A 3D CNN-based Multi-task Learning for Cataract screening and left and right eye classification on 3D AS-OCT images

Abstract

Cataract is the leading cause for visual impairment and blindness. Cataract screening can effectively improve the recovery rate of cataract, and the left and right eye classification is a significant step in cataract screening. Anterior segment optical coherence tomography (AS-OCT) is a non-contact, high-resolution ophthalmic imaging technique, which can quickly obtain pathological information of cataract and left and right eye position information through three-dimensional (3D) imaging. In order to improve the efficiency of cataract screening, we propose a multi-task three-dimensional convolutional neural network (MT-CNN) for automatic cataract detection and left and right eye classification simultaneously based on the 3D AS-OCT images. The MT-CNN is designed based on the hard sharing mechanism, achieving better performance with fewer parameters than single-task learning. The results on an AS-OCT image dataset show that the 3D CNN model obtains better classification performance than the 2D CNN model. Compared with the single-task 3D CNN model, MT-CNN achieves higher accuracy under the premise of greatly parameters reduction and computational complexity reduction.