Data privacy protection domain adaptation by roughing and finishing stage

• Main Authors: Yuan, Liqiang, Erdt, Marius, Li, Ruilin, Siyal, Mohammed Yakoob
• 其他作者: School of Electrical and Electronic Engineering
• 格式: Article
• 語言: English
• 出版: 2023
• 主題: Engineering::Electrical and electronic engineering; Unsupervised Domain Adaptation; Model Generalization
• 在線閱讀: https://hdl.handle.net/10356/172253

The automatic segmentation of organs or tissues is crucial for early diagnosis and treatment. Existing deep learning methods either need massive annotation data or use Unsupervised Domain Adaptation (UDA) approaches with labeled source domain data to train a model for unlabeled target domain data. The methods mentioned above require accessing the data of the source domain. However, in medical imaging, source domain data from a hospital is usually restricted due to patient data privacy regulations. Therefore, it is crucial to perform privacy-preserving domain adaptation, which simultaneously improves the model’s performance on target domain data and preserves the privacy of source domain data. Aiming to achieve Source Free Unsupervised Domain Adaptation (SFUDA), we propose a two-stage framework Roughing and Finishing to extract the relevant features and knowledge from the pre-trained model and protect the patient’s privacy. Specifically, in the Roughing stage, batch statistics are updated using unlabelled target domain data to alleviate feature distribution shifts. Then, a non-parametric weighted entropy-minimization is used to reduce the domain gap further. In the Finishing stage, we aim to enhance the model’s generalization ability by increasing the noise robustness of the feature representation, leading to a more generalizable target feature representation. Besides, we propose pseudo-label training based on a confidence-weighted feature distance to utilize the knowledge from confident samples. The proposed method is evaluated on two medical domain transfer challenges, (1) MRI to CT liver segmentation and (2) cross-domain fundus image segmentation. Extensive experiments and ablation studies demonstrate the superiority of the proposed methods over state-of-the-art SFUDA methods by a large margin. The proposed method outperforms previous UDA approaches that access the source domain data.