ESTRO 2024 - Abstract Book

S3061

Physics - Autosegmentation

ESTRO 2024

The models were tested using the kV images acquired as part of conventional RT from two fractions of each patient, giving 70 test images per patient (35 per fraction). The cGAN segmentation with template matching was compared to the ground truth segmentation for the analysis. The ground truth was labelled by medical physicists in each kV image by performing a rigid contour shift using the open-source Contour Alignment Tool. 7 The model performance was quantified using the centroid geometric error and Dice similarity coefficient (DSC).

Fig. 1 Simulated real-time clinical study of the deep learning model. A patient-specific network is trained prior to the patient’s treatment using DRR data. The generator network from the cGAN is used during the treatment to segment the target. The location of the segmented target can be used for motion management.

Results:

For all patients, DRR generation and model training took less than 4 hours each. The distribution of the centroid errors and DSC for all patients are represented in Fig. 2. The mean and standard deviation of the centroid error was -0.4 ± 2.8 mm and -0.6 ± 1.4 mm in the anterior-posterior/lateral (AP/LAT) and superior-inferior (SI) directions, respectively (Fig. 2a). The predicted and ground truth segmentations have high agreement, with a mean DSC of 0.87 ± 0.08 (Fig. 2b). The time taken for the model to generate the segmentation was less than 10 ms per image on average.