A segmentation method for tuberculosis lesion in chest X-ray based on YOLO and MedSAM

DENG Donghua; HUO Yingyu; LIU Xin

doi:10.19301/j.cnki.zncs.2025.07.017

Abstract：Artificial intelligence has made great progress in the field of medical image analysis,and this paper proposes a collaborative segmentation framework for TB-YOLO-MedSAM,which aims to solve the problems of scarcity of training data,insufficient feature extraction ability,and strong dependence on manual annotation frame prompts in the segmentation of tuberculosis lesions in chest X-ray images.The improved TB-YOLOv8 model is introduced to realize the automatic detection of lesion area and the generation of prompt frames,and the parameters of MedSAM are efficiently fine-tuned by combining low-rank adaptive technology,and an efficient multi-scale attention module and residual bidirectional feature pyramid structure are designed to enhance the multi-scale feature fusion ability.The Foshan Fourth Hospital-TB dataset was selected for experimental verification.The results show that the proposed method has a Dice coefficient and an intersectional and conjugation ratio(IoU)of 0.746 and 0.701,which are better than the original MedSAM model of 0.340 and 0.234,respectively,and significantly better than the mainstream segmentation models such as U-Net,TransUNet,and SwinUnet.This method not only realizes the accurate segmentation of tuberculosis lesions,but also reduces the number of trainable parameters of the model,which provides a feasible solution for the low-computing power pulmonary tuberculosis lesion segmentation task based on chest radiography(CXR)in primary hospitals.