Shape and boundary-aware multi-branch model for semi-supervised medical image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Supervised medical image segmentation requires extensive labeled data, which is often scarce and costly to obtain.
• Limited labeled data leads to poor model performance, including overfitting and low accuracy in medical image analysis.
• Existing methods struggle with the labor-intensive nature of medical image annotation.

Purpose of the Study:

• To develop a novel shape and boundary-aware deep learning model for medical image segmentation.
• To leverage semi-supervised learning to effectively utilize both labeled and unlabeled data.
• To enhance segmentation accuracy, especially in boundary regions, by incorporating shape and distance map information.

Main Methods:

• A V-Net architecture was employed for Pixel-wise Segmentation Map (PSM) and Signed Distance Map (SDM) prediction.
• Multi-scale features were enhanced around target boundaries using SDM information and a Pyramid Pooling Module (PPM).
• Task consistency loss was applied to PSM and SDM outputs to effectively mine unlabeled data.

Main Results:

• The proposed model demonstrated superior performance across three challenging medical image datasets (LA2018, BraTS2019, ISIC2018).
• The boundary-aware features and SDM integration improved segmentation accuracy, particularly in complex regions.
• The semi-supervised approach significantly enhanced model generalization compared to purely supervised methods.

Conclusions:

• The novel semi-supervised model offers a practical and superior solution for medical image segmentation with limited labeled data.
• Integrating shape and boundary information effectively addresses segmentation challenges in medical imaging.
• The model's ability to utilize unlabeled data holds significant promise for advancing automated medical image analysis.