Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Clinical Target Volume Analysis From a Web-Based, AI Platform

T Netherton1*, R Mumme1, A Olanrewaju1, R Douglas1, L Zhang1, D Rhee1, C Cardenas2, H Simonds3, A Jhingran1, B Beadle4, L Court1, (1) The University of Texas MD Anderson Cancer Center, Houston, TX, (2) The University of Alabama at Birmingham, Birmingham, AL, (3) Stellenbosch University, Stellenbosch, ,ZA, (4) Stanford University, Stanford, CA


PO-GePV-M-10 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: The contouring of clinical target volumes (CTVs) is time intensive and depends on the gross tumor volume, physician experience, and many other factors. We are developing a tool to automatically contour CTVs for head and neck cancer and cervical cancer radiotherapy planning. Here we compare auto-contoured CTVs with clinical CTVs.

Methods: (1) Head and neck lymph node levels II-IV, II-V, Ia-V, 1b-V, and retropharyngeal nodes were automatically contoured using deep learning models, and the auto-CTVs retrospectively compared to nodal components from treated CTVs (n = 50). Patients included definitive and post-operative cases with gross disease throughout the head and neck (e.g. oral cavity, larynx, oropharynx). (2) Cervix primary and nodal CTVs and ITVs were automatically contoured using deep learning models, and compared to clinical CTVs (n = 67) and ITVs (n = 67). Patients included those who previously received cervical radiotherapy.

Results: The head and neck CTV comparison yielded average DSC, precision, and recall values of 0.65 ± 0.13, 0.80 ± 0.11, and 0.57 ± 0.16, respectively. Visual evaluation indicated that differences were mostly around the GTV volumes, where variation in clinical practice is expected. The cervical cancer CTV comparisons yielded average DSC, precision, and recall values of 0.78 ± 0.06, 0.84 ± 0.10, and 0.75 ± 0.09, respectively. Cervix ITV comparisons were 0.67 ± 0.11, 0.74 ± 0.13, and 0.63 ± 0.13, respectively.

Conclusion: Automatically generated target volumes agree with manually contoured volumes, with most differences around the gross tumor volume, as expected.

Funding Support, Disclosures, and Conflict of Interest: Multiple authors are members of the Radiation Planning Assistant Team at UT MD Anderson


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email