Representative Slice Selection and Multi-View Projection Learning for Pulmonary Tuberculosis Infectiousness Identification Using CT Volume

Pulmonary tuberculosis (PTB) is a major global health threat. Diagnosing PTB infectiousness is vital for clinical decision-making, but existing etiological examination methods do not meet the requirements for speed, accuracy, and cost effectiveness. Developing deep learning models for infectiousness identification based on computed tomography (CT) volume measurements holds promise for meeting these requirements. However, with limited samples and coarse annotations, the large amount of information in the CT volume poses a challenge for models to distinguish information related to patient-level labels, which often leads to severe model overfitting. In this study, A dual-branch framework is developed for identifying PTB infectiousness using CT volume. To address the issue of imbalance between the CT volume information and the patient-level labels, we propose a method for selecting representative slices to reduce redundant information and adopt a multiple-instance learning framework to improve label supervision. Furthermore, we incorporate multi-view projection information to compensate for the deficiency of global information caused by using single-dimensional slices as the input. Experimental results demonstrate that our strategy effectively mitigates overfitting and achieves desirable performance on an external test set, with an area under the receiver operating characteristic curve of 80.48%. This performance is superior to that obtained for models using the 3D CT volume or 2D projection images alone as the input.