Rapid and Explainable Dibh Patient Selection Using Fully Automated, Planning-Free Deep Learning Framework In Left-Sided Whole-Breast Radiotherapy
Abstract
Purpose
Simple clinical criteria for deep inspiration breath-hold (DIBH) eligibility in left-sided whole-breast radiotherapy, such as the ability to sustain a 20-30 second breath-hold, are commonly used but do not reliably predict dosimetric benefit. Robust assessment requires acquisition of both free-breathing (FB) and DIBH CT scans followed by manual target delineation and treatment planning, making the process labor-intensive and time-consuming. This study presents a fully automated, planning-free deep-learning (DL) framework to estimate dosimetric outcomes and support rapid, explainable DIBH patient selection.
Methods
A DL framework was developed to optimize left-sided whole-breast radiotherapy planning workflow by predicting both 3D dose distributions and the corresponding tangential beam orientations. This approach utilized 3D CT images along with anatomical contours (left-breast, heart, left-lung) generated via a commercial auto-segmentation tool. The model was trained on 36 treatment plans (19 DIBH, 17 FB) with data augmentation to enhance robustness and assessed on an independent test set of 8 treatment plans (4 DIBH, 4 FB). The predicted dose distributions and beam orientations were evaluated against clinical reference plans using dose-volume histogram metrics and mean absolute error (MAE).
Results
The DL-predicted plans closely matched clinical reference plans in quality. The MAE for V95% of PTV was 0.58%±0.54% (DIBH: 0.49%±0.48%; FB: 0.67%±0.58%). For Dmean of the heart, the MAE was 0.37±0.47Gy (DIBH: 0.23±0.34Gy; FB: 0.51±0.54Gy), and for V20Gy of the left-lung, it was 0.41%±0.34% (DIBH: 0.36%±0.31%; FB: 0.46%±0.35%). The MAE between the predicted and planned beam orientations was 4.0°±1.5° (DIBH: 4.4°±1.7°; FB: 3.6°±1.3°).
Conclusion
The proposed DL framework enables rapid, patient-specific comparison of DIBH and FB dosimetry without dosimetrist involvement. It has the potential to support DIBH patient selection, reduce clinical workload, and streamline the planning process post decision-making, facilitating more efficient and personalized breast cancer care.