A prospective development and evaluation of a 2D convolutional neural network-based auto-segmentation model for cervical cancer radiotherapy.
Authors
Affiliations (8)
Affiliations (8)
- Department of Radiation Oncology, Kasturba Medical College, Manipal Academy of Higher Education, Manipal, India.
- Department of Radiology, Ochsner Clinic Foundation, New Orleans, LA, USA.
- Department of Radiation Oncology, BC Cancer Agency-Abbotsford, University of British Columbia, Vancouver, Canada.
- AIRONC Healthcare Technologies Private Ltd, Manipal, India.
- AIRONC Healthcare Technologies Private Ltd, Manipal, India. [email protected].
- Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India. [email protected].
- Department of Radiation Oncology, Kasturba Medical College, Manipal Academy of Higher Education, Manipal, India. [email protected].
- AIRONC Healthcare Technologies Private Ltd, Manipal, India. [email protected].
Abstract
Accurate delineation of target volumes and organs at risk (OAR) is essential in radiotherapy planning for cervical cancer. Deep learning (DL)-based auto-segmentation has the potential to improve contouring efficiency and workflow. This study reports the prospective development and internal validation of a DL-based auto-segmentation model- Deep contour (DC) for cervical cancer radiotherapy. In this prospective single-institution study, a 2-dimensional convolutional neural network based on the LinkNet architecture, DC, was trained on 190 computed tomography (CT) datasets for abdominal and pelvic OARs and 90 cervical cancer datasets for target volumes. Independent validation was performed on 20 CT datasets. Model performance was evaluated using dice similarity coefficient (DSC), Jaccard Index (JI), 95th percentile Hausdorff distance (HD95), average symmetric surface distance (ASSD), and surface dice coefficient (NSD). Expert internal and external radiation oncologists qualitatively explored clinical acceptability using a Likert scale, and segmentation time was compared with manual contouring. The DC demonstrated the greatest geometric performance for the femur (DSC 0.92 ± 0.03; NSD 0.94 ± 0.04), bowel bag (DSC 0.89 ± 0.03; NSD 0.77 ± 0.12), and bladder (DSC 0.88 ± 0.14; NSD 0.87 ± 0.12). Among the target volumes, the inguinal nodal clinical target volume (CTVn_inguinal) achieved the greatest agreement (DSC 0.77 ± 0.04; NSD 0.76 ± 0.07). Moderate performance was observed for the rectum (DSC 0.75 ± 0.16), liver (DSC 0.73 ± 0.21), and pelvic nodal clinical target volume (CTVn_pelvis) (DSC 0.60 ± 0.10), whereas lower performance was observed for anatomically complex structures such as the duodenum, anal canal, common bile duct, pancreas, and pelvic vessels. Clinical evaluation of two cases revealed a Likert score of III-IV for key pelvic organs, such as the bladder, femur, pelvic bowel bag, rectum, and sigmoid. Auto-segmentation significantly reduced the segmentation time from 77 min to 5 s per dataset (p < 0.001). This prospective validation demonstrates that DC auto-segmentation model can achieve acceptable geometric performance congruent across multiple abdominal and pelvic OARs and reasonable geometric performance for the elective inguinal CTV volume. Further validation on larger datasets and evaluation of clinical workflow integration are warranted. CTRI, TRN: CTRI/2024/02/063055, Registration date: February 22, 2024.