Mutual enhancing learning-based automatic segmentation of CT cardiac substructure.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Current thoracic cancer radiotherapy (RT) segmentation treats the whole heart as one organ, increasing cardiac toxicity risks.
• Manual segmentation of cardiac substructures is time-intensive and complex due to anatomical variations.

Purpose of the Study:

• To develop an accurate and automatic deep learning (DL) strategy for segmenting cardiac substructures.
• To specifically improve the segmentation of smaller substructures, such as coronary arteries.

Main Methods:

• A novel DL-based mutual enhancing strategy using three subnetworks: Retina U-net, classification, and segmentation modules.
• The method was evaluated on three diverse datasets (institutional, MM-WHS, ACDC) with manual contouring and comparison against four other network architectures.
• Segmentation accuracy was assessed using Dice similarity coefficient, Jaccard index, Hausdorff distance, and other metrics.

Main Results:

• The proposed DL method achieved significantly higher segmentation accuracy for small cardiac substructures, particularly coronary arteries (e.g., CA-LADA, CA-RCA).
• Comparable results to mask scoring R-CNN were obtained for large substructures (heart chambers), outperforming 3D U-net and mask R-CNN.
• The method demonstrated improved accuracy for small substructures without significantly compromising large substructure segmentation.

Conclusions:

• A new DL-based mutual enhancing strategy enables automatic and accurate segmentation of cardiac substructures.
• This approach enhances the segmentation of critical small substructures like coronary arteries, crucial for reducing cardiac toxicity in RT.
• The method offers a potential tool for rapid, physician-reviewed substructure segmentation, improving clinical workflow efficiency.