Automatic Segmentation Method For CT Images Of Liver Tumors Based On Deep Learning

The most effective way to treat liver cancer is precision hepatectomy,the basic step of which is digital image segmentation of the liver and tumor.The traditional liver parenchyma segmentation method has disadvantages such as low accuracy,slow speed and large computational effort when performing liver segmentation.However,since the abdominal CT image contains not only the liver region,but also other organs such as kidney,gallbladder,spleen and stomach,which have similar contrast between them and have blurred borders with the liver,these can affect the segmentation results of the liver and liver tumors.Therefore,it is of great interest to design and implement a model with high robustness,automatic fine segmentation,and a simple structure with a small number of parameters.In this paper,automatic segmentation of CT images of liver and automatic segmentation of CT images of liver tumor are investigated respectively,and the main contents are as follows:(1)For the liver segmentation task,this paper proposes a segmentation network MSAA-Net combining multi-scale features and an improved attention-aware U-Net.We extracted features at different scales on a single feature layer and performed attention perception in the channel dimension.This study demonstrates that this architecture improves the performance of U-Net while greatly reducing the computational cost.To address the problem that U-Net’s jump connection is difficult to merge and optimize for objects of different sizes,we design a multiscale attention gate structure(MAG),which enables the model to automatically learn to focus on targets of different sizes.In addition,MAG can be extended to all structures containing jump junctions,such as variants of U-Net and FCN.The structure of this paper was evaluated on the 3Dircadb dataset,which has a DICE similarity coefficient of 94.42% for the liver and a model parametric number equivalent to only 38.4% of the Attention U-Net.The experimental results show that MSAA-Net achieves a very competitive performance.(2)For the tumor segmentation task,this paper proposes a liver tumor segmentation method based on the results of liver segmentation.The algorithm uses a contour extraction-based method to eliminate the background region outside the liver,and then uses Gaussian filtering and median filtering to eliminate the liver boundaries and noise points.The tumor region located inside the liver is segmented out by a threshold-based image segmentation method.Finally,the segmentation results were optimized by morphological operations.The liver tumor segmentation algorithm proposed in this paper was evaluated on the 3Dircadb dataset and finally achieved a DICE coefficient of 92.37%,which successfully improved the accuracy of the automatic segmentation work of liver tumor CT images.