Multi-Scale Adversarial Learning and Difficult Supervision for Kidney and Kidney Tumor Segmentation
Shenhai Zheng (Chongqing University of Posts and Telecommunications),* Qiuyu Sun (Chongqing University of Posts and Telecommunications), Xin Ye (Chongqing University of Posts and Telecommunications), Weisheng Li (Chongqing University of Posts and Telecommunications), Laquan Li (Chongqing University of posts and telecommunications)
The 33rd British Machine Vision Conference
Abstract
Automatic segmentation of kidney and kidney tumor areas plays an important role in radiotherapy and clinical practice. In recent years, deep learning methods have been widely used in the segmentation task and have achieved remarkable achievements. However, automatic segmentation of kidney and kidney tumors is still challenging due to their diverse shapes, complex types and unpredictable locations. Inspired by the deep supervision strategy, in this paper, we propose a cascade based approach with multi-scale adversarial learning and difficult supervision to address these challenges. On the whole, the proposed method follows the typical cascade strategy, where coarse segmentation is performed first and then fine segmentation is implemented. For the coarse segmentation part, we use Res-UNet to obtain regions of interest for kidneys and masses (include tumor and cyst). In the fine segmentation part, we propose a Multi-Scale Adversarial Learning Difficulty Supervised UNet (MSALDS-UNet) as our fine-segmented network, which consists of a segmentation network and multiple discriminators. It applies adversarial learning strategies at multiple scales of the segmentation network to improve the final segmentation performance. This is similar to the motivation of deep supervision. In addition, we also propose a difficult region supervised loss applied in MSALDS-UNet to utilize the structured information to better segment hard-to-segment regions such as fuzzy boundaries. This proposed approach can transform multi-class segmentation tasks into multiple simple binary segmentation problems. A thorough validation on the dataset provided by the 2021 Kidney and Kidney Tumor Segmentation Challenge (KiTS21) shows that our model achieves satisfactory results in kidney and kidney tumor segmentation.Video
Citation
@inproceedings{Zheng_2022_BMVC,
author = {Shenhai Zheng and Qiuyu Sun and Xin Ye and Weisheng Li and Laquan Li},
title = {Multi-Scale Adversarial Learning and Difficult Supervision for Kidney and Kidney Tumor Segmentation},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0879.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by