ESTRO 2021 Abstract Book

S563

ESTRO 2021

Conclusion Both our deep learning based models decrease brachial plexus delineation variability. Combining our first learning based model with anatomical criteria leads to model improvement with a high rate of clinical acceptability. Such an approach could be used to improve automatic delineation tools for other complex OARs. PD-0732 Can we reduce clinician intervention in breast target volume auto-segmentation approvals? C. Welgemoed 1,2 , E. Spezi 3 , D. Gujral 1,2 , R. McLauchlan 1,4 , E. Aboagye 2 1 Imperial College Healthcare NHS Trust, Radiotherapy, London, United Kingdom; 2 Imperial College London, Department of Surgery and Cancer, London, United Kingdom; 3 Cardiff University, School of Engineering, Cardiff, United Kingdom; 4 Imperial College London, Department of Physics, London, United Kingdom Purpose or Objective Volume delineation in breast and regional nodal radiotherapy is vital for the accurate execution of treatment. Current practice involves manual delineation, resulting in large inter-and intra-observer variations. Auto- segmentation methods could help standardise and improve observer variations but vary in accuracy and require clinician intervention before clinical application. These accuracy variations are best described by comparing an auto-segmentation to the “ground truth,” calculating the over-and underlap of the volumes. This study explored the correlation between a geometric accuracy measure and clinician evaluation. A potential geometric accuracy threshold value for clinically acceptable outlines would reduce the number of structures that require checking and save clinician time. Materials and Methods 100 CT scans of previously treated patients were randomly selected. Structures including the breast and level one to four lymph nodes were delineated and checked by a consultant breast oncologist, adhering to the ESTRO guidelines 1 . Following delineation, these CT-structure sets were utilised as “ground truth” delineations and atlas templates, from which 30 atlas-based auto-segmentations were created. We applied the Jaccard Conformity Index (JCI) to calculate over-and underlap between the volumes, describing auto-segmentation accuracy. Three clinicians evaluated the auto-segmentations, documenting the number of corrected slices per structure, the clinical significance of corrections, and using a 4-level grading system to define the CT scan acceptability. Results The auto-segmentation accuracy results and corrected CT slice numbers per structure are displayed in figure 1 and corrected, clinically significant slices per structure in figure 2 (a) Our analysis indicated significant correlations between JCI values and the number of corrected CT slices, and JCI values and clinically significant clinician corrections for the breast structure: p = 0.007 (ρ = -.484), and p = 0.018 (ρ = .430), respectively. The JCI values for breast cases with clinically significant corrections ranged between 0.5 and 0.9 (0.5 indicating poor conformance and 0.9, good conformance to the “ground truth”), justifying the need for clinician intervention, despite a strong correlation. Clinicians, A and B graded all CT scans as “good with limited correcting required.” Clinician C graded 24 CT scans as “acceptable, modification required” and 6 cases as “not acceptable, delete.” No scans were rated as “excellent, no correction required.” Figure 2 (b).