Research on CT Images of Lung Lesions Based on Variable Convolution Kernel

CT image segmentation of lung lesions has a wide range of research prospects and application value in the fields of clinical medicine and computer-aided diagnostic technology. However, considering that the CT images of lungs have problems such as unclear margins of the lesion area, irregular shapes, and large differences in the CT images of lungs between different individuals, it is still a challenge to segment the CT images of lungs lesions in a more accurate way. In this paper, an improved CT image segmentation algorithm for lung lesions based on U-Net (AC-Net) is proposed. Firstly, based on the U-Net architecture, the Transformer module is introduced, which utilizes the global attention mechanism of the Transformer module to obtain global information more effectively, considering that the traditional CNN structure has limited sensory field and is not good at establishing remote information connections. Secondly, the model adds a global attention mechanism (GAM) in the decoder part, which reduces and amplifies global interactions by retaining spatial and channel information to reduce information. Finally, the convolutional kernel used in the model is Variable Kernel Convolution (AKConv), which refines the segmentation results through flexible convolutional shapes to better capture the edge shapes of the image. The model was compared with experiments on the covid19-ct-scans public dataset, and the experimental results showed the effectiveness of the improved method in this paper, and demonstrated higher segmentation accuracy on CT images of lung lesion regions, which provides a new reference for CT image segmentation of lung lesion regions.