S1590
Physics - Autosegmentation
ESTRO 2026
hamper their use in online adaptive radiotherapy. Input reconstruction-based methods (IR) are a low- cost alternative, that reconstructs the input image (e.g., CT) and computes uncertainty as the difference between the reconstructed and input images in a single-inference. Although this has been successfully applied to regression tasks (head-and-neck dose prediction [1]), the extension to segmentation tasks remains unexplored.We propose a two-inference method that efficiently adapts IR for segmentation and benchmark it with two IR architectures against state- of-the-art UQ methods. Material/Methods:
see Figure 2). In contrast, IRsDice reaches correlation comparable to state-of-the-art UQ methods, while keeping low-inference cost (N=2).Single-inference model used for IR has comparable sDice to multiple- inference methods for BraTS and Pancreas datasets, but a median 0.1 lower for LUNG and AbdominalCT1k.IR architectures are comparable, but occasional differences in correlation between models for IRsDice and IR2sDice suggest the architecture should be optimized on a case-by-case basis. Conclusion: The proposed IRsDice provides a lightweight UQ method competitive with state-of-the-art, without modifying the segmentation model or being dependent on its training parameters or data augmentation. It is compatible with all IR architectures and requires only two forward passes and preserves the segmentation performance of the original model. References: [1] Margerie Huet-Dastarac, Dan Nguyen, Eleonore Longton, Steve Jiang, John Lee, and Ana BarragánMontero. Can input reconstruction be used to directly estimate uncertainty of a dose predictionu- net model? Medical Physics, 51(10):7369–7377, July 2024.[2] Matthew Ng, Fumin Guo, Labonny Biswas, Steffen Erhard Petersen, Stefan K. Piechnik, StefanNeubauer, and Graham A. Wright. Estimating uncertainty in neural networks for cardiac mrisegmentation: A benchmark study. IEEE Transactions on Biomedical Engineering, 70:1955– 1966,2020. Keywords: Uncertainty quantification Digital Poster 4001 Evaluation of a Commercial AI-Based Cranial Tumor Segmentation Tool for Brain Metastases in Stereotactic Radiosurgery Larissa Kilian 1,2 , Julia Bauer 1 , Reinhard Schild 1,2 , Diana Sladek 1,2 , Daniel Zips 1 , Gueliz Acker 2,3 , Carolin Senger 1,2 1 Department of Radiation Oncology, Charité - Universitaetsmedizin Berlin, Berlin, Germany. 2 Charité CyberKnife Center, Charité - Universitaetsmedizin Berlin, Berlin, Germany. 3 Department of Neurosurgery, Charité - Universitaetsmedizin Berlin, Berlin, Germany Purpose/Objective: Manual segmentation of brain metastases (BM) in MRI is a labor-intensive task requiring substantial clinical expertise and time. Recent randomized evidence further supports the use of stereotactic radiosurgery (SRS) for patients with up to 20 BM, highlighting its expanding role [1]. AI-based auto-segmentation has emerged as a promising tool to accelerate clinical workflows and improve reproducibility [2]. However, the clinical performance of commercial AI
We apply IR method to a Res-UNet (Figure 1) and evaluate it for tumor segmentation error prediction in four datasets: BraTS (brain MRs); Pancreas, LUNG, and AbdominalCT1k (thorax/abdominal CTs). While previous work used mean squared error (IRMSE) as uncertainty metric [1], we propose a new approach tailored to segmentation tasks: a second segmentation is obtained from the reconstructed image, and the surface Dice (sDice) between the two segmentations quantifies the uncertainty (IRsDice).We compare two reconstruction model architectures (IR and IR2, see Figure 1) and evaluate them (inferences N=2) against multiple-inference alternatives, where sDice between each prediction and the average prediction quantifies uncertainty (as in [2]). These methods then have higher cost: MCDO (N=20), DE (N=10), and TTA (N=8).Spearman correlation between each uncertainty metric and the sDice between the prediction and ground truth evaluates the performance of all UQ methods. Results:
MSE is unsuitable as a metric for IR-based UQ in segmentation (IRMSE correlation < 0.3 for all datasets,
Made with FlippingBook - Share PDF online