Abstract:

Aiming at the problem of insufficient expression ability of liver tumor features and limited global contextual information transmission, an improved UNet liver tumor segmentation method is proposed. Firstly, a low rank reconstruction convolution is designed to optimize the large number of parameter problems caused by traditional convolution operations, and use it to construct a convolution kernel reconstruction module that uses residual structure to improve the encoder decoder, so that the encoder retains more detailed information and the decoder recovers information more effectively, thereby enhancing the expression ability of liver tumor features. Then, to enrich the transmission of global contextual information, a three branch spatial pyramid pooling module is designed to optimize the bottleneck structure of information transmission and break the limitation of a single path. Secondly, a multi-scale feature fusion module is designed to optimize the reuse mechanism of encoder information, enhance the modeling ability of the model for global contextual information, and improve its efficiency in extracting liver tumor features at different scales. Finally, the performance of our method was tested on the LiTS2017 and 3DIRCADb datasets. The experimental results show that our method achieves Dice and IoU values of 97.56% and 95.25% in the liver segmentation task on the LiTS2017 dataset, and 89.71% and 81.58% in the liver tumor segmentation task, respectively. The Dice and IoU values in the liver segmentation task of the 3DIRCADb dataset reached 97.63% and 95.39%, respectively, while the Dice and IoU values in the liver tumor segmentation task reached 89.62% and 81.63%, respectively. This method can effectively alleviate the problem of insufficient expression ability of liver tumor features, and further enhance the model's ability to capture global contextual information.

Key words: liver tumor segmentation, convolutional kernel reconstruction, spatial pyramid pooling, multi scale feature fusion