CSMViT: A Lightweight Transformer and CNN fusion Network for Lymph Node Pathological Images Diagnosis

To address the burdensome and time-consuming nature of manual diagnosis of pathological sections, this study proposes an automated pathological image detection system. This system can directly detect pathological images and accurately locate lesion tissues, providing a reference for pathological diagnosis. We propose an improved MobileViT model for feature extraction in the system, which we have named CSMViT. Considering the complexity and multi-scale characteristics of pathological images, we made three significant modifications to the MobileViT model. First, the original MV2 module was replaced with an improved Ghost module to reduce the model's parameter count, enhance detection accuracy, and accelerate inference speed. Second, we improved the backbone structure of the network to achieve multi-scale feature learning, which not only further reduces the parameter count but also allows for more effective capture of features at different scales. Lastly, we introduced a new CSA module that can simultaneously accept two feature maps of different sizes as input. Through internal attention mechanisms and feature fusion, this module achieves cross-scale feature learning. Experimental results indicate that the CSMViT model achieved accuracy, F1-score, and specificity of 99.42%, 99.4%, and 99.6%, respectively. Additionally, the detection accuracy of CSMViT for the entire pathological image is 84%, representing an 8% improvement over the original network. Notably, the FLOPs of CSMViT is 1.461G, which is a 72.19% reduction compared to the original network, significantly decreasing the model's complexity. These results thoroughly demonstrate the effectiveness and substantial value of CSMViT in pathological image detection.