2219 - Enhancing X-Ray Based Adaptive Radiotherapy: Pilot Study on AI-Driven Contour Quality

T. Bai, C. Y. Liao, S. Wang, D. H. Moon, J. Visak, D. D. M. Parsons, D. J. Sher, and M. H. Lin; Department of Radiation Oncology, University of Texas Southwestern Medical Center, Dallas, TX

Purpose/Objective(s): Contouring efficiency is a critical aspect of online adaptive radiotherapy as a large contributor to the total time a patient remains immobilized on the couch. Our prior work showed the contour propagation-based system holds the promise to alleviate this challenge. A novel version of the system with a fully AI engine for organs-at-risk (OARs) auto-contouring exhibits further potential to address this challenge. In this pilot study, we aim to evaluate the performance of the new AI system by quantifying the contour efforts reduction and explore how to best hybrid them towards efficient workflow. Materials/

Methods: We retrospectively analyzed standard CBCT data from 776 treatment sessions involving 14 OARs based on 24 head-and-neck cancer patients. For each OAR, we compared the AI contours against the physicians’ contours. As a surrogate for manual online contour editing time, we calculated the added path length for each contour with a 1.5 mm margin (APL-1.5mm). This added margin length indicates stylistic, non-dosimetrically impactful changes. We then computed the ratio of APL-1.5mm to the total length of the physicians’ contours to isolate which contours needed large editing time (i.e., a ratio of 1 indicates no edits). We further divided each structure into three sections to parse critical editing locations as: missing, inferior/superior, middle slices, and utilized APL-1.5mm reduction to measure the efforts distribution. Lastly, we visually examined contour difference.
Results: Contouring efforts reduction ranges from 13% to 95% with a median value of 80%. For brachial plexus, spinal canal, constricts, and esophagus, the primary editing efforts are from missing slices. For mandible, oral cavity and parotids, the editing is concentrated on the middle slices. For the remaining structures, the effort distribution is relatively uniform. Visual examination shows contouring discrepancy for oral cavity, esophagus, and spinal cord stems from different contouring style.
Conclusion: Our initial experience indicates that the new system can decrease the effort required for contouring to varying degrees across different structures, with specific focuses during online adaptation. For OARs where the contouring styles deviate from institutional protocol, it may be advisable to strategically disable the AI and utilize the deformed contours instead. Other OARs’ performance is expected to be further improved with the new high-quality CBCT. Abstract 2219 – Table 1

	Contour Effort Reduction	Contour Edits Location
		Missing	Inf/Sup	Middle
Mandible	95%	11%	19%	70%
BrachialPlex_L	86%	48%	12%	39%
BrachialPlex_R	84%	56%	15%	29%
Brainstem	78%	40%	28%	32%
Spinal Canal/Cord	70%	82%	6%	11%
Cavity_Oral	13%	8%	23%	70%
Constrict_Sup	89%	65%	12%	22%
Constrict_Mid	87%	71%	18%	11%
Constrict_Inf	86%	79%	12%	9%
Esophagus	70%	49%	26%	30%
Glnd_Submand_L	77%	27%	26%	47%
Glnd_Submand_R	78%	31%	29%	40%
Parotid_L	80%	20%	25%	55%
Parotid_R	84%	18%	28%	54%