Learning With Explicit Shape Priors for Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Deep Learning

Background:

• Medical image segmentation is crucial for analysis and surgical planning.
• UNet-based networks dominate but struggle with limited receptive fields, hindering long-range dependency modeling for organs or tumors.
• Existing methods fail to simultaneously address limited receptive fields and segmentation head dependency.

Purpose of the Study:

• To introduce a novel shape prior module (SPM) to enhance UNet-based medical image segmentation.
• To explicitly incorporate global and local shape priors to improve segmentation performance.
• To provide a versatile module that can be integrated into various network architectures.

Main Methods:

• Proposed a novel shape prior module (SPM) incorporating explicit global and local shape priors.
• Global shape priors offer coarse representations for modeling long-range contexts.
• Local shape priors provide finer guidance, reducing reliance on the segmentation head.

Main Results:

• The proposed SPM achieved state-of-the-art performance on three challenging public datasets.
• SPM effectively models long-range dependencies and alleviates the dependence on learnable prototypes in segmentation heads.
• Demonstrated the plug-and-play capability of SPM with both CNNs and Transformer-based backbones.

Conclusions:

• The novel shape prior module (SPM) significantly improves medical image segmentation accuracy.
• SPM offers a flexible and effective solution for enhancing existing segmentation models.
• The explicit introduction of shape priors is a promising direction for future medical image analysis research.