Two-Stage Segmentation Framework Based on Distance Transformation.

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Anatomy

Background:

• Deep learning aids clinical analysis through lesion segmentation and diagnosis.
• Partial volume effects blur medical image edges, hindering accurate segmentation of organs and lesions.
• Accurate segmentation is crucial for reliable medical image analysis.

Purpose of the Study:

• To develop a novel framework for improving medical image segmentation accuracy, specifically for left atrium MRI.
• To investigate the utility of distance maps as weight maps to enhance network focus on Region of Interest (ROI) edges.
• To explore various distance map generation methods for optimizing network learning.

Main Methods:

• A novel deep learning framework was designed to embed distance maps into a two-stage network.
• Distance transformation was applied to ROI edges to create weight maps.
• Multiple distance map generation techniques were proposed and evaluated.
• The framework was applied to left atrium MRI segmentation.

Main Results:

• The proposed framework effectively improved left atrium MRI segmentation performance.
• Experimental results validated the hypothesis that distance maps enhance learning of ROI edge regions.
• The integration of distance maps as weight maps proved feasible and effective.

Conclusions:

• The novel deep learning framework incorporating distance maps is effective for enhancing medical image segmentation.
• This approach successfully addresses challenges posed by blurred edges due to partial volume effects.
• The method shows significant potential for improving diagnostic accuracy in clinical medical analysis.