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

Radiation-induced cardiotoxicity contributes to non-cancer mortality in patients treated with thoracic radiotherapy. Radiation dose to specific cardiac substructures is more predictive of adverse cardiac outcomes than whole-heart dose. This study aimed to develop a robust, automated cardiac substructure segmentation framework tailored to thoracic RT imaging.

We curated CardiacSubstructSeg, a contrast-enhanced RT simulation CT dataset comprising 69 patients with left-sided lung cancer and expert manual annotations of 21 cardiac substructures. We proposed DINO-CardiacSeg, a novel segmentation framework built upon a CT domain-adaptive self-supervised pretraining. Model performance was assessed using five-fold cross-validation and compared with state-of-the-art methods. Transferability was assessed via fine-tuning and evaluation on the TotalSegmentator cardiac subset (n = 603).

On CardiacSubstructSeg, DINO-CardiacSeg achieved the highest overall performance across all metrics (65.95% ± 7.91% DSC and 7.91 ± 4.40 mm HD95). Performance gains were consistent across small and low-contrast structures. In transfer-learning evaluation on TotalSegmentator, DINO-CardiacSeg maintained superior performance (DSC: 90.81% ± 8.23%, HD95: 6.07 ± 13.17 mm), demonstrating effective adaptation after fine-tuning to a dataset with differing annotation protocols and imaging characteristics.

DINO-CardiacSeg enables accurate and scalable segmentation of clinically relevant cardiac substructures on thoracic RT simulation CT through CT-specific foundation model pretraining. This approach outperforms existing methods on internal evaluation and fine-tuning to a public cardiac CT dataset, supporting its potential for substructure-level cardiac dosimetry and improved cardiotoxicity risk stratification in lung radiotherapy.