Advancing Pharyngeal Constrictor Muscle Auto-Segmentation: A Comparative Analysis

Anju Kaimal, Konstantinos Zormpas-Petridis, Justine Tyler, Christopher M. Nutting, Matthew Blackledge

Keywords: Auto-contouring, Head and neck cancer, radiotherapy planning

Abstract: Accurate auto-segmentation of the pharyngeal constrictor muscle (PCM) is crucial for precise treatment planning in head and neck cancer. This study compares three deep-learning segmentation methods: (i) 2D U-Net, (ii) 3D UNet with data fingerprinting (nnUNet), and (iii) our proposed 2D statistical curve fitting method using transfer learning. Using a total of 168 planning CT images for training and validation, results indicate varying effectiveness, with the nnUNet exhibiting the lowest mean surface distance (2.73mm, inter-quartile range = 1.02mm), statistical curve fitting obtained a mean-surface distance of 8.25 mm with IQR of 6.33 mm. Though this was not as accurate as nnUNet, but was significantly superior than a conventional 2D UNet model.

Submission Number: 138