Automated Triage for Deep Inspiration Breath-Hold Using AI-Based Segmentation and Heart Dose Prediction
Abstract
Purpose
Deep inspiration breath-hold (DIBH) can reduce cardiac dose in left-sided breast cancer radiotherapy, but its benefit varies substantially between patients. DIBH is also more resource-intensive than free breathing (FB), increasing treatment time and workflow complexity, making unnecessary use undesirable. Accurate pre-treatment identification of patients who truly benefit from DIBH is therefore essential. This study compares the accuracy of mean heart dose prediction using two AI-based approaches—RapidPlan (Varian Medical Systems, Palo Alto, CA) and a deep learning (DL) model—for DIBH triage.
Methods
A DL-based dose prediction model using an HD-UNet architecture was trained, validated, and tested on 113/20/25 left-sided breast cancer patients treated with regional nodal irradiation. A RapidPlan model was trained on 27 patients and evaluated on an independent test set of 20 patients, reflecting typical clinical implementation. For both approaches, anatomical structures were generated using AI-Rad Companion Organs RT (Siemens Healthineers, Erlangen, Germany), enabling a fully AI-driven, post-CT workflow suitable for hands-off DIBH triage. A pilot clinical validation was performed on four patients with both FB and DIBH clinical treatment plans to assess predicted DIBH benefit.
Results
On the test cohorts, the mean error in predicted mean heart dose was 38.98 cGy for the DL model and 54.51 cGy for RapidPlan. In the four-patient subset, the DL model more accurately predicted DIBH-associated dose reduction, with a lower mean absolute error (45.1 cGy) and negligible bias (+0.9 cGy). RapidPlan showed larger errors (60.2 cGy) and a systematic negative bias (−58.9 cGy).
Conclusion
Both AI-based approaches achieved accuracy within clinically reasonable ranges for mean heart dose prediction. The DL model demonstrated superior performance for estimating DIBH benefit, supporting further evaluation for selective DIBH triage and avoidance of unnecessary, time-intensive DIBH treatments. RapidPlan, while slightly less accurate, offers the advantage of full integration into clinically available software.