ESTRO 2022 - Abstract Book

S300

Abstract book

ESTRO 2022

A. Boschetti 1 , C. Votta 1 , A. Re 1 , D. Piro 1 , M. Marras 1 , A. D'Aviero 1 , F. Catucci 1 , D. Cusumano 2 , C. Di Dio 1 , S. Menna 2 , M. Iezzi 3 , F.V. Quaranta 1 , C. Flore 1 , E.G. Sanna 1 , D. Piccari 1 , G.C. Mattiucci 2 , V. Valentini 2 1 Mater Olbia Hospital, Radiation Oncology Unit, Olbia (SS), Italy; 2 Fondazione Policlinico Universitario Agostino Gemelli IRCCS, Dipartimento di Diagnostica per Immagini, Radioterapia Oncologica ed Ematologia, Roma, Italy; 3 Istitituto di Radiologia, Dipartimento di Diagnostica per Immagini, Radioterapia Oncologica ed Ematologia, Roma, Italy Purpose or Objective Accurate delineation of organs at risk (OARs) is a crucial step of treatment planning but a significant bottleneck in a center workflow. Manual delineation is a time-consuming process and often suffers from significant inter-observer variability. Deep-learning based auto-segmentation has the potential to improve this step. We aimed to compare deep-learning-generated auto-contours (AC) with the ones made manually (MC) by expert Radiation Oncologists from a single center. Materials and Methods Radiotherapy planning computed tomography (CT) scans of patients undergone treatment in thoracic region were considered. MC of OARs were delineated by different radiation oncologists. The same scans were processed by a commercial deep learning auto-segmentation based software to generate AC. Two different protocols were used: breast protocol including thyroid, spinal cord and both lungs’ delineations, and thoracic protocol comprising additionally esophagus, aorta, and trachea contours. The same contouring guidelines used by AC software were referred to perform MC delineation. The MC were compared with AC using the Dice Similarity Coefficient (DSC) and 95% Hausdorff distance transform (DT). Results Thirty-three CT were analyzed (breast protocol n=25, thoracic protocol n=8). DSC and DT are showed in Fig. 1 and 2. Minimal differences were showed if we consider DSC of lungs (0.97 for both left and right lung and DT, respectively, 17.91 and 19.13 mm) and heart (mean DSC and 95% DT were 0.93 and 10.05 mm) The comparison of spinal cord delineation showed slightly worse DSC (0.85) and DT 6.4 mm), probably related to better accuracy of Artificial Intelligence in intercept density differences compared to human eye. Thyroid (n=25) was the organ with more noticeable differences: in this case the DSC and DT were respectively 0.7 and 9.87 mm. Esophagus, trachea and aorta were contoured on 8 scans. Their DSC were respectively 0.77, 0.82, 0.84, and DT were 14.81, 12.55 and 95.30 mm. No unacceptable AC were noticed.

Conclusion Although this is a preliminary analysis with a limited number of patients, deep-learning auto-segmentation seems to provide acceptable segmentation for thoracic OARs and even in less accurate organs, it could provide a starting point for review and manual adjustment. Data suggest that this could become a useful time-saving tool to optimize workload and resources in radiation therapy. Further studies to confirm its clinically viability are needed.