Coarse to Fine Segmentation Method Enables Accurate and Efficient Segmentation of Organs and Tumor in Abdominal CT

Hui Meng, Haochen Zhao, Deqian Yang, wang song ping, Zhenpeng Li

Keywords: segmentation, abdominal CT, coarse to fine

Abstract: Automatic segmentaion of organs and tumor in abdominal CT scans is essential for cancer diagnosis and treatment monitoring. However, there does not exist an accurate and efficient method for universal organ and tumor segmentation in abdominal CT scans. Therefore, we propose a coarse to fine segmentation (CFS) method based on pseudo labels. Specifically, the CFS consists of a coarse segmentation model (CSM), a tumor segmentation model (TSM), and an organ segmentation model (OSM). The CSM is trained to segment abdominal regions in CT scans. The TSM and the OSM are trained to generate segmentation masks of organs and tumor. The outputs of the TSM and the OSM are merged to generate the final segmentation results. To improve efficiency of the CFS, we optimize the inference process by streamlining intricate steps. On validation set of FLARE23 challenge, our method achieves mean DSC of 91.59% and mean NSD of 95.74% on organ segmentation, and mean DSC of 47.12% and mean NSD of 39.94% on tumor segmentation. The mean inference time is 24.12s, and the mean area under the GPU memory-time curve is 39543.46MB.

Submission Number: 7