SEGMENTATION OF LIVER AND LIVER TUMOR BASED ON ATTENTION RESIDUAL

The long-term manual diagnosis of liver medical images is likely to cause fatigue to the doctor, leading to misdiagnosis and missed diagnosis. Aimed at the above phenomenon, an improved Unet network is proposed for automatic segmentation of liver and liver tumors. We improved the Unet model and introduced the attention residual structure and feature multiplexing structure to improve the utilization efficiency of the feature information in the input image. We improved the loss function and added the under-segmentation and over-segmentation penalty factors to the Dice coefficient to improve the model’s predictive ability. The experimental results on the public dataset show that the segmentation similarity coefficients of the algorithm in the liver and liver tumors reaches 0.962 and 0.713, respectively, which is better than the existing segmentation models and has strong robustness.