ESTRO 2024 - Abstract Book

S3000

Physics - Autosegmentation

ESTRO 2024

how much manual correction autocontours require and whether subsequent changes make a clinically significant impact on the dose to the patient [1]. In this work the AI contouring package, Siemens (Erlangen, Germany) Direct Organs is compared to the gold standard and contours generated via thresholding for the lung and the heart dosimetrically and geometrically to determine whether correction is warranted.

Material/Methods:

Gold standard contours were created by skilled radiographers for a arbitrarily selected series of 30 right breast and 31 left breast female patients; the heart entirely manually contoured and the lungs using an edited contour based on density thresholding in ProSoma (Medcom, France). Tangential plans were created in Philips (Netherlands) Pinnacle with either 26 Gy in 5# or 40.05 Gy in 15# according to departmental protocol. The 26Gy plans were rescaled to 40.05 Gy for direct comparison of plans at the same dose level (OAR constraints are linearly scaled between prescriptions). AI contours for the heart and the lungs were generated using Direct Organs, and for the lungs the unedited ProSoma contour was redrawn. The Hausdorf Distance (HD) was calculated in Pinnacle and the DVH statistics extracted for all contours. The DVH statistics for the autocontours were compared to the manual contours via the Wilcoxon rank sum test to determine if the autocontours would have given a reliable estimate of dose without correction. In addition, an attempt was made to investigate the specific cases of outlier patients (obese, chest wall, male) to investigate the performance of the algorithms for these cases via the HD.

Results:

Figure 1 shows the HD distance for the autocontours, which is sometimes very large and larger than the interobserver error (shown on figure) [2]. On inspection most of the errors were in the bronchial tree, with the ProSoma contour occasionally contouring the opposite lung. Thus there are errors in the contours, but the Direct Organs contours have smaller errors than ProSoma contours. Table 1 shows the mean DVH statistics for the organs with the ProSoma contour showing the largest deviation from the gold standard. However, the DVH statistic differences are not statistically significant between the autocontours and gold standard although the ProSoma contours do approach significance. For the outlier patients the means HD and standard deviations overlap with the main group of patients, implying the packages are robust for these patients.