Unsupervised domain adaptation for medical image segmentation by pre-training and 3D multi-scale feature fusion

Medical image segmentation is crucial for modern healthcare, yet traditional methods often rely on large annotated datasets that are difficult to obtain. To address this, we propose a pre-training and 3D multi-scale feature fusion framework (PreMS-UDA) that enhances cross-modality segmentation by effectively utilizing 3D image information, overcoming the limitations of existing 2D-based UDA approaches. Our method first performs encoder pre-training separately on source and target domain images, incorporating a group quantization module to preserve more feature information. In the domain adaptation stage, a 3D multi-scale feature fusion module enables the model to learn both 2D domain and 3D spatial information by fusing features from multiple slices. Experimental results show that PreMS-UDA outperforms state-of-the-art UDA methods in abdominal multi-organ segmentation and prostate segmentation, demonstrating its significant improvement in medical image segmentation. Our code is provided at the following repository: https://github.com/Aya-natsume/PreMSUDA . • Introduces a novel framework for domain adaptation in cross-modality segmentation. • Utilizes group quantization to retain complex anatomical structural features. • Integrates 3D multi-scale feature fusion for improving segmentation performance. • Outperforms state-of-the-art methods in abdominal and prostate segmentation.