AMOTS: Partially supervised framework for abdominal multi-organ and tumor segmentation via aspect-aware complementary

Achieving precise segmentation of the abdominal multi-organ and pan-cancer tumors is crucial for disease diagnosis and treatment planning in clinical fields such as surgery and radiotherapy. However, the diversity of abdominal organs and tumor types, as well as partially labeled datasets, significantly increases the difficulty in training and improving segmentation accuracy. While existing methods attempt to focus on the segmentation of multiple organs and lesions simultaneously, they mainly concentrate on addressing the issue of missing labels but do not simultaneously consider improvements at the network level. Thus, this paper introduces a cascaded framework, AMOTS. First, to enhance the model's feature extraction and analysis capabilities in various extreme scenarios, the framework employs a lightweight convolutional network in the first stage to rapidly localize regions of interest, followed by two structurally identical Aspect-Aware Complementary Networks (AACNet) in the second stage for fine segmentation of organs and tumors, respectively. AACNet includes two novel modules: the Directional Separation Focus Module (DSFM) for multi-directional boundary recognition and the Multi-View Slice Cross Attention Module (MVSCM) for global interaction enhancement. Secondly, ambiguity hard mining and pseudo-label supervision strategies are introduced to tackle the challenge posed by label imbalances. This approach progressively enhances the model's recognition accuracy for unlabeled classes during training. Extensive experiments on large multi-source public datasets (FLARE2023 and MOTS) demonstrate that AMOTS surpasses other methods in segmentation accuracy. Our code is available at www.github.com/zzm3zz/AMOTS.