TB-YOLO-Mamba: A Lightweight Deep Learning Framework Based on State Space Models for Mycobacterium tuberculosis Detection

Target detection of Mycobacterium tuberculosis technology improves the accuracy and timeliness of tuberculosis diagnosis and reduces the risk of tuberculosis transmission. At present, the Mycobacterium tuberculosis detection models based on deep learning have many parameters, and the accuracy of the detection result needs to be improved. To achieve efficient detection of Mycobacterium tuberculosis, a lightweight model named TB-YOLO-Mamba is proposed in this study. Aiming at the problem of a mass of parameters and high computational complexity, C2f-Faster module is designed using the property of Partial Convolution (PConv) that only convolutions part of the input channels. The module minimizes unnecessary computations and memory accesses to effectively extract spatial features. To enhance the precision of Mycobacterium tuberculosis target recognition, the TB-YOLO-VSS double branch module is designed. Firstly, the local feature refinement branch combines the advantages of convolutional layers in capturing local features. Secondly, the global feature enhancement branch is designed by utilizing the excellent ability of State Space Models (SSMs) to revealing long dependencies. Finally, Coordinate Attention(CA) is introduced to capture the spatial relationship between the two branches. Experimental results show that compared with Faster R-CNN, SSD, RetinaNet, YOLOv5-CTS and other Mycobacterium tuberculosis target detection algorithms, the average accuracy of TB-YOLO-Mamba reaches 87.2%. Compared to the baseline, the number of TB-YOLO-Mamba parameters decreases by 20%, and FLOPs decreases by 13%. It is proved that this method can detect Mycobacterium tuberculosis more quickly and accurately.