S1610
Physics - Autosegmentation
ESTRO 2026
1. Kensen, C.M., et al., Physics and Imaging in Radiation Oncology, 2024. 32: p. 100648.2. Elmahdy, M.S., et al., 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI). 2020. IEEE.3. Isensee, F., et al., Nature methods, 2021. 18(2): p. 203-211.4. Wasserthal, J., et al., Radiology: Artificial Intelligence, 2023. 5(5): p. e230024. Keywords: autosegmentation, patient-specific fine- tuning
Radiomics dataset for tumor segmentation. For OOD detection, we used scans from the NSCLC Radiogenomics (N=140) as the ID cohort, and pulmonary embolism (N=1400), negative COVID-19 (N=120), and kidney cancers (N=1400) as distinct OOD cohorts. For training the OOD detectors, the frozen segmentation model first delineated potential tumor regions on all the scans. Then, we repurposed the frozen encoder from the model to extract multi-scale deep features from tumor-anchored 3D crops (128 x 128 x 128 voxels) across all encoder stages. These features, collected from ID and OOD cohorts, were used to train cohort-specific random forest classifiers (1000 trees, depth=20) with outlier exposure [2]. Once trained, these random forest classifiers were frozen and used on new incoming scans to be classified as either ID or OOD (Figure 1 e). OOD detection performance was measured using area under the ROC curve (AUROC) and false positive rate at 95% true- positive rate (FPR95). We benchmarked against standard confidence-based approaches including MaxSoftmax [3], MaxLogit [4], energy-based scores [5], and random forest classifiers trained on hand-crafted radiomic features. Results: Our approach consistently and substantially outperformed all comparative methods (Table 1). For chest pathologies, our approach achieved AUROC/FPR95 of 96.74/18.54% for pulmonary embolism and 94.59%/24.36% for negative COVID-19 scans, exceeding the next-best benchmark by 6 and 4 points, respectively. For kidney cancer, we reached near-perfect separation with 99.97 AUROC and 0.03% FPR95, and while the confidence-based approaches approached our AUROC on this task, their FPR95 remained substantially higher, underscoring the superior discriminative power of the deep-features based random forest OOD detector.
Digital Poster 4734
Deep feature-based out-of-distribution detection improves safety of automated lung cancer tumor segmentation in radiotherapy workflows Aneesh Rangnekar, Harini Veeraraghavan Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, USA Purpose/Objective: Reliably accurate detection and segmentation of cancerous lesions from computed tomography (CT) scans is essential for automated treatment planning and cancer treatment response assessment. Automated tumor segmentation is critical for radiotherapy treatment planning and response assessment. While deep learning models achieve high accuracy on in-distribution (ID) data, they are vulnerable to producing confidently incorrect segmentations when applied to out-of-distribution (OOD) cases (Figure 1 a-d), such as non-cancerous pathologies or different anatomical sites, thereby posing safety risks in clinical workflows.
Conclusion: Our approach provides a practical, interpretable solution for detecting OOD scans in radiotherapy workflows. Future work will extend to additional tumor sites and variations across multi-institutional datasets. References: [1] Jiang, J., Tyagi, N., Tringale, K., Crane, C. and Veeraraghavan, H., 2022, September. Self-supervised 3D anatomy segmentation using self-distilled masked
Material/Methods: A Swin Transformer encoder, pretrained using the self- distilled masked image transformers (SMIT [1]) on 10,000 unlabeled 3D CT scans of head to pelvis disease sites, was fine-tuned with a convolutional U- Net decoder on 317 lung cancer CTs from the NSCLC
Made with FlippingBook - Share PDF online