ESTRO 2024 - Abstract Book

S3153

Physics - Autosegmentation

ESTRO 2024

Figure 1. Dice Similarity Coefficient comparing the manually segmented prostate gland, the centre gland and peripheral zone, to the auto segmented contours across a range of slice thicknesses.

The model was trained to an overall accuracy of 72.0% using this dataset, with the central gland having a DSC of 79.8% and the peripheral zone 70.5%. A degradation in model performance was observed when images were resampled to larger slice thicknesses, this is shown in figure 1 which presents the prostate gland DSC for all slice thicknesses. The number of samples which were autosegmented with a non-zero volume at slice thicknesses of 7 mm and 8 mm were 12, compared to the initial 15. The model was only stable for 3 mm slice thickness as all metrics offered a statistically similar (p > 0.05) set of segmentation scores compared to the autosegmentations at 2 mm, with 8 mm slice thickness being the only slice thickness where all segmentation metrics were significantly different to the metrics at 2 mm. The absolute volume difference was found to be comparable to the original slice thickness across slice thicknesses up to 8 mm, but the degrading DSC and HD highlights the segmentation volume may not contain the correct image voxels. Table 1. Median segmentation comparison values are presented for each slice thickness compared to the original manual segmentations. P-values in brackets were calculated using a Wilcoxon rank sum test comparing segment comparison scores for autosegmentations on resampled images to autosegmentations the original images of slice thickness 2 mm. The number of samples per slice thickness which the model autosegmented (non-zero DSC) are listed.