Multi-pathology classification of retinal OCT scans

Abstract

Purpose : Optical coherence tomography (OCT) scans are commonly used for eye disease diagnosis. Each OCT volume scan consists of hundreds of individual B-scans. In a typical OCT review workflow, clinicians need to examine each B-scan individually. Automatic detection of different pathologies in B-scans could improve the workflow for disease diagnosis. In this study, we developed an efficient multi-label Convolutional Neural Network (CNN) that flags the presence of eight different retinal abnormalities in individual B-scans.

Methods : Multi-head CNN was trained to perform eight-label binary classification on each input image. Figure 1 shows the architecture of the network and abnormality labels. We trained the entire network end-to-end with a pre-trained model (ResNet-50 trained on ImageNet). Later, by freezing the shared representation sub-net, we fine-tuned each classification head separately to maximize the sensitivity/specificity for each abnormality.
76,544 B-scans (598 eyes of 598 subjects; 128 B-scans from each eye) captured by CIRRUS™ HD-OCT 4000 and 5000 (ZEISS, Dublin, CA), labeled by two experts and adjudicated by a third expert were used in this study [1]. The dataset was split into train, validation, and test sets with 80%/10%/10% split at the patient level. However, the data is highly imbalanced (shown in Figure 2a), which affects the model's performance. We used a macro soft F_β-score with a beta of 2 as a loss function to address this issue in training and to increase the sensitivity and specificity simultaneously.

Results : Despite the highly unbalanced and limited training data, our method achieved mean sensitivity of 86.8% and specificity of 87.4%, and a mean accuracy of 87.6% by having a common representation sub-net and specialized classification head. Figure 2b shows the detailed accuracy metrics for each abnormality in the hold-out test set.

Conclusions : By designing a multi-head CNN and using a custom loss function, we were able to train a single model to perform eight-label classification on OCT B-scan images to improve the accuracy and efficiency of clinical workflow.

[1] Yu et al. IOVS 2020; 61(9): PB0085

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

 

Figure 1: Architecture of the multi-head classification CNN

 

Figure 2: Distribution of the data and evaluation results of the model