2885 - Beyond the Hardware: AI-Driven Cardiac Substructure Dose Assessment for Ventricular Tachycardia Radioablation

Purpose/Objective(s): Stereotactic body radiation therapy (SBRT) for ventricular tachycardia (VT) is a noninvasive, high-dose treatment that can reduce VT burden, decrease implantable cardioverter-defibrillator (ICD) shocks, and improve arrhythmia control in patients with limited treatment options. As clinical use expands, evaluating dose to cardiac substructures (CS) may help optimize efficacy and safety, but ICD-related image artifacts complicate accurate CS segmentation. This work applies deep-learning segmentation to segment CS and enable substructure-level dose analysis in cardiac radioablation.

Materials/Methods: Nineteen patients with ICDs who received SBRT (25 Gy, 1 fraction) for recurrent VT following prior intervention were evaluated. CT simulation (120kVp, 100-200 mAs, ~1 mm in-plane resolution, 1-3 mm slice thickness) was conducted using gating (n=5) or 4D with compression (n=14), with (n=10) and without (n=9) intravenous contrast, and reconstructed with (n=9) and without (n=10) metal artifact reduction. A semi-automatic pipeline for ICD artifact detection and masking was developed using K-means clustering and region growing. nnU-Net based segmentation, pre-trained on thoracic cancer patients with no ICDs, was applied to predict 20 CS including whole heart (WH), chambers, great vessels (GVs), coronary arteries (CAs), valves, and conduction nodes (CNs) on CT volumes. Model performance was compared before and after artifact masking coupled with post-processing (smoothing and interpolation) based on agreement with ground truth using Dice Similarity Coefficient (DSC), 95% Hausdorff Distance (HD95), and Wilcoxon Signed-Rank tests (p<0.05). Dosimetric evaluation (Dmean, maximum (D0.03cc)) was performed for CS.

Results: Median planning target volumes were 124.8cc (range, 50.5-306.1cc). Artifact masking and post-processing significantly improved segmentation performance, resulting in an average DSC of 0.65±0.25 (baseline, 0.55±0.26, p<0.05) and average HD95 of 12.9±14.1mm (baseline, 22.3±33.5mm, p<0.05) across all 20 CS. Table 1 summarizes doses where Dmean highest for the circumflex CA. All CS not reported had Dmean <5Gy and D0.03cc <8Gy (except for Aorta; D0.03cc, 10.9Gy).

Conclusion: DL segmentation of CS for VT treatment planning was feasible despite significant metal artifact burden. Semi-automatic artifact masking significantly improved CS segmentation accuracy to enable dosimetric evaluation for future patient outcome analysis.