An end-to-end multi-task deep learning framework for bronchoscopy image classification

Lung cancer and tuberculosis (TB) are leading causes of mortality from lung diseases. Bronchoscopy plays a crucial role in diagnosing these conditions and determining appropriate treatment plans for patients. During bronchoscopy, clinicians often need to decide promptly whether to perform a lung biopsy upon observing abnormal symptoms. Since biopsies can lead to side effects such as excessive bleeding and infection, clinicians must make these decisions judiciously. Computer-aided diagnosis systems (CADx) can serve as virtual assistants, potentially preventing unnecessary procedures. This paper proposes a deep learning-based CADx system for diagnosing TB and lung cancer via bronchoscopy. Unlike normal and abnormal cases, lung cancer and TB are not easily distinguishable during bronchoscopy procedure. To address this challenge, a multi-task model with two branches, utilizing DenseNet and incorporating a Squeeze and Excitation (SE) module, is presented. Evaluated on a dataset of 515 images, the model achieved an impressive overall accuracy of 90.6%, surpassing a competing method. Sensitivities for cancer, TB, and normal cases were 91.3%, 81.5%, and 96.2%, respectively.