Deep learning-based automatic segmentation of cardiac substructures for lung cancers.

Area of Science:

• Radiotherapy and Medical Imaging
• Artificial Intelligence in Healthcare
• Cardiovascular Anatomy

Background:

• Accurate segmentation of cardiac substructures is vital for minimizing radiation-induced heart disease during lung cancer radiotherapy.
• Current manual segmentation methods are time-consuming and prone to inter-observer variability.

Purpose of the Study:

• To develop and validate deep learning-based auto-segmentation models for nineteen cardiac substructures.
• To assess the accuracy and clinical acceptability of AI-driven cardiac substructure delineation.

Main Methods:

• An nnU-Net auto-segmentation model was trained on 100 non-small cell lung cancer patients with manually delineated cardiac substructures.
• Model performance was evaluated using Dice Similarity Coefficient (DSC) and dose metrics on an independent dataset of 42 patients.
• Subjective evaluation by four physicians assessed the clinical acceptability of auto-segmented contours.

Main Results:

• The AI model achieved high average DSCs for major cardiac structures (e.g., 0.95 for whole heart, 0.91 for chambers).
• Average absolute errors in mean/max doses to substructures were within acceptable clinical ranges.
• 94% of auto-segmented contours were rated as clinically acceptable by physicians.

Conclusions:

• The developed nnU-Net model effectively delineates cardiac substructures, including coronary arteries, with high accuracy.
• This AI approach shows significant promise for improving radiation dose assessment to cardiac substructures in lung cancer patients.
• The model can aid in reducing the risk of radiation-induced heart disease through precise auto-segmentation.