Deep-learning based diagnostic quality assessment of choroid layer in OCT scans

Abstract

Purpose : Various vision-related ailments including age-related macular degeneration (AMD) and central serous chorioretinopathy (CSCR) are caused due to the dysfunctions manifested in the highly vascular choroid layer of the posterior segment of the eye. Accurate detection of choroidal structural changes plays a crucial role in disease diagnosis. Optical Coherence Tomography (OCT) imaging enabled clinicians to visualize choroid layer. Consequently, many methods have been developed to assist clinicians in diagnosis. However, the performance of these algorithms is largely constrained by the quality of the OCT scan. This study aims to achieve automated quality assessment of choroidal features (CQA) in OCT scans. We examine deep learning architectures to discriminate between good and bad quality OCT scans and also generate corresponding visual explanations via Grad-CAM.

Methods : A retrospective dataset of 1094 healthy and 3080 diseased OCT B-scans taken from Heidelberg Retina Angiography (HRA) Spectralis OCT machine was used in this experiment. The images were graded subjectively into two classes `good' and `bad' by a trained expert (K. Ozimba) based on various parameters including visibility of the choroid, the contrast between choroidal luminal (vessel) and stromal regions, contrast between the choroid and sclera, especially at choroid sclera interface. Inspired by the efficacy of deep learning features in image quality assessment, we employed variants of two ubiquitous architectures: residual networks (ResNets) and EfficientNet. We evaluated the model performance via stratified 5-fold cross-validation and demonstrated the model's discriminative ability through visual explanations generated by Grad-CAM.

Results : The average accuracy over 5-folds was found to be 97.69%, 96.99%, and 97.92%, respectively, for ResNet18, EfficientNet B0, and EfficientNet B3. Grad-CAM-generated visual explanations on representative images also demonstrate the superiority of EfficientNet B3 over the other models.

Conclusions : We have demonstrated the efficacy of transfer learning in CQA-OCT images by deploying and comparing the performance of state-of-the-art deep learning models, namely, ResNet18 and EfficientNet (B0 & B3). Further, we have also provided explainable visual feedback using Grad-CAM that corroborates closely with clinicians.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Proposed deep-learning-based framework for choroidal feature quality assessment in OCT.