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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Accurate abdominal organ and tumor segmentation is critical for clinical applications like surgery and radiotherapy.
• Challenges include organ/tumor diversity and partially labeled datasets, hindering segmentation accuracy.
• Existing methods often address missing labels but neglect network-level improvements.

Purpose of the Study:

• To introduce AMOTS, a cascaded framework for precise abdominal multi-organ and pan-cancer tumor segmentation.
• To enhance feature extraction and segmentation accuracy, particularly in challenging scenarios with limited labels.
• To improve recognition of unlabeled classes through novel training strategies.

Main Methods:

• A cascaded framework employing a lightweight convolutional network for initial localization.
• Two Aspect-Aware Complementary Networks (AACNet) for fine segmentation, featuring Directional Separation Focus Module (DSFM) and Multi-View Slice Cross Attention Module (MVSCM).
• Ambiguity hard mining and pseudo-label supervision strategies to address label imbalance and enhance unlabeled class recognition.

Main Results:

• AMOTS demonstrated superior segmentation accuracy compared to existing methods on large public datasets (FLARE2023 and MOTS).
• The proposed DSFM and MVSCM modules effectively improved boundary recognition and global interaction.
• The ambiguity hard mining and pseudo-label supervision strategies enhanced performance on imbalanced datasets.

Conclusions:

• The AMOTS framework offers a significant advancement in abdominal multi-organ and tumor segmentation.
• The novel AACNet architecture and training strategies effectively address segmentation challenges, including data scarcity and imbalance.
• AMOTS provides a robust and accurate solution for clinical applications requiring precise medical image segmentation.