Name
Abstract | ||
|---|---|---|
•We introduce a new task named Noisy Supervised Domain Adaptation (NSDA), which is both challenging and meaningful for real-world applications in medical image segmentation.•We propose a novel self-cleansing unsupervised domain adaptation (S-CUDA) method to address the NSDA problem for medical image segmentation, which trains two peer adversarial networks simultaneously in a co-training and co-cleansing style. For each peer network, we divide the training samples into the high-confidence clean labeled set (i.e., clean data) and noisy labeled set (i.e., noisy data) based on its per-sample loss. The identified clean data are then used to train the peer networks, which keep diverging to another one, to filter different type of errors and avoid confirmation bias in the self-training.•We propose to clean and reuse the identified high-confidence noisy labeled samples to provide more diversity and useful information to learn a robust model. For the identified high-confidence noisy data, we clean and refine their labels using the network’s predictions as pseudo labels, which will be further reused in the following learning iterations.•We design a novel loss function to use both the original noisy labels and the updated cleaned labels from the peer network to break up the chain of “error propagation”. The new loss function can help the cross-reviewing framework alleviate the “learning trap” problem exiting in self-training process (i.e., the quality of pseudo labels does not get improved and the new supervision from them becomes weaker and weaker).•We conduct extensive experiments on public datasets, i.e., two retinal fundus image datasets and a public spinal cord gray matter segmentation challenge dataset, to demonstrate the effectiveness of our proposed S-CUDA framework. The comprehensive results demonstrate the superior performance of our method compared with existing state-of-the-art methods. |
Year | DOI | Venue |
|---|---|---|
2021 | 10.1016/j.media.2021.102214 | Medical Image Analysis |
Keywords | DocType | Volume |
Medical image segmenation,Noisy labels,Domain shift,Cross denoising,Label correction,Adversarial learning | Journal | 74 |
ISSN | Citations | PageRank |
1361-8415 | 0 | 0.34 |
References | Authors | |
2 | 5 | |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Luyan Liu | 1 | 0 | 0.34 |
| Zhengdong Zhang | 2 | 0 | 1.01 |
| Shuai Li | 3 | 175 | 31.37 |
| Kai Ma | 4 | 49 | 18.48 |
| Yefeng Zheng | 5 | 1391 | 114.67 |