Deep learning based diagnostic quality assessment of choroidal OCT features with expert-evaluated explainability

Abstract

Various vision-threatening eye diseases 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. In the current clinical practice, screening choroidal structural changes is widely based on optical coherence tomography (OCT) images. Accordingly, to assist clinicians, several automated choroidal biomarker detection methods using OCT images are developed. However, the performance of these algorithms is largely constrained by the quality of the OCT scan. Consequently, determining the quality of choroidal features in OCT scans is significant in building standardized quantification tools and hence constitutes our main objective. This study includes a dataset of 1593 good and 2581 bad quality Spectralis OCT images graded by an expert. Noting the efficacy of deep-learning (DL) in medical image analysis, we propose to train three state-of-the-art DL models: ResNet18, EfficientNet-B0 and EfficientNet-B3 to detect the quality of OCT images. The choice of these models was inspired by their ability to preserve the salient features across all the layers without information loss. To evaluate the attention of DL models on the choroid, we introduced color transparency maps (CTMs) based on GradCAM explanations. Further, we proposed two subjective grading scores: overall choroid coverage (OCC) and choroid coverage in the visible region(CCVR) based on CTMs to objectively correlate visual explanations vis-à-vis DL model attentions. We observed that the average accuracy and F-scores for the three DL models are greater than 96%. Further, the OCC and CCVR scores achieved for the three DL models under consideration substantiate that they mostly focus on the choroid layer in making the decision. In particular, of the three DL models, EfficientNet-B3 is in close agreement with the clinician’s inference. The proposed DL-based framework demonstrated high detection accuracy as well as attention on the choroid layer, where EfficientNet-B3 reported superior performance. Our work assumes significance in bench-marking the automated choroid biomarker detection tools and facilitating high-throughput screening. Further, the methods proposed in this work can be adopted for evaluating the attention of DL-based approaches developed for other region-specific quality assessment tasks.