A geometric and dosimetric comparison of three AI-based autocontouring packages in the head and neck region

Introduction: AI-based autocontouring products claim to be able to segment organs with accuracy comparable to humans. We compare the geometric and dosimetric performance of three AI-based autocontouring packages (Autocontour 2.5.6, (“RF”); Annotate 2.3.1, (“TP”) and RT-Mind_AI 1.0, (“MM”)) in the head and neck region. Methods: We generated 14 organ at risk (OAR) autocontours on 13 computed tomography (CT) image sets. They were compared with clinical (human-generated) contours. The geometric differences were quantified by calculating Dice coefficients and Hausdorff distances. The autocontours were compared visually with the clinical controus by an expert physician. The autocontour sets were also ranked for accuracy by two physicians. The dosimetric effects were evaluated by recalculating treatment plans on the autocontoured CT sets. Results: RF and TP slightly outperformed MM in geometric metrics (the percentage of OARs having mean Dice coefficients > 0.7 was RF 57.1 %, TP 64.3 % and MM 50.0%). The physician judged RF and TP contours to be more anatomically accurate, on average, than the manual contours (manual contour mean accuracy score 2.49, RF 2.28, MM 3.24, TP 1.93). The mean scores given to the autocontours by the two physicians were better for RF and TP, compared to MM (RF 1.86, MM 2.36, TP 1.77). The dosimetric differences were similar for all three programs and were not strongly correlated with the geometric differences. Conclusions: The performance of the three autocontouring packages in the head and neck region is similar, with TP and RF slightly outperforming MM. The correlation between geometric and dosimetric metrics is not strong, and dosimetric evaluation is therefore recommended before clinical use of autocontouring software.