Robust framework for adaptive field-of-view automatic segmentation in vaginal brachytherapy for endometrial cancer

Area of Science:

• Medical Imaging
• Radiation Oncology
• Artificial Intelligence

Background:

• Vaginal brachytherapy (VBT) for endometrial cancer lacks efficient and reproducible automatic segmentation tools.
• Current manual segmentation methods are time-consuming and prone to inter-observer variability.
• Automated segmentation is crucial for optimizing VBT treatment planning and delivery.

Purpose of the Study:

• To develop and validate an automated framework for segmenting the clinical target volume (CTV) and organs of interest (OOIs) in VBT.
• To adapt the framework for varying clinical contouring protocols and partially labeled datasets.
• To assess the performance and clinical transferability of the automated segmentation method.

Main Methods:

• A three-step framework based on nnUNet was developed and adapted for segmenting CTV/OOIs on pre-treatment CT scans.
• The method accommodated different contouring protocols by utilizing region-of-interest labels and/or subsets of structures.
• Model development utilized a dataset of 289 endometrial cancer patients treated between 2014 and 2021.

Main Results:

• The automated segmentation achieved a Dice similarity coefficient of 87.7% for CTV and ranged from 72.6% to 88.9% for OOIs (small bowel, bladder, rectum).
• Dose calculation metrics showed acceptable deviations, with average absolute D2cm3 deviations of 27.5, 25.5, and 21 cGy for small bowel, bladder, and rectum, respectively.
• Clinical transferability assessment yielded median scores of 4 for CTV and 3 for bladder/rectum, indicating good clinical utility.

Conclusions:

• The developed framework provides high-quality segmentation of CTV and OOIs from CT scans, enabling reliable dose calculations.
• This automated approach outperforms existing state-of-the-art methods.
• The framework has the potential to significantly reduce complexity and treatment time in VBT protocols.