DIVE-ART: A tool to guide clinicians towards dosimetrically informed volume editions of automatically segmented volumes in adaptive radiation therapy

Roberfroid, Benjamin [UCL] Lee, John Aldo [UCL] Geets, Xavier [UCL] Sterpin, Edmond [UCL] Barragan Montero, Ana Maria [UCL]

Purpose: Automatic segmentation algorithms are key for online adaptive radiotherapy. However, the lack of quality assurance tools requires manual review of auto-segmented volumes. Our goal is to reduce these manual operations thanks to a Dosimetrically Informed Volume Edition (DIVE) map. Methods: The DIVE map for a given auto-segmented volume is generated as follows. First, the volume is deformed so that N deformation scenarios (Sd,d=1,..N) are generated. Second, the 3D dose distribution for the initial (S0) and deformed (Sd) scenarios is predicted with a deep convolutional neural network (DCNN). Third, an “impact of correction” (IOC) scalar value is computed for each Sd as the weighted sum of relevant dose-volume metrics differences between S0 and Sd. These IOC values are then spatially aggregated to form the DIVE-map. Our DCNN was trained with 180 adaptive treatment fractions. DIVE performance for auto-segmented bladder and rectum was evaluated on 30 fractions. The evaluation assessed 1) the dose-volume metrics from automatically generated plans for either conventionally- or DIVE-corrected contours; 2) the decrease in the amount of corrections when using the DIVE-map. Metrics were evaluated on the conventionally corrected contours. Results: Using the DIVE-map, the number of corrected voxels was reduced by at least 50% in more than half of the 30 fractions (16 for rectum, 18 for bladder), while keeping the same dosimetric quality as the plan with conventionally corrected organs (differences in V60Gy below +2%). Conclusion: The proposed method successfully reduced the need for corrections in auto- segmented organs without impairing plan dosimetric quality.