topK dice loss for medical image segmentation
Seyed mohsen hosseini (University of Tehran, University of Tehran)
Seyed mohsen hosseini (University of Tehran, University of Tehran)
The 35th British Machine Vision Conference
Abstract
In image segmentation tasks, the class-imbalance problem affects the performance of neural networks. This problem is especially severe in medical images where usually the task involves segmenting a small foreground in a large background. Different methods have been developed to address this problem, e.g. using region based losses like the Dice loss. There is also a difficulty-imbalance problem, because the majority of samples of each class are easy to classify and hard samples are in minority. This leads to an ineffective learning, as the easy samples dominate the training process. A common strategy to address this problem, is reweighting the cross entropy loss, in order to focus the training on hard samples. A novel loss for addressing both the class-imbalance and difficulty-imbalance problems is proposed in this work. Unlike previous methods that address these problems separately, the proposed method tackles them at the same time using a modified dice loss. Experiments on different medical segmentation tasks show that the proposed method outperforms popular existing methods for addressing class and difficulty imbalance problems.Citation
@inproceedings{hosseini_2024_BMVC,
author = {Seyed mohsen hosseini},
title = {topK dice loss for medical image segmentation},
booktitle = {35th British Machine Vision Conference 2024, {BMVC} 2024, Glasgow, UK, November 25-28, 2024},
publisher = {BMVA},
year = {2024},
url = {https://papers.bmvc2024.org/0897.pdf}
}
Copyright © 2024
The British Machine Vision Conference is organised by