 37 eyes including 21 had severe NPDR and 16 had moderate NPDR. Subclinical RNV was present in 14 eyes (37.8%). The eyes with RNV had significantly higher mid-peripheral and total NPA indices but not in the posterior region (mid-peripheral NPA: 31.97% ± 7.02% vs. 24.80% ± 6.60%, p=0.041; total NPA: 27.96% ± 6.36% vs. 21.61% ± 5.65%, p=0.046). The total NPA index showed the highest diagnostic accuracy for subclinical RNV detection (AUC: 0.761, with a sensitivity of 64.3% and a specificity of 87% at a cutoff value of 28.84%). Conclusion: Widefield SS-OCTA can detect subclinical RNV. The eyes with higher mid-peripheral NPA indices are more likely to have subclinical RNV, indicating that the NPA index may be a useful biomarker for identifying eyes at risk of RNV.