ESTRO 2024 - Abstract Book

S4493

Physics - Machine learning models and clinical applications

ESTRO 2024

In this work, environments are defined by means of 91 series of target mask images (81 for training and 10 for validation) following noisy breathing signals of 450 timesteps with an average amplitude of 4 pixels. The trajectories are considered as the exact positions of the Gross Tumor Volume (GTV) center of mass, corresponding to a circle with a 2-pixels radius. Each synthetic 2D image (32*32 pixels) includes GTV , Planning Target Volume ( PTV , GTV enlarged by one pixel), PTV surroundings ( PTVs , zone covering two pixels around the PTV) and Organs At Risk ( OAR , the remaining part of the image). During treatment, the exact position is known every 200ms from binary masks reproducing fluoroscopy image acquisitions (see Figure 1). To model the position uncertainty pattern between two measures, a prediction of the target position is created by adding a Gaussian noise with increasing standard deviation (N(0, t*0.15)) to the exact trajectories. Two sets of positions are saved, the exact and the predicted positions. The daily dose distribution of one proton therapy treatment fraction (30 fractions of 2Gy) is determined by a sequence of 450 actions. An action can either be a beam’s movement (a pixel's up, down, right, left, no movement) or the delivery of a gaussian dose representing a proton pencil beam dose (5-pixels zone and std of 0.57) perpendicular to the image. The doses are accumulated on an image, called Accumulated DoseMap , centered on a reference position (see Figure2(B)) and used to assess treatment accuracy. These dose mappings on a reference position are made both on exact and predicted position. The Desired DoseMap is the objective dose defined with 2Gy in the PTV enlarged by one pixel and 1Gy one pixel around to obtain a homogeneous and sufficient dose in the PTV. Rewards are inversely proportional to the squared difference, on PTV + PTVs, between the Desired and the Accumulated DoseMap based on the exact tumor position. The observations representing the state of the environment are composed of three different matrices shown in Figure 1: Uncertainty Matrix (UM) representing the growing uncertainty on the exact position of the tumor's center of mass between two measurements thanks to Gaussian smoothing of the PTV mask at the predicted tumor position. • DoseMaps Difference (DD), the difference between the Desired and the Accumulated DoseMap based on the tumor predicted position. • Beam Position (BP), a binary matrix with a 1 on the pixel representing the Beam Position . •

These matrices are sent to a deep neural network which learns to select the best action considering the current state of the environment. BP and DD are combined before entering the network to extract dose information while BP and UM are combined to obtain positional accuracy.