YOLOv10-FMW: A Multiscale Convolutional Detection Method for Pulmonary Tuberculosis.

Early detection and treatment of minimal pulmonary tuberculosis (mPTB) are essential for preventing disease progression and transmission. However, the detection of mPTBs on computed tomography (CT) remains challenging because the lesions are often small, low in contrast, morphologically heterogeneous, and characterized by subtle manifestations. This study aimed to develop and evaluate a dedicated deep learning framework for detecting mPTB lesions on chest CT. The model was trained and validated using data collected from three hospitals and externally evaluated using an independent dataset from a fourth hospital. We proposed YOLOv10 Faster-MSCM-WIoUv3 (YOLOv10-FMW), an improved YOLOv10-based detector. The backbone combines a faster block and an efficient multiscale attention mechanism to strengthen fine-grained feature representation. The neck incorporates a multiscale convolution module (MSCM) to facilitate cross-scale feature fusion for lesions with varied sizes and morphologies, whereas the WIoUv3 is used as the bounding-box regression loss to improve lesion localization. Patient-level data partitioning was applied to prevent potential data leakage, and tenfold cross-validation was performed to assess model stability. On the external test set, YOLOv10-FMW achieved an mAP@0.5 of 0.848, a precision of 0.845, a recall of 0.775, and an AUC of 0.828. These findings suggest that YOLOv10-FMW demonstrates promising detection performance and preliminary external generalizability in detecting mPTB. It holds promise as a computer-aided tool for early clinical detection.