Pancreas segmentation is very difficult since pancreas occupies only a very small fraction less than 0.5% of a CT volume and suffers from high anatomical variability. Most existing methods use a two-stage framework: the coarse and the fine. We argue that both stages have the same purpose of improving pancreatic-pixel classification accuracies. Inspired by this observation, we transfer fine-model weights to the coarse. If we directly copy the pre-trained fine model, the performance is low due to the domain gap of the different input images, so we further propose a momentum update strategy for transferring models. Our momentum update stands on the other observation that input images are in three different domains: the small image cropped by the ground-truth bounding box (
bash f2.sh
We appreciate it if you cite the following paper:
@InProceedings{TangMICCAI2022,
author = {Yumou Tang and Kun Zhan and Zhibo Tian and Mingxuan Zhang and Saisai Wang and Xueming Wen},
title = {Curriculum knowledge switching for pancreas segmentation},
booktitle = {ICIP},
year = {2023}
}
If you have any questions, feel free to contact me. (Email: ice.echo#gmail.com
)