Domain-adapted foundation model for automated cardiac CT substructure segmentation for thoracic radiotherapy

Area of Science:

• Medical imaging
• Radiotherapy
• Artificial intelligence

Background:

• Radiation-induced cardiotoxicity is a major cause of non-cancer mortality in patients receiving thoracic radiotherapy.
• Dose to specific cardiac substructures, not just whole-heart dose, predicts adverse cardiac events.
• Accurate segmentation of cardiac substructures is crucial for precise radiation dosimetry.

Purpose of the Study:

• To develop a robust, automated framework for cardiac substructure segmentation in thoracic radiotherapy CT imaging.
• To improve the accuracy and efficiency of cardiac substructure segmentation for radiation dose assessment.
• To enhance cardiotoxicity risk stratification in lung cancer patients undergoing radiotherapy.

Main Methods:

• Curated a dataset (CardiacSubstructSeg) of 69 patients with manual annotations of 21 cardiac substructures.
• Developed DINO-CardiacSeg, a novel framework using CT domain-adaptive self-supervised pretraining.
• Evaluated performance using five-fold cross-validation and compared with state-of-the-art methods.
• Assessed transferability by fine-tuning and evaluating on a larger public dataset (TotalSegmentator).

Main Results:

• DINO-CardiacSeg achieved superior performance on the internal dataset (65.95% DSC, 7.91 mm HD95).
• Consistent performance gains were observed across small and low-contrast cardiac structures.
• Demonstrated effective transfer learning on a public dataset, maintaining high performance (90.81% DSC, 6.07 mm HD95).

Conclusions:

• DINO-CardiacSeg provides accurate and scalable segmentation of cardiac substructures on thoracic radiotherapy CT scans.
• The framework leverages CT-specific foundation model pretraining for improved performance.
• This approach supports substructure-level cardiac dosimetry and better cardiotoxicity risk stratification in lung radiotherapy.