3185 - Using Deep Learning to Delineate the Elective Pelvic Nodal Volumes in Patients with Prostate Cancer: A Tool Aimed to Facilitate Peer Review
Presenter(s)
B. M. Anderson1, A. Karunaker2, M. Foster2, S. K. Das3, S. Saraiya2, S. Sud1, G. H. Goldin4, M. C. Repka1, L. B. Marks1, and L. Mazur1; 1Department of Radiation Oncology, University of North Carolina, Chapel Hill, NC, 2University of North Carolina, Chapel Hill, Chapel Hill, NC, 3University of North Carolina, Chapel Hill, NC, 4University of North Carolina Hospitals, Chapel Hill, NC, United States
Purpose/Objective(s): Accurate and consistent delineation of the elective pelvic nodal clinical target volume (CTV) is crucial for patients receiving radiation therapy (RT) to the prostate and elective pelvic nodes. However, inter-physician variability remains a challenge, leading to inconsistencies in treatment planning. We herein qualitatively and quantitatively evaluate an in-house deep learning-based segmentation model trained to delineate the elective pelvic nodal volumes in patients with prostate cancer.
Materials/Methods: Planning CT images from 152 patients treated with RT to the prostate and elective pelvic nodes between 2021 and 2024 were retrospectively analyzed. Patients were randomly split 8:2 (training:test) for model development and validation. Several fully convolutional neural network (FCNN) architectures were investigated. We created a graphical user interface that integrates into the treatment planning system and presents the model predictions and the manually defined CTV.
Within the test cases: Quantitative model performance (compared to the clinically used segmentations) was assessed using Dice Similarity Coefficient (DSC), and maximum surface distance. Qualitative analysis of 14 randomly selected test cases was performed by an experienced expert physician who rated each model-based segmentations as (a) clinically usable with ‘<2 minutes of edits’, (b) with ‘<10 minutes of edits’; or (c) ‘not usable.Results: The deep-learning-based model achieved a median DSC of 0.80 and median max surface distance of 3.3mm. Qualitative assessment reported 57%(8/14) of segmentations were 'Usable with 2 minutes of edits', 36% (5/14) as ‘<10 minutes of edits’, and 7% (1/14) were ‘Not Usable’.
Conclusion: Deep learning-based segmentation of the elective pelvic nodal CTV demonstrates high concordance with expert-defined contours and has the potential to facilitated standardized contouring practices. These findings support the use of our tool for facilitating peer review discussions and promoting contouring consistency in clinical practice. A formal experiment to assess the impact of these deep learning-based segmentations on peer review is planned.