ESTRO 2022 - Abstract Book

S642

Abstract book

ESTRO 2022

Forty-nine patients that received liver treatments between 2017 and 2020 with proton pencil beam scanning were divided into a training (n=31), validation (n=7), and test set (n=11) for the AI (Fig. 1a). The AI is based on casting beam-angle selection as a multi-label classification problem. To account for angular boundary discontinuity, the underlying convolution neural network was trained with a novel Circular Earth Mover’s Distance based regularization and multi-label circular- smooth label technique. After training, the network generates a prediction for the probability of each of the 72 classes that represent 5° steps in beam angles (Fig. 1b). The final AI angles are then produced by post-processing with an analytical algorithm that emulates proton planning clinical practice. I.e., the probability is modified for each beam angle based on the distance to the target and organs-at-risk occluding the target (Fig 1c). Performance was evaluated by comparing AI beam angles with the ground truth, i.e., angles selected by human planners. Additionally, knowledge-based treatment planning was employed to automatically create treatment plans and assess the impact of beam angle choice on several dosimetric parameters.

Results For 8 of the 11 cases in the test set, AI-selected beam angles agreed with those determined by human planners to within 20 degrees (median angle difference = 10°; mean = 18.6°). Moreover, out of the total 22 beam angles predicted by the model, 15 (68%) were within 10 degrees of the human-selected angles (see table).

Case No.

Human Beam Angles [°] Predicted Beam Angles [°] AI Angles after post-Processing [°]

39 40 41 42 43 44 45 46 47 48 49

75, 130

0, 70

0, 70

0, 90

90, 170 40, 110

90, 170 40, 110

20, 105

45, 115, 180

40, 70, 110

40, 70, 170

90, 115 90, 180 45, 345 75, 190 80, 340 85, 175 0

40, 125 90, 170 90, 355 40, 110 90, 355 90, 160 0, 90

90, 125 90, 170 90, 355

0, 40

110, 160 90, 355 90, 160

Mean Difference Median Difference

26.2 12.5

18.6

10.0 The high correlation of human and AI predicted beam angles resulted in comparable plans in terms of dosimetric parameters (Fig 2). Select cases showed significant angle differences, dosimetric differences and their implications to overall plan quality will be discussed in terms of plan robustness, distal stoppage, and optimization settings.