Abstract:

Accurate segmentation of liver and lesions is an important premise for computer-aided doctors to diagnose liver cancer and formulate corresponding treatment plans. Aiming at the problem of high-precision segmentation algorithm that the process is complicated and the segmentation needs to be completed in multiple steps, end-to-end U-net segmentation algorithm for predicting liver and lesion areas is proposed. Firstly, it improves the basic U-net model and adds feature reuse ideas to improve the efficiency of the network model for feature utilization. Secondly, the loss function is improved, and the cross entropy and the Dice coefficient are combined. At the same time, the under-division penalty factor is fine-tuned to improve the model effect and the detection ratio of the lesion. Finally, the optimization results are processed before and after the addition. Experiments in the MICCAI_2017_LiTS dataset show that the end-to-end network can still achieve the algorithmic accuracy of complex networks and multi-step segmentation networks.

Key words: liver, lesion, CT image, segmentation algorithm

摘要：

准确的肝脏病灶分割是计算机辅助医生进行肝癌诊断和制定相应治疗计划的重要前提，针对高精度分割算法普遍存在的过程复杂、分割需多步完成的问题，提出了一种同时预测肝脏区域和病灶区域的端到端U-net分割算法。改进基础U-net模型，加入特征复用思想以提高网络模型对于特征的利用效率；对损失函数进行改进，采用交叉熵和Dice系数相结合，同时加入欠分割惩罚因子微调，提高模型预测能力及病灶的检出比例；加入[[-200,200]]的阈值化预处理和提取肝脏最大联通域、删除病灶扁平预测结果的后处理来优化结果。在MICCAI_2017_LiTS数据集的实验表明，端到端网络依然可以达到复杂网络、多步分割网络的算法精度。

关键词: 肝脏, 病灶, CT图像, 分割算法