基于深度监督残差网络的肝脏及肝肿瘤分割

The Liver and Liver Tumor Segmentation Based on Deeply Supervised Residual Unet

针对医生手动对肝脏肿瘤CT图像分割耗时、耗力,且易受主观判断影响的问题,该研究提出一种深度监督残差网络(Deeply Supervised Residual Unet,DS-ResUnet)算法,以实现对腹部增强CT图像中肝脏及肝脏肿瘤区域进行全自动分割的目的。首先,利用公开发布的MICCAI2017肝脏肿瘤分割(LiTS)挑战赛数据集,并使用python及TensorFlow开源框架进行数据分析;然后,构建深度监督残差网络对肝脏及肝肿瘤图像进行自动分割;最后,通过平均Dice系数、全局Dice系数、Jaccard系数、平均对称表面距离(ASSD)、95%豪斯多夫距离(HD95)、准确率和召回率七个评价指标对所提出算法与Unet模型的性能进行比较分析。结果显示,所提出的DS-ResUnet算法在肝脏分割上的七个评价指标结果依次为96.06%、95.08%、92.54%、1.98 mm、12.87 mm、96.11%、96.06%,优于Unet模型的结果(95.71%、94.52%、91.91%、2.41 mm、14.21 mm、95.48%、96.01%);在肝肿瘤分割上的七个评价指标结果依次为67.51%、76.65%、54.21%、6.65 mm、25.34 mm、80.39%、64.27%,也优于Unet模型的结果(60.67%、73.47%、47.39%、9.43 mm、39.38 mm、79.61%、58.01%)。这表明所提出的算法有效地提高了分割效果,实现了从3D腹部增强CT图像中全自动分割肝脏和肝肿瘤区域的目的。

For the problem that doctors manually segment the liver tumor from CT image is time-consuming,labor-intensive,and susceptible to subjective judgment,we propose a deeply supervised residual Unet(DSResUnet)that incorporates residual link and deep supervision into Unet for more precise segmentation.The proposed method was evaluated on the public MICCAI 2017 liver segmentation(LiTS)challenge dataset with Dice coefficient,Jaccard coefficient,average symmetrical surface distance(ASSD),95%Hausdorff distance(HD95),precision and recall.The experimental results show that the results on the above 7 evaluation indicators of liver segmentation with the proposed DS-ResUnet are 96.06%,95.08%,92.54%,1.98 mm,12.87 mm,96.11%,and 96.06%,respectively,achieve superior results on almost all metrics to the widely-used Unet(95.71%,94.52%,91.91%,2.41 mm,14.21 mm,95.48%,96.01%).The results on the above 7 evaluation indicators of liver tumor segmentation with the proposed DS-ResUnet are 67.51%,76.65%,54.21%,6.65 mm,25.34 mm,80.39%,and 64.27%,respectively,also better than that of the Unet(60.67%,73.47%,47.39%,9.43 mm,39.38 mm,79.61%,58.01%).Therefore,the proposed DS-ResUnet improves the segmentation results and achieves automatic segmentation of liver and liver tumor regions from the 3D abdominal enhanced CT image.