ESTRO 2024 - Abstract Book

S3136

Physics - Autosegmentation

ESTRO 2024

VFSS of the HNC patients. For each frame, we created a multi-class binary mask that included four segmented regions of interest and six areas corresponding to a 2D Gaussian (sigma = 4 pixels) centered in each landmark. We trained a 2D U-Net within nnU-Net framework with five-fold cross-validation for 350 epochs each fold. For inference, we predicted each frame separately and then applied some additional post-processing steps to ensure the stability of the predictions throughout the video. To evaluate the accuracy of the model in the remaining 1552 labeled frames, we computed the average distance between the ground truth and the predicted center of masses for each landmark, and the average surface distance for the segmentations. The second aim was to extract several parameters from the labeled VFSS characterizing the swallowing process and compare the results between the HNC and healthy groups with a student t-test. For this, we isolated a single deglutition event from the healthy and pathological labeled VFSS and measured seven swallowing parameters: tongue posterior movement, opening of upper esophageal sphincter, elevation of the hyoid bone, peristaltic movement of the posterior pharyngeal wall, closure of the pharyngeal light, and amount of bolus remaining in the epiglottal and esophageal areas before and after swallowing. The dynamic behavior of these parameters was analyzed both with the distance or area change in pixels and with the time required for stage completion. The total time of deglutition was also measured.

Results:

Regarding the auto-labeling algorithm, each fold of the 2D U-Net model was trained in approximately 3 hours, yielding a total training time of 15 hours. The model was evaluated in the 1552 frames not used for training, yielding an average distance error of 2.045 pixels for the landmarks and 0.860 pixels as the mean average surface distance for the segmentations. Figure 1 compares the ground-truth and predicted regions of interest of two VFSS frames belonging to different HNC patients.