S1562
Physics - Autosegmentation
ESTRO 2026
planning. The correction process could become more efficient if therapists were alerted to areas containing possible segmentation errors.Uncertainty estimation techniques can be used to assess deep learning segmentation quality4. These techniques are promising, as they show reasonable correlation with segmentation errors3. However, because predictive uncertainty is inherently high around class boundaries, separating distinct areas which possibly require correction is challenging, limiting utility in clinical workflows. Therefore, we propose a refined method for presenting uncertainty in distinct and spatially coherent uncertainty areas. Material/Methods: This study analyzes a dataset of 578 pre-treatment 3D T2-weighted MRI scans, with automatic segmentations of the prostate, bladder, rectum, and femurs generated by the clinically deployed convolutional neural network architecture DeepMedic2. Clinical experts review and manually correct these segmentations, which results in the correction maps. The corrections are then divided into distinct connected components. Model uncertainty is estimated by Monte-Carlo Dropout1. We perform 50 stochastic forward passes per pre-treatment segmentation to estimate predictive uncertainty.To reduce the effect of inherent uncertainty on class boundaries and emphasize spatial uncertainty coherence, the uncertainty maps are convolved with a Gaussian kernel with standard deviation of σ voxels. Subsequently, the maps are thresholded to create binary uncertainty maps, which are then labeled as separate connected components. Each connected component is counted as a unique uncertain area. Optimal thresholds are calculated from the mean optimal voxel-wise F1-score on a separate calibration set. The remaining validation set is used in the experiments to indicate expected clinical performance. Results:
Fig. 1 shows the impact of increasing σ on thresholded uncertainty maps. The uncertain areas become more spatially coherent and less concentrated around class boundaries.
Fig. 2 shows that smoothing the uncertainty maps increases the mean F1-score and reduce the number of distinct uncertain areas. However, as σ increases, the recall decreases, indicating that more corrections are missed. Adversely, uncertain areas will overlap
Made with FlippingBook - Share PDF online