ESTRO 2024 - Abstract Book

S3141

Physics - Autosegmentation

ESTRO 2024

Keywords: Deep learning, adaptive region-specific loss

2913

Digital Poster

AI-segmentation of Pelvic Bone Substructures for Bone-Sparing Radiotherapy

Lars Nyvang 1 , Camilla JS Kronborg 2,3 , Karen-Lise G Spindler 1,3 , Tine B Nyeng 1 , Anne IS Holm 1 , Maja V Sand 1 , Line Ø Ring 1 , Martin Kyndt 4 , Ditte S Møller 1,3 1 Aarhus University Hospital, Department of Oncology, Aarhus, Denmark. 2 Aarhus University Hospital, Danish Centre for Particle Therapy, Aarhus, Denmark. 3 Aarhus University, Department of Clinical Medicine, Aarhus, Denmark. 4 MIM Software Inc., MIM, Cleveland OH, USA

Purpose/Objective:

The Danish Anal Cancer Group (DACG) II trial (NCT05385250) investigates the use of radiotherapy minimizing bone exposure thereby decreasing the risk of insufficiency fractures in the treatment of anal cancer. Bone sparing radiotherapy requires time consuming delineation of bone substructures (figure 1) and given the potential applicability in pelvic radiotherapy, there is a pressing need for methods that expedite the segmentation process and seamlessly integrate it into clinical practice. To address this challenge, we have developed and assessed an automated algorithm designed for the rapid segmentation of pelvic bone substructures (PBS).

Material/Methods:

We established a training dataset that included PBS (comprising the complete pelvic bones, sacroiliac (SI) joints for both the left (L) and right (R) sides, sacral alae (L/R), acetabulum (L/R), symphysis (L/R), and femoral heads (L/R)) for a cohort of 69 patients diagnosed with anal cancer. An AI algorithm based on U-Net (referred to as AI) was developed through a collaboration with MIM software. To evaluate the performance of the AI model, ground truth PBS datasets were manually delineated by an experienced oncologist on CT scans from 10 additional patients with anal cancer and compared to the AI segmentations.