基于Mask RCNN和U-Net结合的三阶段肾脏与肿瘤分割方法

A Three-Stage Kidney and Tumor Segmentation Method Based on the Combination of Mask RCNN and U-Net

针对腹部计算机断层扫描(Computed Tomography,CT)图像中,因组织细节对比度低,肾脏与肿瘤形状不规则造成自动分割困难的问题,提出了一种基于Mask RCNN和U-Net结合的三阶段肾脏与肿瘤分割方法.首先,利用Mask RCNN网络识别断层序列图像中的肾脏,记录肾脏出现和结束时的断层数,缩小目标范围;其次,进行肾脏与肿瘤的分割,汇总含有肿瘤的断层切片,采用以U-Net为基础,下采样增加密集连接,上采样使用双三次插值的网络,获得更准确的全局位置特征和局部细节特征;然后,再继续进行肿瘤分割,将结果与上一阶段融合;最后,使用基于三维连通域的方法进一步优化分割结果.实验结果表明,所提方法在KiTS19数据集上肾脏与肿瘤分割的平均Dice系数分别为0.95720和0.81636,与其他基于CNN的方法相比,在分割精度及准确率上均有所提升,有助于实现肾脏与肿瘤自动分割.

The automatic segmentation for abdominal computed tomography(CT)images was difficulty due to low contrast ratio of tissue details and irregular shapes of kidney and tumor.A three-stage kidney and tumor segmentation method based on the combination of Mask R-CNN and U-Net was proposed.Firstly,the kidneys in the tomographic sequence images was identified by using the Mask R-CNN network,and the number of slices was recorded when the kidney appeared and disappeared in CT image,so the target range was narrowed.Secondly,the kidney and tumor were segmented,and the tomographic slices containing the tumor were summarized.The more accurate global location features and local detail features were obtained by using the network which based on U-Net,and increased dense connections in down-sampling,and used the bicubic interpolation in up-sampling.Thirdly,the segmentation of the tumor was continued,and the results was amalgamated with the previous stage.Finally,the segmentation results were further optimized by using the method based on three-dimensional connected domains.Experiments show that the proposed method has an average Dice coefficient of 0.95720 and 0.81636 for kidney and tumor segmentation on the KiTS19 dataset.Compared with other CNN-based methods,the segmentation accuracy is improved,which helps to achieve the automatic segmentation of kidney and tumor segmentation.