ESTRO 2024 - Abstract Book

S4438

Physics - Machine learning models and clinical applications

ESTRO 2024

2. Second, the 3D dose distributions for the initial (S 0 ) and the different deformed (S d ) scenarios are predicted with a deep convolutional network (DCN)[1][2]. 3. Third, an “impact of correction” (IOC) scalar value is computed for each S d as the weighted sum of relevant dose-volume metrics differences between S 0 and S d . 4. Fourth, all IOC values are assigned to all voxels inside their corresponding deformed organs to form a 3D map for each S d . Final DIVE-map results from averaging out these 3D IOC maps for all S d within the same deformation amplitude.

Our DCN was trained with 180 adaptive treatment fractions from 9 prostate cancer patients previously treated with ETHOS adaptive treatment (Varian)[3].

DIVE-maps were generated for auto-segmented bladder and rectum (using ETHOS auto-segmenter) from 30 treatment fractions of 3 other patients. To evaluate the capability of our DIVE-map to indicate dosimetrically relevant regions, different triggering correction values ( t ) were evaluated, i.e. t =1, t =1.5, and t =2.0. We followed a slice-by-slice correction strategy based on the values of the DIVE-map and the manually corrected contours (considered our reference ground-truth): If all misclassified voxels (with respect to manually corrected contour) of a given slice had corresponding DIVE-map voxels values below the threshold t , no correction of auto-segmented contour was performed on the slice. In contrast, when there was at least one misclassified voxel with value above the threshold t , the entire slice was corrected exactly the same way as in the manually corrected contour. The plans were then automatically re-generated with the different corrected contours (using ETHOS Intelligent Optimization Engine). For each correction threshold t we specifically analyzed the V60Gy for 3 key OARs: the bladder, the rectum and the anal canal. The difference in PTV coverage was also reported. The decrease in the quantity of corrections for each threshold t was also evaluated. An example of generated DIVE-map for bladder and rectum can be found in Figure 1. Results of generated plans with DIVE-corrected bladder and rectum in terms of V60Gy and PTV coverage differences are display on Figure 2. First, It can be noticed that PTV coverage was not strongly impacted for any correction method since the largest delta values are about 0.5%. Concerning OARs, for uncorrected contours, 13 over 30 fractions presented a V60Gy difference above +2%. This was reduced to 6 fractions for the DIVE-corrected contours with threshold t=2.0. However, the best trade off between good dosimetric quality and reduced correction volume was found with the DIVE-map threshold of 1.5, since no fraction showed any V60Gy above +2% difference while contours corrections could significantly be reduced. Using our DIVE-map with this reasonable correction threshold of 1.5, the clinical quality could be preserved while decreasing correction volumes of rectum and bladder for all adaptive sessions that were part of this study. Our DIVE map tool identified 6 and 11 volumes out of the 30 where there was no correction to carry out for the rectum and the bladder, respectively. Moreover, in 16 and 18 fractions out of the 30, the number of corrected voxel was reduced by more than 50%. Results:

• Figure 1: Examples of generated DIVE-map. The higher the DIVE-map voxels values, the higher the dosimetric impact of a potential correction.