A cascading approach using se-resnext, resnet and feature pyramid network for kidney tumor segmentation

Abstract

Accurate segmentation of kidney tumors in CT images is very important in the diagnosis of kidney cancer. Automatic semantic segmentation of the kidney tumor has shown promising results to wards developing advance surgical planning techniques in the treatment of kidney tumor. However, the relatively small size of kidney tumor volume in comparison to the overall kidney volume, and its irregular distribution and shape makes it difficult to accurately segment the tu mors. In addressing this issue, we proposed a coarse to fine segmentation which leverages on transfer learning using SE-ResNeXt model for the initial segmentation and ResNet and Feature Pyramid Network for the final segmentation. The processes are related and the output of the initial results was used for the final training. We trained and evaluated our method on the KITS19 dataset and achieved a dice score of 0.7388 and Jaccard score 0.7321 for the final segmentation demonstrating promising results when compared to other approaches.