Deep learning-based automated detection of retinal diseases using optical coherence tomography images.

Feng Li,Hua Chen,Zhi-Zheng Wu,Kai-Qian Zhou,Min-Shan Jiang,Xue-Dian Zhang,Zheng Liu

doi:10.1364/boe.10.006204

Abstract

Retinal disease classification is a significant problem in computer-aided diagnosis (CAD) for medical applications. This paper is focused on a 4-class classification problem to automatically detect choroidal neovascularization (CNV), diabetic macular edema (DME), DRUSEN, and NORMAL in optical coherence tomography (OCT) images. The proposed classification algorithm adopted an ensemble of four classification model instances to identify retinal OCT images, each of which was based on an improved residual neural network (ResNet50). The experiment followed a patient-level 10-fold cross-validation process, on development retinal OCT image dataset. The proposed approach achieved 0.973 (95% confidence interval [CI], 0.971-0.975) classification accuracy, 0.963 (95% CI, 0.960-0.966) sensitivity, and 0.985 (95% CI, 0.983-0.987) specificity at the B-scan level, achieving a matching or exceeding performance to that of ophthalmologists with significant clinical experience. Other performance measures used in the study were the area under receiver operating characteristic curve (AUC) and kappa value. The observations of the study implied that multi-ResNet50 ensembling was a useful technique when the availability of medical images was limited. In addition, we performed qualitative evaluation of model predictions, and occlusion testing to understand the decision-making process of our model. The paper provided an analytical discussion on misclassification and pathology regions identified by the occlusion testing also. Finally, we explored the effect of the integration of retinal OCT images and medical history data from patients on model performance.

Highlights

Our approach used an ensemble of four classification model instances as shown in Fig. 4, each of which was based on an improved ResNet50
In the multiclass comparison between choroidal neovascularization (CNV), diabetic macular edema (DME), DRUSEN, and NORMAL, our model achieved an accuracy of 0.973 with a sensitivity of 0.963 and a specificity of 0.985
We prepared and processed a relatively big dataset of retinal optical coherence tomography (OCT) images captured in real-world setting, and presented a novel ensemble of four classification model instances to detect three most common blinding retinal diseases from OCT images automatically and reliably, each of which was based on an improved ResNet50

Summary

Introduction

Lemaître et al [14] proposed a method based on extracted local binary pattern features from OCT images and dictionary learning using bag of words models for DME detection, which achieved a sensitivity and a specificity of 81.2% and 93.7%, respectively. Srinivasan et al [16] employed histogram of oriented gradients (HOG) descriptors and SVM classifiers for the detection of age-related macular degeneration (AMD), DME, and normal retina, which acquired 100%, 100%, and 86.67% accuracy at the OCT level, respectively. These classification approaches highly relied on features explicitly defined by ophthalmologists using their domain knowledge, resulting in time-consuming, weak generalization ability, and even unfeasibility in large datasets

Methods

Results

Discussion

Conclusion