198 - Deep Learning-Based Quick MLC Sequencing for MRI-Guided Online Adaptive Radiotherapy

A. Ahunbay¹, E. S. Paulson², S. Hamdan³, L. Buchanan⁴, E. E. Ahunbay², and Y. Zhang²; ¹University of Wisconsin-Madison, Madision, WI, ²Department of Radiation Oncology, Medical College of Wisconsin, Milwaukee, WI, ³Yale University, New Haven, CT, ⁴Thompson Proton Center, Knoxville, TN

Purpose/Objective(s): One of the bottlenecks of MR-guided Online Adaptive Radiotherapy (MRgOART) is the time-consuming daily online replanning process. For full blown adaptation (Adapt to Shape), a two-phase optimization process is needed: Phase 1 involves a fluence based optimization with non-deliverable theoretical fluences, Phase 2 focuses on segment-based optimization with final deliverable MLC leaf sequences. There can be large dosimetric degradation from phase 1 to 2 due to leaf sequencing as the optimizer does not see the effects of leaf sequencing in the first phase, and recovery can take up to 10 minutes of iterations. This work aims to replace the time-consuming leaf sequencing process with a fast deep learning (DL)-based method that can provide deliverable segments almost instantaneously, potentially accelerating and enhancing the online optimization process. Materials/

Methods: Daily MRIs and plans from 225 daily fractions of 45 pancreatic cancer patients undergoing MRgOART on a 1.5T MR-linac were used. The architecture included the following two major steps: 1) a conditional Generative Adversarial Network (cGAN) model to predict the shapes of each segment from a fluence map, which recurrently predicts each segments shape using previously predicted segments along with the fluence map; and 2) a linear matrix equation module to optimize the MU weights of segments. Multiple models with different number of segments per beam (4-6) were generated and run and the final segments for each beam were selected with the smallest Mean Absolute Error with the target fluence. The deliverable leaf sequences were imported into the treatment planning software for dose calculation. Comparisons were performed with the original plans using 3D Gamma analysis with a 2%/2mm and 1%/1mm criterion.

Results: The average number of segments per beam in the DL cGAN method was 5.6 as compared to 7.8 in the original plan, resulting in reduced estimated delivery time of 8.7 mins vs 10.6 mins. The mean gamma passing rate was 99.9% using 2%/2mm criteria and 93.1% for 1%/1mm criteria, close to the 99.7% and 96.7% with repeated calculations using 1% statistical uncertainty. The sequencing for one plan took ~5s using a GTX1660TIGPU.

Conclusion: The developed models can generate an optimized deliverable leaf sequence quickly and accurately from a fluence map with a reduced number of segments. The beam-wise MLC sequence prediction has the potential to facilitate real-time beam-wise plan adaptation. It can seamlessly integrate into the online replanning workflow, greatly expediting the daily plan adaptation process.

Table 1

Case	Gamma Passing Rate/%				# segs/beam		MU		Estimated Delivery Time(min)
	Orig vs Repeated calc		Orig vs DL
	1%/1mm	2%/2mm	1%/1mm	2%/2mm	orig	DL	orig	DL	orig	DL
1	97.7	99.8	93.4	100.0	7.9	6	1910	1882	9.45	8.1
2	93.4	99.6	91.7	99.8	8.1	5	1766	1729	9.3	7
3	95.1	99.5	91.8	99.9	8	4.9	2071	2040	9.87	7.68
4	99.4	100.0	95.7	100.0	7.5	5.9	1415	1393	12.1	10.2
5	98.0	99.8	93.2	100.0	7.6	6	1431	1424	12.2	10.3
mean	96.7	99.7	93.1	99.9	7.8	5.6	1719	1694	10.6	8.7