A multi-centre evaluation of deep learning based radiotherapy planning for left-sided node-negative breast cancer

Abstract

Background and Purpose:

Deep learning based planning (DLP) has the potential to improve consistency and efficiency in radiotherapy treatment planning. However, its clinical applicability remains limited, partly due to the need to translate a predicted dose into a deliverable dose. This study evaluated the generalisability of an institution specific DLP solution across multiple institutions by assessing its performance and developing a standardised translation parameter set.

Materials and Methods: Four institutions provided clinical treatment plans of 15 patients with left-sided node-negative breast cancer. Treatment plans delivering 40.05 Gy were generated using a deep learning prediction model trained on data from one institution. External validation was performed using national consensus criteria, by applying the initial parameter settings (InitialMimick) to datasets (n = 45) from three other institutions. A standardised parameter set (GenericMimick) was then developed based on data (n = 12) from all four institutions, whereafter it was evaluated on the remaining 48 patients of the dataset. Results: InitialMimick plans showed higher average dose values in the planning target volume for the Dmean (40.5 vs. 40.1 Gy) and D2% (42.4 vs. 41.4 Gy), with fewer cases meeting all clinical goals (15/45) compared to clinical plans (25/45). After parameter adjustment, GenericMimick plans resulted in more plans meeting all goals (28/ 48), comparable to the clinical plans (30/48), with Dmean of 40.3 vs. 40.1 Gy and D2% of 41.9 vs. 41.5 Gy. Mean differences in organs at risk mean doses were less than 0.2 Gy. Conclusion: DLP with a standardised translation parameter set demonstrated general applicability across multiple institutions.