DTN-MTLF: A dual-teacher network based multi-task learning framework for unbiased glaucoma diagnosis

Glaucoma is a chronic progressive optic neuropathy that is the leading cause of irreversible blindness worldwide. Developing precise automatic screening algorithms is essential for early detection and treatment of glaucoma. However, this is not an easy task and faces many challenges. Therefore, this study proposes a dual-teacher network based multi-task learning framework (DTN-MTLF). The proposed method includes a dual-teacher network and a student network. The dual-teacher network is used to generate distilled labels and evidence map labels, and achieve unbiased screening of glaucoma with the support of curriculum modules. The student network is a dual-branch multi-task convolutional neural network (CNN) that can diagnose glaucoma and predict evidence map. The proposed method can effectively alleviate the diagnostic bias of glaucoma and generate high-confidence evidence maps. The accuracy, sensitivity, specificity, AUC and \(\text {F}_2\) ${\text{F}}_2$ -score of student network on the LAG dataset for diagnosing glaucoma are 0.981, 0.978, 0.980, 0.997 and 0.976 respectively. These experimental results indicate that the proposed method is superior to the state-of-the-art glaucoma diagnosis methods. Moreover, DTN-MTLF also enables real-time inference and generates reliable evidence maps to help clinicians explore the lesion area in detail.