ESTRO 2024 - Abstract Book

S3024

Physics - Autosegmentation

ESTRO 2024

To test this methodology, three datasets of male pelvis MR scans were available, used for online adaptation on a 1.5 T MR-linac: 1) A labeled training dataset (11 pts, 58 scans) used for the selection of the features that correlate with segmentation performance, 2) An unlabeled dataset of the receiving hospital (148 pts, 1299 scans) to train the anomaly detection using the features provided, 3) A labeled test dataset of the receiving hospital (12 pts, 60 scans) to verify whether representative cases for clinical practice were correctly selected. A 3D nnU-Net [1] segmentation model for male pelvic MR was available and applied to all three dataset to predict rectum and bladder segmentations. For each segmentation, image and shape features (e.g. sphericity, major axis length, volume) were calculated [2, 3]. For rectum and bladder individually, 5 features were selected based on their correlation with the segmentation performance (surface dice 3mm) on dataset 1. Then an isolation forest was trained on dataset 2 using the selected features, and applied to dataset 3 to obtain anomaly scores. These scores can then be used to select cases for manual evaluation. As a reference, we calculated the expected segmentation performance spread over dataset 3 when randomly selecting N cases for manual evaluation (for each N cases, random selection was performed 100 times and the average value is presented).

Results:

Figure 1. shows the segmentation performance (surface dice 3mm) of dataset 3 plotted against the anomaly scores per case. Not all cases with a high anomaly score are incorrect, however, with increasing anomaly score the range of the segmentation performance scores increases. Figure 2. shows the standard deviation of the surface dice plotted against selecting the N worst cases based on the anomaly score, with random choice as a reference. Case selection based on the anomaly scores provides a bigger spread in segmentation performance than when randomly selecting cases.