3397 - Auto-Segmentation of Elective Nodal Clinical Target Volumes for Anal Cancer Using Artificial Intelligence

M. A. Kamal Aldin¹, B. Law², P. H. Stone³, S. H. Benedict³, P. Park³, J. Hunt¹, L. Tchelebi⁴, S. K. Jabbour⁵, Q. Chen⁶, Y. Rong⁶, and A. M. Monjazeb¹; ¹University of California Davis Comprehensive Cancer Center, Sacramento, CA, ²UC Davis medical school, Sacramento, CA, ³UC Davis, Department of Radiation Oncology, Sacramento, CA, ⁴Northwell, Lake Success, NY, ⁵Rutgers Cancer Institute of New Jersey, Department of Radiation Oncology, New Brunswick, NJ, ⁶Department of Radiation Oncology, Mayo Clinic AZ, Phoenix, AZ

Auto-segmentation of elective nodal clinical target volumes for anal cancer using artificial intelligence

Purpose/Objective(s): The application of artificial intelligence (AI) for automatically segmenting organs at risk and target volumes in radiation therapy planning is rapidly developing. We aim to develop an AI model that creates elective nodal clinical target volumes (nCTV) in anal cancer as a template for pelvic nodal auto-segmentation and as a standard workflow applicable to other disease sites. The primary aim of this investigation is to determine if refining the quality of the training data can generate accurate AI models using smaller data sets.

Materials/

Methods: We identified 70 anal cancer patients contoured between 12/20/2010 and 12/1/2022. CTV nodal volumes were edited for quality to align with the NRG Oncology contouring atlas. Of 70 cases, 40 cases were assigned to a training group and 30 cases were assigned to a testing group. Three training cases were excluded due to incomplete data. Thirty testing cases were contoured by AI and manually. The differences in AI generated vs. manually contoured CTVs were quantitatively evaluated using novel scripts for the median, mean and range of (1) Hausdorff distance 100^th and 95^th, (2) volumetric differences, and (3) dice similarity index measures. The quality of AI vs. manual contours were evaluated using a subjective 10-point scale (10=requires no editing, 5=major editing, 1=complete editing) by 3 board certified GI radiation oncologists.

Results: Of 70 cases, 21 were male and 49 female; 53 used IV contrast; 64 used oral contrast; 64 were simulated in the prone position, and 6 supine. In the testing set, AI generated contours that volumetrically recapitulated manual contours (Table 1). GI radiation oncologists preferred manual contours in 56% of cases, AI contours in 14%, and found them to be equivalent in 30%. The mean quality score was 7.6(+/-1.1) for manual vs. 6.9(+/-0.87) for AI generated.

Conclusion: We developed a new model for auto-segmentation of nCTV based on the NRG Oncology anorectal contouring atlas. This model used fewer training cases than most AI contouring models and yielded metrics similar to manual contours, showcasing that refining training data can improve AI models. Physician experts preferred manual contours in most cases but, AI generated contours were usually similar and acceptable, and sometimes even preferrable. This tool may significantly reduce physician effort to contour pelvic lymph node CTVs but physician editing is still required to account for inter-patient heterogeneity in volumes. Strategies for improving the metrics and quality of this tool will be further investigated. Abstract 3397 – Table 1

Table 1	Median	Mean	Range Lower Limit	Range Upper Limit
Hausdorff Distance(cm) 100th	4.93	4.69	2.44	8.88
Hausdorff Distance (cm) 95th	1.94	1.77	1.20	3.60
Volumetric Difference(cc)	155.30	139.66	16.29	621.73
DICE similarity index	0.79	0.80	0.68	0.84