S1620
Physics - Autosegmentation
ESTRO 2026
Mahalanobis distances[3] calculated using the development cohort distribution. Structure-specific Mahalanobis distance thresholds were established from histogram distributions of the development set, using empirically-determined cut-offs to flag potential outliers. Additionally, structures with multiple disconnected volumes were flagged as outliers, except for naturally discontinuous structures (e.g. rib cage). Flagged cases were visually inspected and classified as segmentation errors or atypical anatomy to assess method performance. Results: Six shape features were selected: bounding box dimensions (x, y, z), elongation, surface-to-volume ratio, and sphericity. Mahalanobis distance thresholds of 5 (most structures) and 7 (bony structures, body) identified outliers with median rates of 1.4% (IQR: 0.7- 3.9%) in development and 1.8% (IQR: 0.9-4.1%) in test cohorts across all structures (Fig.1). Visual inspection revealed outliers included both segmentation errors (Dev. median:47.9% (IQR: 16.1-69.6%) Test median 39.8% (IQR: 18.7-72.3%) of outliers) and anatomically valid outliers (Dev. median: 52.1% (IQR: 30.4-83.9%), Test median: 60.2% (IQR: 27.7-81.3%) of outliers). Detected outliers exhibited atypical anatomy due to: (a) morphological variations, (b) incomplete scan coverage at superior/inferior CT boundaries.
The approach accepts false positives from anatomical outliers as an acceptable trade-off to minimize undetected segmentation errors that could compromise dose-response analysis validity. This enables automated segmentation QA across large- scale lung cancer datasets. A study of the False Negative rate is pending further research. References: 1. Wasserthal J, Breit HC, Meyer MT, et al. TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images. Radiology: Artificial Intelligence. 2023;5(5):e230024. doi:10.1148/ryai.230024 2. Finnegan RN, Chin V, Chlap P, et al. Open-source, fully-automated hybrid cardiac substructure segmentation: development and optimisation. Phys Eng Sci Med. 2023;46(2):377-393. doi:10.1007/s13246-023-01231-w 3. Rousseeuw PJ, van Zomeren BC. Unmasking multivariate outliers and leverage points. Journal of the American Statistical Association. 1990;85(411):633-639. doi:10.1080/01621459.1990.10474920 Keywords: Autosegmentation, Quality Assurance Digital Poster 5181 From manual to fully automated: AI segmentation accuracy for MR-Linac adaptive workflows evaluated by dosimetric criteria David Tilly 1,2 , Saya Tulin Mayeedi 3 , Nina Tilly 1,2 1 Immunology, Genetics and Pathology, Uppsala University, Uppsala, Sweden. 2 Medical Physics, Uppsala University Hospital, Uppsala, Sweden. 3 Medical Radiation Science, Stockholm University, Uppsala, Sweden Purpose/Objective: To determine the segmentation accuracy required for AI-driven fully automated online adaptive MR-linac workflows in prostate radiotherapy by evaluating the dosimetric impact of treatment plans based on AI- generated segmentations against clinical criteria. Material/Methods: A treatment planning study was performed using prostate cancer patients treated with ultra-hypo- fractionated radiotherapy (6.1 Gy × 7 fractions) on an MR-linac. A previously developed U-Net-based AI model provided segmentations for the clinical target volume (CTV) and organs of interest (OOI) bladder and rectum. Twenty-four fractions were selected from the model’s test set to represent a similar Dice range as the overall test set. Rivaling treatment plans were generated for two segmentation types:AI-generated segmentation (AI)Expert segmentation segmentation (Ground truth (GT))Plans based on AI segmentations were evaluated using GT segmentation and compared to GT-based plans. An isotropic 3 mm margin was
Conclusion: Morphometric outlier detection showed promise in automated QA identifying problematic segmentations.
Made with FlippingBook - Share PDF online