Efficient Epistemic Uncertainty Estimation in Cerebrovascular Segmentation

TL;DR

This work tackles the trustworthiness of DL-based cerebrovascular segmentation by introducing an efficient epistemic uncertainty framework built on a Bayesian-approximate ensemble. The model uses a 3D U-Net backbone with dynamic layer-wise Bernoulli ensemble components and a two-stage training scheme, emphasizing uncertain regions to improve learning. Experiments on COSTA/IXI data show that the ensemble can outperform baselines on in-distribution data and provide uncertainty maps that correlate with segmentation errors, with higher uncertainty observable on out-of-distribution data and enabling uncertainty-guided corrections. The approach offers a practical path toward uncertainty-aware clinical segmentation, with future work suggested on incorporating aleatoric uncertainty and backbone improvements to further boost generalization and trust.

Abstract

Brain vessel segmentation of MR scans is a critical step in the diagnosis of cerebrovascular diseases. Due to the fine vessel structure, manual vessel segmentation is time consuming. Therefore, automatic deep learning (DL) based segmentation techniques are intensively investigated. As conventional DL models yield a high complexity and lack an indication of decision reliability, they are often considered as not trustworthy. This work aims to increase trust in DL based models by incorporating epistemic uncertainty quantification into cerebrovascular segmentation models for the first time. By implementing an efficient ensemble model combining the advantages of Bayesian Approximation and Deep Ensembles, we aim to overcome the high computational costs of conventional probabilistic networks. Areas of high model uncertainty and erroneous predictions are aligned which demonstrates the effectiveness and reliability of the approach. We perform extensive experiments applying the ensemble model on out-of-distribution (OOD) data. We demonstrate that for OOD-images, the estimated uncertainty increases. Additionally, omitting highly uncertain areas improves the segmentation quality, both for in- and out-of-distribution data. The ensemble model explains its limitations in a reliable manner and can maintain trustworthiness also for OOD data and could be considered in clinical applications

Figures (5)

• Figure 1: Efficient Ensemble Architecture: Ensemble members are created by selecting the layers randomly at each iteration. In step one the model is trained to minimize cross entropy loss, in step two N samples are generated and their variance is used to estimate $L_U$.
• Figure 2: Segmentation error maps of all networks: First row: Baseline results; Second row: Ensemble results. True positive segmented voxels are displayed in red, false positives in yellow, and false negatives in green
• Figure 3: Qualitative results of segmentation and UQ for the three architectures under consideration. Sagittal view of one MRA image. From left to right, we show voxelwise variance (in purple/yellow) and segmentation error overlaid on the input image for our 3D U-Net, DVN and the half-sized U-Net. In the error map, true positives are displayed in red, false positives in yellow, and false negatives in green.
• Figure 4: Left (A): clDice vs. variance threshold for our network; Right (B): Average variance of test images for all datasets.
• Figure 5: Segmentation results of OOD datasets