D Boukerroui¹, J Baker^2,3*, Y Mcquinlan¹, A Riegel^2,3, Y Cao^2,3, M Gooding¹, L Potters^2,3, (1) Mirada Medical Ltd., Oxford, U.K, (2) Department of Radiation Medicine, Northwell Health, Lake Success, NY 11042, (3) Donald and Barbara Zucker School of Medicine at Hofstra/Northwell, Hempstead, NY 11549

• 
  J Baker

PO-GePV-M-63 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: Most studies on commissioning of auto-contouring have focused on global measures for assessing contouring performance or on time-savings following implementation, neglecting the clinical impact on contouring style and practice. This study aims to quantify the extent of expected manual adjustments when fully manual contouring of organs-at-risk (OAR) in the Head and Neck is to be replaced by an AI solution for a large cohort.

Methods: 716 Head and Neck cancer cases manually contoured (MC) in clinical routine practice were used. Fully automatic contours (AC) using deep-learning contouring (DLCExpert™, Mirada Medical Ltd., Oxford, UK) were generated and compared to the clinical contours. The amount of expected adjustment to the AC to perfectly match the corresponding MC was then computed. Contours were deformably registered to a common reference shape per organ for localized assessment analysis. Results were expressed as the median, 10ᵗʰ and 90ᵗʰ percentile of adjustment over the patient cohort and visualized using 3D renderings on the reference structures.

Results: The overall expected editing was to enlarge the ACs (medians were positive except for the Larynx). The medians were within axial voxel spacing (1mm) for 6 OARs and within the median slice spacing (3mm) for all OARs except for Esophagus, Optic Chiasm and Crico. Large adjustments were expected in some locations for most OARs. The spatial 3D rendering showed that many expected edits can be explained by interpretation of contouring guidelines (Optic chiasm, inferior part of spinal cord). However, others reflect one of variability in auto-contouring performance or in manual contouring.

Conclusion: Specific regions for expected large edits were identified per OAR in the Head and Neck. A similar analysis comparing ACs with clinical contours obtained by editing the ACs will permit assessing any impact of the introduction of AI based contouring solution on contouring variability in the clinic.

Funding Support, Disclosures, and Conflict of Interest: DB and MG are employees of Mirada Medical Ltd. DB was funded by Eureka Eurostars project ARTWORK Grant E!113263, InnovateUK, Grant 600485

Keywords

Quality Assurance, Segmentation, Statistical Analysis

Taxonomy

IM/TH- Image Analysis (Single Modality or Multi-Modality): Computer/machine vision

Contact Email