Efficient deep learning-driven workflows for accelerated single-fraction radiation therapy planning of a solitary tumour in the lung

Abstract

Background and Purpose Same-day, single-fraction radiotherapy requires efficient planning. We introduce two deep learning (DL) dose prediction-driven workflows to generate clinically acceptable plans for treating metastases in the lung. Materials and methods Consecutive patients treated for a tumour in the lung (11/2018–11/2024) were considered. One workflow used the computed tomography density map and a synthetic dose built from the tumour segmentation as inputs for the DL model [without-organs of interest (OOI)]. The other workflow additionally used predicted OOI segmentations (with-OOI). Optimisation objectives were extracted from the predicted dose by the treatment planning system (TPS). Generated plans were compared with clinical plans with the number of achieved goals, conformity index CI100, and control point modulation complexity score (cpMCS). Plan complexity increases with decreasing cpMCS. Processing times were recorded. Results Training and testing datasets contained 200 and 56 treatments, respectively. Without- / with-OOI workflows generated plans in 8.8 and 9.6 min. The 3%/3 mm gamma passing rates between predicted and clinical doses were median ± interquartile range 99.4 ± 0.7% / 99.3 ± 0.8%. After dose optimisation in the TPS, 25% of plans had a higher number of achieved clinical goals relative to clinical plans in both workflows. The conformity was improved (CI100 = −5.3% / −3.9% relative to the clinical value, p-values <0.001) at the cost of increased complexity (cpMCS = −19.4% / −19.0%, p-values <0.001). Conclusions Workflows driven by DL dose prediction could generate a clinically acceptable plan in less than ten minutes for same-day, single-fraction treatment of a tumour in the lung.