CK-ATTnet: Medical image segmentation network based on convolutional kernel attention

The medical image partition model has a wide range of application prospects in medical diagnosis and treatment and has become an important auxiliary method to improve the diagnostic level by medical imaging analysis. After the feature extraction ability of the convolutional neural network (CNN) reached a bottleneck, the form of feature extraction represented by Transformer has made significant achievements in the medical image domain in recent years. However, the structure of Transformer is relatively fixed, the cost of computer resources is large, and it is difficult to adjust the model structure according to the complex medical imaging segmentation task. To better adapt to the limitation of clinical diagnostic equipment on the parameter scale of the network model, this paper proposed a CK-ATTnet based on the convolutional kernel attention mechanism. CK-ATTnet uses the depthwise separable convolution attention mechanism, which completely innovates the way that the attention mechanism is used on the original image in the traditional model. In addition, CK-ATTnet realizes the new normal form of applying the attention mechanism to the convolutional kernel for feature extraction for the first time. This design further improves the local feature acquisition ability of the convolutional kernel and does not require additional hardware enhancement requirements after applying the attention mechanism to the model. Compared with other CNN models, CK-ATTnet can extract more accurate and fine-grained features; compared with Transformer-based models, it has fewer learning parameters. Experimental results show that CK-ATTnet exhibits better segmentation performance and fewer learning parameters than other models in multiple datasets and has very good application prospects.