2238 - Automated Delineation of Primary Gross Tumor Volume in Cervical Cancer Using a 3D Res-U-Net Model: Performance and Challenges

J. Chen¹, Z. Mo², L. C. Jia², J. Pan¹, S. Liu¹, C. Wang¹, J. Wang¹, C. Ji¹, J. Yang³, M. Cai¹, and J. Ma¹; ¹Department of Radiation Oncology, The First Affiliated Hospital of Xi’an Jiaotong University, Xi’an Jiaotong University, Xian, China, ²Shenzhen United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, China, ³The First Affiliated Hospital of Xian Jiaotong University, Xian Jiaotong University, Xian, China

Purpose/Objective(s): Cervical cancer is the fourth highest cause of cancer mortality and one of the most common gynecologic malignancies worldwide, accounting for a significant amount of the worldwide burden of womens cancer. Accurate delineation of the primary gross tumor volume (GTVp) in pre-treatment magnetic resonance (MR) images of cervical cancer plays a pivotal role in guiding clinical target volume contouring, as well as in contributing to definitive diagnoses and prognostic predictions. Materials/

Methods: In this retrospective, single-center study, pre-treatment T2-weighted MR sequences were collected from 198 cervical cancer patients treated between January 2013 and October 2018. All GTVp contours were manually delineated by two radiation oncologists. Due to the difficulty in identifying GTVp on axial plane MR images, manual contouring was cross-referenced with sagittal plane MR images. A 7-layer 3D Res-U-Net architecture was developed to automate the GTVp delineation process on axial plane MR images. The reliability of our model was evaluated through comparative experiments with nnUNetV2 as the benchmark. The configuration of nnUNetV2 was determined by an automated search strategy, whereas our proposed model diverges from nnUNetV2 primarily in aspects of the architectural design, input spacing, and input patch size. The performance of model was evaluated using dice similarity coefficient (DSC) and average symmetric surface distance (ASSD). The robustness and reproducibility of the model were assessed via five-fold cross-validation. Additionally, to evaluate the efficiency of the model, the mean time spent on manual delineation and modifications for model predictions were recorded for a subset of 50 patients.
Results: In the five-fold cross-validation with 198 patients, the benchmark model nnUNetV2 achieved a mean DSC of 0.76 ±0.17and a mean ASSD of 2.5 ±3.2. Additionally, the Res-U-Net model yielded a mean DSC of 0.78 ±0.16and a mean ASSD of 2.3 ±3.1. The average time for manual GTVp delineation was approximately 7 minutes per patient, while the average time for modifying model predictions was significantly less at about 3 minutes.
Conclusion: The 3D Res-U-Net model has demonstrated promising performance in the delineation of GTVp on axial plane MR images; however, its reliability with small tumors necessitates further enhancement. In addition to supplementary data, the fusion of axial and sagittal plane MR images is another approach worth exploring.