Cardiac substructure segmentation with deep learning for improved cardiac sparing.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiation Oncology

Background:

• Radiation-induced heart disease is a risk in radiation therapy.
• Cardiac substructures are crucial for accurate dose calculation but poorly visualized on CT.
• Current methods lack efficiency and accuracy in cardiac substructure segmentation.

Purpose of the Study:

• To develop a novel deep learning (DL) pipeline for cardiac substructure segmentation.
• To leverage MRI's soft tissue contrast with CT for improved visualization.
• To enable state-of-the-art segmentation using a single, non-contrast CT input for radiation therapy planning (RTP).

Main Methods:

• A 3D deep learning model was trained on paired cardiac MRI and CT data from 25 breast cancer patients.
• Ground truth delineations of 12 cardiac substructures were established using MR/CT registration.
• The model incorporated deep supervision and a Dice-weighted multi-class loss function, with post-processing using 3D conditional random field (CRF).

Main Results:

• The DL model achieved accurate segmentations for chambers (DSC=0.88), great vessels (DSC=0.85), and pulmonary veins (DSC=0.77).
• Mean distance to agreement across substructures was below 2.0 mm.
• DL provided improved segmentation compared to a multi-atlas method, with faster contour generation (~14 seconds per patient).

Conclusions:

• Deep learning offers significant efficiency and accuracy gains for cardiac substructure segmentation.
• This approach has high potential for integration into radiation therapy planning.
• Improved cardiac substructure segmentation can lead to better cardiac sparing and reduced radiation-induced heart disease.