Source-Free Semi-Supervised Domain Adaptation for Tuberculosis Recognition

Jie Ma, Zhijun Xu, Xinyu Xiong, Ma-yi-di-li Ni-jia-ti, Dongyu Zhang, Weijun Sun, Feng Gao, Guanbin Li

Abstract: Tuberculosis (TB) as one of the major threats to human health worldwide, leads to millions of deaths every year. Despite numerous recent research efforts towards computer-aided TB diagnosis, these methods often suffer from data bias, or domain shift, across different imaging devices and hospitals. This limitation leads to poor performance in real-world scenarios, especially for multi-domain data. Moreover, patient privacy and data security concerns pose significant barriers to data accessibility and exacerbate the difficulty of domain adaptation. To mitigate these problems, in this paper, we propose a Bilateral-Branch Consistency Network (BBCN) for TB recognition under the Source-Free Semi-Supervised setting. The BBCN consists of a Source Model Adaptation branch (SMA) and a Target Model Learning (TML) branch that effectively enhances the cross-domain feature representation ability. A consistent regularizer is further proposed acting as a cross-domain regularization interaction machine between two branches. Experimental results demonstrate the effectiveness of our method.