Towards Generic Abdominal Multi-Organ Segmentation with multiple partially labeled datasets

Abstract

An increasing number of publicly available datasets have facilitated the exploration of building universal medical segmentation models. Existing approaches address partially labeled problem of each dataset by harmonizing labels across datasets and independently focusing on the labeled foreground regions. However, significant challenges persist, particularly in the form of cross-site domain shifts and the limited utilization of partially labeled datasets. In this paper, we propose a GAMOS (Generic Abdominal Multi-Organ Segmentation) framework. Specifically, GAMOS integrates a self-guidance strategy to adopt diffusion models for partial labeling issue, while employing a self-distillation mechanism to effectively leverage unlabeled data. A sparse semantic memory is introduced to mitigate domain shifts by ensuring consistent representations in the latent space. To further enhance performance, we design a sparse similarity loss to align multi-view memory representations and enhance the discriminability and compactness of the memory vectors. Extensive experiments on real-world medical datasets demonstrate the superiority and generalization ability of GAMOS. It achieves a mean Dice Similarity Coefficient (DSC) of 91.33% and a mean 95th percentile Hausdorff Distance (HD95) of 1.83 on labeled foreground regions. For unlabeled foreground regions, GAMOS obtains a mean DSC of 86.88% and a mean HD95 of 3.85, outperforming existing state-of-the-art methods.