Advancing Medical Imaging: A Domain-Adaptive Approach to Distinguish Lung Cancer and Pulmonary Cryptococcosis

Pulmonary cryptococcosis (PC) is a subacute or chronic visceral mycosis caused by Cryptococcus neoformans, a yeast. This infection typically affects the lungs and central nervous system, with symptoms associated with pneumonia, meningitis, or involvement of the skin, bones, or internal organs. Computed Tomography (CT) is an important investigation for cryptococcosis and other lung diseases. Radiologists determine whether the patient has pulmonary cryptococcosis or lung cancer based on CT findings and clinical symptoms. However, the CT manifestations of pulmonary cryptococcus and lung cancer are similar, which often leads to misdiagnosis and makes patients suffer. In this paper, we present a domain adaptive classification method for addressing domain differences between datasets from different sources. This method combines a channel attention mechanism with adversarial networks to alleviate domain differences and enhance the model's generalization ability. Specifically, the channel attention aims to increase sensitivity to key features and enhance feature representation. By incorporating domain adaptation functions, the source domain is integrated with the target domain to focus on capturing feature differences. We validated this approach on a clinical dataset and achieved excellent results.