ESTRO 2025 - Abstract Book

S2530

Physics - Autosegmentation

ESTRO 2025

was assessed according to a four-point Likert scale measure (FPLS) on a test set consisting of 18 CBCT images from 6 different patients. The FPLS qualitatively measures how acceptable the predicted contour for each organ is, ranging from 1 (Unacceptable) to 4 (Acceptable). The predicted contours were then analyzed to monitor cachexia dynamics by longitudinally evaluating tissue volume and Hounsfield Units (HU) as a measure of mass density within each tissue. Finally, the resulting analysis was presented intuitively in an interactive dashboard to facilitate clinical decision-making. Results: The segmentation model provides accurate contours with an average FPLS of 3.1 across all contours, indicating that the predicted contours have only minor errors. Figure 2 presents the analysis of NSCLC cachexia for a patient throughout chemo-radiotherapy. The workflow proved to be effective for monitoring NSCLC cachexia in patients, and interviews with clinicians highlighted the usability of the dashboard. The AI-driven approach enabled detailed automated tracking of muscle and adipose tissue changes over the course of treatment.