Diagnosing tic disorders from videos using multi-phase learning

Xiaojing Xu, Ruizhe Zhang, Zihao Bo, Junfeng Lyu, Yuchen Guo, Feng Xu

Abstract: Worldwidely, the number of individuals with tic disorder has reached 59 million, while the prevalence of this disorder is still rapidly increasing. In this work, we proposed a multi-phase learning method for diagnosing childhood tic disorders from facial videos. To handle the problem of limited data annotation, we design an Entropy Gain (EG) metric to generate and select samples with pseudo labels and propose a multi-phase learning algorithm to efficiently leverage the EG-labeled data in a "from easy to difficult" manner. In our method, we use aligned facial landmarks as a compact data representation to further protect patient privacy and achieve efficient learning. Through extensive experiments on the test dataset, we demonstrate that our method behaves extraordinarily better compared to baseline approaches, improving AUC by 3.9 %, and facilitating expedited diagnostic assessment for medical practitioners.

External IDs:doi:10.1016/j.compeleceng.2025.110216