What question did this study set out to answer?

This research aims to establish a framework for evaluating the clinical acceptability and reproducibility of auto-segmentation models in head and neck cancer.

June 8, 2026Physics and Imaging in Radiation Oncology0 citationsOpen Access

Auto-segmentation of organs-of-interest clinical acceptability & reproducibility framework in head and neck cancer

JMJoseph MarsillaUniversity Health Network MWMattea WelchPrincess Margaret Cancer Centre JSJoshua SirajUniversity Health Network

Key Points

This research aims to establish a framework for evaluating the clinical acceptability and reproducibility of auto-segmentation models in head and neck cancer.
Introduced an Auto-Segmentation Clinical Acceptability & Reproducibility Framework to benchmark deep learning models.
Standardized model training and assessment workflows with 479/44/59 split for training/tuning/testing.
Used a 5-point Likert scale for expert assessment of AI-generated organs-of-interest.
16 out of 19 AI-generated organs-of-interest categories were deemed clinically acceptable by experts with minor revisions.
Framework codebase made publicly available for other researchers to benchmark methods.
Expert assessments support the transparency and reproducibility of auto-segmentation tools.

Abstract

Background and purpose Auto-segmentation of organs-of-interest (OOI) in cancer patients is essential for facilitating radiotherapy planning and reducing inter-observer variability. Deep learning-based auto-segmentation models have shown promise, but limited transparency and reproducibility hinder their generalizability and clinical acceptability, limiting their use in clinical settings. Materials and methods We introduced an auto-Segmentation Clinical Acceptability & Reproducibility Framework, a comprehensive framework designed to benchmark open-source deep learning models for auto-segmentation of 19 essential OOIs in head and neck cancer (HNC). Reproducibility was achieved through harmonized data curation, standardized model training (training/tune/test of 479/44/59) and assessment workflows. New models can be benchmarked against 12 pre-trained open-source deep learning models, while estimating clinical acceptability using a 5-point Likert scale. Results The framework codebase is openly available for benchmarking OOI auto-segmentation methods. During development, expert assessment of the best performing model labelled 16/19 AI-generated OOI categories as clinically acceptable with only minor revisions. Conclusions The framework facilitates benchmarking and expert assessment of AI-driven auto-segmentation tools, addressing the need for transparency and reproducibility in this domain. Through its emphasis on clinical acceptability, our framework fosters the integration of AI models into clinical environments, specifically within radiation therapy.

问 AI

Bookmark

View Full Paper