ESTRO 2024 - Abstract Book

S3085

Physics - Autosegmentation

ESTRO 2024

2068

Digital Poster

Impact of the availability of preliminary auto-contours on clinically used ones

Lucia Goanta, Adriana Faiella, Alessia Farneti, Giuseppe Sanguineti

IRCCS Regina Elena National Cancer Institute, Radiation Oncology, Rome, Italy

Purpose/Objective:

The correct delineation of target volumes and surrounding organs-at-risk (OAR) is time consuming and at risk of interobserver variability. Several autosegmentation models are emerging as useful tools within the routine clinical workflow. However, their implementation may lead to automation bias with potential overreliance on the proposed automated OAR. 1 The purpose of the present study is to investigate this issue.

Material/Methods:

We compared the difference between automatically generated and physician validated contours stratifying by the presence or absence of a preliminary auto-contour, assuming that if the difference is the same in the two scenarios, the impact of the availability of a preliminary contour is negligible. To this scope we decided to analyse the bladder and the rectum, two structures for which we have preliminary shown a strong agreement during the contouring phase among 3 senior radiation oncologists (ADF, ALF, GS). We first carried out a prospective study on 60 consecutive patients treated at our facility for localized prostate cancer (AUTO-GUIDED sub-group). Following an internal workflow, the simulation CT was preliminary sent to a dedicated software where preliminary auto-contours (Pr-AC) for both the bladder and the rectum were generated. Afterwards, the Pr-ACs were reviewed and modified by ADF, ALF, GS, as necessary. The final, clinically validated contours (CV-C1) were taken as reference to which the automated ones were compared in terms of Dice Similarity Coefficient (DSC), Jaccard (JC), Hausdorff distance (HD) and Mean Distance to Agreement (MDA). 2 The second step of our study consisted in the analysis of a second cohort of 60 patients with localized prostate cancer treated at our facility by the same physicians before the availability of the auto-contouring tools (NO AUTO-GUIDED sub-group). Within this cohort, we compared the contours for both the rectum and the bladder as drawn per clinical practice (CV-C2) without the support of the auto-segmented tool to those generated automatically. The latter ones were obtained post-hoc exclusively for research purposes and were not manually edited at all (Ph-AC). Also within this sub-group, CV-C2 and Ph-AC contours were compared in terms of DSC, JC, HD and MDA.

Finally, the differences in the selected metrics between Pr-AC and CV-C1 versus Ph-AC and CV-C2 were assessed using the Mann-Whitney U test and the Kruskar Willis test. P values <0.05 were considered statistically significant.

Results: