Chest X-ray Lesion Detection Based on Improved YOLOv7

Chest x-ray examination is one of the important methods of clinical examination. Due to the complex structure of the chest, irregular lesions and other factors, it is difficult to clearly display the features, and the performance of chest X-ray lesion detection is limited. Therefore, this paper takes the characteristics of chest x-ray images as the starting point. According to the characteristics of chest x-ray lesions, the YOLOv7 algorithm has been improved. By introducing the MVB (MobileViT Block) module into the backbone feature extraction network, the correlation of multiple locations can be correlated, more global and accurate information can be extracted, and the irregular shape of the lesion can be effectively processed. At the same time, in view of the complex background noise of the chest radiograph image, the GAM (Global Attention Mechanism) is introduced in the feature pyramid fusion stage. This method can increase the importance of object focus and suppress the interference of background noise. Finally, through comparative experiments, this paper found that the improved chest x-ray lesion detection algorithm of mAP@.5 reached 0.62, and compared with the benchmark model YOLOv7 algorithm, the detection rate of the improved algorithm increased by 2.6%. In addition, the ablation experimental results show that the improvements made in this paper can effectively alleviate the problem in chest X-ray images.