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

Abstract

Background and purpose: The lack of robust automatic tools for segmentation in vaginal brachytherapy (VBT) limits efficiency and reproducibility in clinical practice. We aimed to develop a framework for automatic segmentation of the clinical target volume (CTV) and organs of interest (OOIs) for endometrial cancer patients that undergo vaginal cuff brachytherapy.

Material and methods: We developed a three-step framework based on nnUNet and adapted to our context to segment the CTV/OOIs on pre-treatment computed tomography (CT) scans. Our method was adaptive to different contouring protocols used in clinical practice, where images were partly labelled, by either providing labels within a region of interest and/or considering only a subset of present structures. A dataset of 289 patients treated between 2014 and 2021 was used for model development (139/35/115 for training, validation and testing).

Results: The Dice similarity coefficient of the CTV was 87.7%, while OOI Dice coefficients were 72.6%, 88.9%, 86.3% for the small bowel, bladder and rectum, respectively. The average absolute D2cm3 deviations were 27.5(±58.7), 25.5(±21.0), and 21(±19.1) cGy for the small bowel, bladder and rectum, respectively. The average absolute D90 % deviation was 17.0(±19.6) cGy. Clinical transferability was assessed using a 1-4 scale, achieving a median of 4 for the CTV, 3 for bladder and rectum, and 2 for the small bowel.

Conclusions: Our method segmented the CTV and OOIs from CT scans with high quality and reliable dose calculations within the relevant range. Our framework outperformed state-of-the-art methods and potentially could help reduce complexity and time in VBT protocols.

Keywords: Artificial intelligence in oncology; Automatic segmentation; Dosimetry evaluation; Endometrial cancer; Vaginal cuff brachytherapy.