Semi-supervised pre-training based multi-task network for thyroid-associated ophthalmopathy classification

Abstract

Thyroid-associated ophthalmopathy (TAO) is a blinding autoimmune disorder, and early diagnosis is crucial in preventing vision loss. Orbital CT imaging has emerged as a valuable tool for diagnosing and screening TAO. Radiomic is currently the most dominant technique for TAO diagnosis, however it is costly due to the need for manual image labeling by medical professionals. Convolutional Neural Network (CNN) is another promising technique for TAO diagnosis. However, the performance of CNN based classification may degrade due to the limited size of collected data or the complexity of designed model. Utilizing pretraining model is a crucial technique for boosting the performance of CNN based TAO classification. Therefore, a novel semi-supervised pretraining based multi-task network for TAO classification is proposed in this paper. Firstly, a multi-task network is designed, which consists of an encoder, a classification branch and two segmentation decoder. Then, the multi-task network is pretrained by minimizing the prediction difference between two segmentation decoders through a semi-supervised way. In this way, the pseudo voxel-level supervision can be generated for the unlabeled images. Finally, the encoder and one light-weighted decoder can be initialized by the pretrained weights, and then they are jointly optimized for TAO classification with the classification branch through multi-task learning. Our proposed network model was comprehensively evaluated on a private dataset which consists of 982 orbital CT scans for TAO diagnosis. We also tested the classification generalization performance using an external dataset. The experimental results demonstrate that our model significantly improves the classification performance when compared with current SOTA methods. The source code is publically available at https://github.com/VLAD-KONATA/TAO_CT.