Predicting Achievable Cardiac Dose Reduction with Deep Inspiration Breath-Hold Using Free-Breathing CT and Deep Learning
Abstract
Purpose
Deep-inspiration breath-hold radiotherapy (DIBH) is widely used in left-sided breast cancer radiotherapy to reduce heart dose, but the dosimetric benefit varies among patients. Assessing DIBH eligibility typically requires additional CT simulation and treatment planning, increasing clinical workload. To address this, we developed a two-stage deep learning framework to synthesize DIBH-CT images (sDIBH-CT) and predict dose distribution. This study evaluates the feasibility of predicting DIBH-induced mean heart dose reduction (ΔMHD) using only free-breathing CT images (FBCT), aiming to streamline patient selection.
Methods
Paired FBCT and DIBH-CT datasets from 132 left-sided breast cancer patients were retrospectively analyzed (training: 92, validation: 19, test: 21). A 3D generative adversarial network was implemented to synthesize sDIBH-CT from FBCT. Concurrently, a 3D U-Net dose prediction model was trained on clinical DIBH-CTs. The predicted ΔMHD was calculated as the difference between MHD predicted by the trained 3D U-Net on the original FBCT and that on the synthesized sDIBH-CT. Reference ΔMHD was defined as the difference in MHD between clinical plans based on DIBH-CT and FBCT. Performance was validated by comparing predicted ΔMHD against the ground-truth reference plans using median absolute error (MedAE) and Pearson’s correlation coefficient (r). Additionally, the anatomical accuracy of sDIBH-CT and organ contours was visually verified.
Results
The sDIBH-CT images successfully reproduced characteristic anatomical changes associated with DIBH, including thoracic expansion and posterior-inferior cardiac displacement. In the test set, comparison with clinical plan–derived reference ΔMHD showed a MedAE of 0.11 Gy, with an interquartile range of 0.04–0.34 Gy. A strong positive correlation was observed between the predicted and reference ΔMHD (r = 0.743, p < 0.01).
Conclusion
This framework demonstrates the feasibility of estimating achievable cardiac sparing with DIBH using FBCT alone. The results suggest potential for automated patient stratification and DIBH eligibility assessment prior to CT simulation, thereby improving clinical efficiency.