Multi-source Unsupervised Domain Adaptation for Medical Image Recognition

Medical image recognition is pivotal in intelligent healthcare, especially when addressing complex diseases and human anatomical structures. Intelligent models can be trained using a vast number of labeled medical images, which enables the automatic detection of lesions. However, existing models often overlook the domain gap between the training data and the actual clinical environment, resulting in poor performance in new clinical settings. In this paper, we propose an adaptive dynamic multi-stage pseudo-labeling mechanism based on multi-source domain samples for generating pseudo-labels of the unlabeled target domain images. Additionally, we introduce a multi-source domain adaptation (MSDA) framework for medical image recognition with limited labeled training samples for a specific dataset. Training with multi-source samples enhances the model’s generalization and adaptability, in which the diverse samples enable the target model to learn more representative feature maps. Our method achieves high accuracy and robustness in medical image recognition, demonstrating strong adaptability and superior performance across various clinical scenarios.