ESTRO 2024 - Abstract Book

S4456

Physics - Machine learning models and clinical applications

ESTRO 2024

coplanar beam arrangements in brain targets. We compare beam angles chosen by the AI to those generated by human planners. Additionally, plans were created based on human and AI chosen beam angles and resulting plan quality was compared dosimetrically.

Material/Methods:

The AI beam angle prediction was formulated as a key-point detection task, where each key-point represents a beam arrangement, i.e. a pair of gantry and couch angle. The full 360° range of possible gantry angles was divided into 72 steps, with each step spanning 5°. The couch angles were classified into seven categories: -90°,-60°,-30°,0°,30°,60°, and 90°. Consequently, the combinations of gantry and couch angles result in a total of 72x7 potential key-points. 109 brain patient plans were used to train the model. For prediction, the AI processes the CT image, the clinical target volume (CTV), and the contours of several organs-at risk (OARs) as input. It then provides key-point predictions as its output. The AI model's architecture consists of a 3D ResNet encoder with spatial and channel-wise squeeze-and-excitation blocks. This encoder is paired with a decoder that generates prediction maps, facilitating key-point selection. Model performance was evaluated with a key-point distance metric: Within the 72x7 grid, each key-point was assigned an x-coordinate for the gantry angle and a y-coordinate for the couch angle, as depicted in Figures 1&2 (top and right axes). The distance between beam arrangements chosen by humans and those chosen by the AI was calculated as the geometric difference between those key-point coordinates. Two measures were used to score model performance: The F1 score which corresponds to the relative amount of predicted beams within a distance ≤5 key points, and the relative mean distance across all beam arrangements. Metrics were evaluated on an independent test set of 20 patients. To evaluate the clinical viability of the AI prediction, deliverable plans were generated from the human and AI chosen beam angles for the test set. An automated knowledge-based planning solution was employed to avoid variability in human-driven plan optimization. Any dosimetric differences were thus caused by the chosen beam angles alone. Plans resulting from human and AI beam angles were compared in terms of dose-volume histogram (DVH) parameters for Body, CTV, Brainstem, Cochlea, and Optics and tested for significance with a paired t-test (p<0.05). Plans were normalized to provide the same CTV coverage (V100%=95%).

Results:

Average prediction time for each case was 0.6 seconds. Figures 1&2 show the count of human and AI chosen beam arrangements across the test set and visually illustrate good agreement. Quantitatively, the model obtained an F1 score of 0.7083 and a relative mean distance of 3.8 key-points.