Automated INL/OPL subsidence detection in intermediate AMD with deep neural networks

Abstract

Purpose : The subsidence of the inner nuclear layer (INL) and outer plexiform layer (OPL) is an important imaging biomarker associated with the initial outer retinal atrophy in patients with intermediate age-related macular degeneration (AMD). Deep neural networks (DNNs) for optical coherence tomography (OCT) can help patient screening and automated localization of this biomarker.

Methods : The method predicts bounding boxes with potential subsidences on each B-scan of an OCT. The system is composed of two DNNs: a detection module (DM) that infers bounding boxes around subsidences with a likelihood score and a classification module (CM) that scores subsidence presence at a B-scan level. The final prediction results from merging the overlapping boxes B-scan-wise, and the final lesion score is the product of the DM and CM predictions. The volume-wise subsidence score is the maximum of the predictions for all B-scans. The model ensemble increases the system's robustness by reducing the number of false-positive detections (FPs) from DM while maintaining its explainability. Both CM and DM use a ConvNeXt backbone fine-tuned for the corresponding tasks. The CM has a binary output from a fully-connected layer; the DM uses the MaskR-CNN architecture. The approach is evaluated via stratified cross-validation with ten patient-wise splits.

Results : 1960 Heidelberg Spectralis OCT scans of 280 eyes (76 positive) from 140 patients (49 positive) were used for training and testing. The average volume-wise subsidence detection area under the Receiver Operating Characteristic (AUC) was 0.97±0.01 (Fig. 1A) and the area under the Precision-Recall Curve was 0.48±0.14. For positive cases, the subsidence locations were correctly identified more than 80% of the time with a single FP detection/scan (Fig. 1B). Representative detection examples are shown in Fig. 2.

Conclusions : DNN systems are efficiently able to perform automated INL/OPL subsidence detection in OCT images. This allows objective and early identification of progression to advanced AMD for screening and risk assessment. Such tools are essential for facilitating access to timely treatment for the wider population.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Fig. 1: Curves for the ten folds in study. A: Receiver Operating Characteristic (ROC). B: Free-response Receiver Operating Characteristic curve (subsidence cases only) for the DM and DM+CM.

 

Fig. 2: Example of TP (left) and FP (right) and the corresponding prediction score.