Click here to

Session: Multi-Disciplinary: Segmentation III [Return to Session]

Deep Learning Neural Network for Automatic Delineation of Para-Aortic Clinical Target Volume

W Lu*, P DCunha, W Chi, M Chen, L Ma, M Kazemimoghadam, Z Yang, X Gu, K Albuquerque, University of Texas Southwestern Medical Center, Dallas, TX


TH-IePD-TRACK 3-6 (Thursday, 7/29/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]

Purpose: To develop and evaluate an artificial-intelligence-based approach that automatically delineates clinical target volume (CTV) of para-aortic lymph nodes (PAN) for locally advanced cervical cancer (LACC) treatment.

Methods: We retrospectively collected 47 planning CTs of LACC patients. For each CT, the Aorta, Inferior Vena Cava (IVC) and PAN-CTV were manually delineated for the elective para-aortic region from the left renal vein to the aortic bifurcation. The PAN-CTV was prophylactically contoured (ignoring gross para-aortic nodes) according to the recent Small et al. (2020) consensus guidelines. For each CT set, the spine structures were automatically delineated using threshold- and morphology-based segmentation. Then, all CTs were registered to a common reference frame based on the delineated spine structures and cropped to dimensions of 15x15x20 cm³ centered along the spine. A U-Net-based segmentation network was developed and trained to automatically segment Spine, Aorta, IVC, and PAN-CTV in the registered and cropped region. The segmented structures were converted into DICOM RTStruct with the original CT association.

Results: We used 32 CTs for training, 8 for validation, and 7 for testing. The Dice-coefficient (DSC), Hausdorff-distance (HD95), and average symmetric-surface-distance (ASD) were used as the evaluation metrics. For PAN-CTV segmentations, our model achieved mean (SD) values of 75.5 (7.3)%, 8.42 (4.52) mm, and 1.55 (0.61) mm for DSC, HD95, and ASD, respectively. All PAN-CTV segmentations were visually checked by an attending radiation oncologist specializing in gynecologic malignancies and deemed clinically acceptable.

Conclusion: An Artificial Intelligence approach comprised of registration and DL network was developed and validated for automatic segmentation of Spine, Aorta, IVC, and PAN-CTV in LACC. Registration was used to improve the boundary definition of the virtual structure, PAN-CTV, for DL segmentation without human intervention. This approach could facilitate efficient workflow and allow for optimal dose delivery to the involved nodal basins in LACC.

Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by NIH grants: R01 CA235723, R01 CA218402.





    IM/TH- Image Segmentation Techniques: Modality: CT

    Contact Email