Classification of Companion Animals’ Ocular Diseases: Domain Adversarial Learning for Imbalanced Data

Abstract

In contrast to the widespread implementation of computer-aided diagnosis of human diseases, the limited availability of veterinary image datasets has hindered its application in animals. Additionally, while most medical imaging data are captured in clinical settings, such as optical coherence tomography and fundus photography, diagnosis based on digital camera or smartphone images can be more beneficial for pet owners. This study specifically focuses on achieving generalization between screening environments, aiming to accurately diagnose diseases using casual images obtained by pet owners, despite the majority of training images being captured with specialized equipment in hospitals. Given these challenges and the significant role of computer-aided diagnosis in veterinary science, this study aims to develop a practical deep-learning framework for classifying ocular surface disease images in companion animals. The dataset used in this study consists of diverse ocular disease images of canines and felines obtained through slit lamps and digital cameras. The proposed approach includes two layers of labels for multitask learning and a gradient reversal layer based on normalized feature maps. We achieved 84.7% and 65.4% accuracy for the total dataset of canine and feline, respectively. For the camera domain in particular, canines and felines reached 86.2% and 73.2% accuracy, respectively.