One bottleneck of MRI-guided Online Adaptive Radiotherapy (MRoART) is the time-consuming daily online replanning process. The current leaf sequencing method takes up to 10 minutes, with potential dosimetric degradation and small segment openings that increase delivery time. This work aims to replace this process with a fast deep learning-based method to provide deliverable MLC sequences almost instantaneously, potentially accelerating and enhancing online adaption.
Approach: Daily MRIs and plans from 242 daily fractions of 49 abdomen cancer patients on a 1.5T MR-Linac were used. The architecture included: 1) a recurrent conditional Generative Adversarial Network (rcGAN) model to predict segment shapes from a fluence map (FM), recurrently predicting each segment's shape; and 2) a linear matrix equation module to optimize the monitor units (MU) weights of segments. Multiple models with different segment numbers per beam (4-7) were trained. The final MLC sequences with the smallest relative absolute errors were selected. The predicted MLC sequence was imported into treatment planning system for dose calculation and compared with the original plans.
Main results: The gamma passing rate for all fractions was 99.7±0.2% (2%/2mm criteria) and 92.7±1.6% (1%/1mm criteria). The average number of segments per beam in the proposed method was 6.0±0.6 compared to 7.5 ± 0.3 in the original plan. The average total MUs were reduced from 1641 ± 262 to 1569.5 ± 236.7 in the predicted plans. The estimated delivery time was reduced from 9.7 minutes to 8.3 minutes, an average reduction of 14% and up to 25% for individual plans. Execution time for one plan was less than 10 seconds using a GTX1660TIGPU.
Significance: The developed models can quickly and accurately generate an optimized, deliverable leaf sequence from a FM with fewer segments. This can seamlessly integrate into the current online replanning workflow, greatly expediting the daily plan adaptation process.

.
Read full abstract