Category-level and Instance-level Synergistic Alignment Applied in Unsupervised Domain Adaptation for Pancreatic Segmentation

In medical image segmentation, significant feature differences exist between medical images with different acquisition parameters or modes, leading to decreased performance of deep neural networks in clinical scenarios. Specifically, in pancreas organ segmentation, the organ’s intricate morphology, the high cost of manual annotation by medical professionals, and challenges posed by multimodal imaging data contribute to the difficulty of pancreas organ segmentation. Therefore, this paper proposes an unsupervised domain adaptation (UDA) approach aimed at leveraging publicly available source domain data with rich label information for pancreas organ segmentation, thereby improving segmentation performance in the target domain. The proposed method is based on Generative Adversarial Networks (GANs) and achieves source-to-target domain image translation through image-to-image transformation and feature alignment. This leads to an end-to-end unsupervised domain adaptation framework for pancreas organ segmentation. The method utilizes a feature fusion module to explore domain-invariant features and employs two orthogonal classifiers for category alignment. Through adversarial learning and deep supervision mechanisms, it integrates feature alignment and category alignment to enhance segmentation performance. In the experimental section, we utilized two publicly available datasets as the source and target domains and validated the effectiveness of the proposed method through performance comparison experiments. The experimental results demonstrate a significant improvement in segmentation performance on unlabeled target images. Specifically, the unsupervised semantic segmentation achieved a Dice coefficient of 62.2% on the publicly available datasets. This provides valuable reference and guidance for research and application in the field of medical image segmentation.