Session Item

Accurate OAR segmentation is essential for radiotherapy but labor-intensive. Automatic OAR delineation can save time and and resources and improve reproducibility in radiotherapy. Our aim was to assess the performance of a commercial automatic segmentation model for HNC patients in various positions, focusing on the implementation for routine clinical use.

The 3D CNN U-Net Deep Learning model for head and neck developed by RaySearch Laboratories AB (RSL, Sweden) was assessed in this study. Autocontouring was performed on 22 OARs for 137 head and neck CT scans of 98 adult and pediatric patients in the following 8 positions, relevant for particle therapy with fixed beam lines: 1) head-first-supine (HFS) straight ; 2) HFS with head hyperextension; 3 & 4) head first decubitus left and right; 5 & 6) HFS with head rotation left and right; 7 & 8) head-first-prone with head rotation left and right. A geometrical comparison of the autocontours and the manual, clinically used segmentations was performed, using the Dice Score Coefficient (DSC) and the Hausdorff Distance (HD) and compared to interobserver variability (IOV), where available . For 20 CT scans in positions 1 and 2, additional qualitative and dosimetric  analyses were performed. Qualitative scoring was performed on a 0-3 scale based on the amount of time saved in manual contouring by three independent observers. ROIs with a median score of ≥2 were considered useful for daily practice. Dosimetric analysis was performed by comparing the average (Davg) and near-maximum (D2%) dose using the Mann-Whitney U test. p<0.05 was considered significant.

Based on the geometric similarity metrics, the model performance in positions 1 and 2  was in the same range as the IOV . E.g., for the brainstem, the mean DSC was 0.86±0.05 and 0.84±0.09 (IOV DSC = 0.88) and the mean HD was 4.16±1.88 mm and 7.49±12.00 mm (IOV HD = 4.0 mm) in the HFS straight and hyperextension group, respectively (figure 1). The model performance for adult and pediatric scans was similar, with only the brain (p=0.015) and the right eye (p=0.046) showing significant differences in DSC between the two groups . Model performance in the other positions was extremely unstable, including cases of left-right confusion and erroneous localization of OARs. For the additional analyses, we found a median score of ≥2 for 13/18 ROIs for the qualitative analysis. The dosimetric analysis yielded no significant difference for any ROIs when comparing D2% and Davg for manual and automatic contours within the same treatment plan (figure 2).

Our study showed that the current geometrical performance of the RSL automatic segmentation model is not suited for use in daily clinical practice in its current form for all patient positions. For HFS straight and hyperextended scans, we found that 13/18 automatic segmentations were suited for use in daily clinical practice from a geometrical, dosimetric and qualitative perspective.