[Paper Review] A Decoupled Uncertainty Model for MRI Segmentation Quality Estimation

Richard Shaw, Carole H. Sudre+3 more|arXiv (Cornell University)|Sep 6, 2021

Radiomics and Machine Learning in Medical Imaging38 references4 citations

TL;DR

This paper proposes a decoupled uncertainty model using a heteroscedastic noise model in a self-supervised Bayesian deep learning framework to estimate task-specific MRI segmentation quality. By augmenting training data with k-space artefacts, the method decouples uncertainty into distinct sources (e.g., motion, RF spikes), enabling robust, real-time quality estimation that outperforms traditional metrics like SNR and CNR in predicting segmentation accuracy.

ABSTRACT

Quality control (QC) of MR images is essential to ensure that downstream analyses such as segmentation can be performed successfully. Currently, QC is predominantly performed visually and subjectively, at significant time and operator cost. We aim to automate the process using a probabilistic network that estimates segmentation uncertainty through a heteroscedastic noise model, providing a measure of task-specific quality. By augmenting training images with k-space artefacts, we propose a novel CNN architecture to decouple sources of uncertainty related to the task and different k-space artefacts in a self-supervised manner. This enables the prediction of separate uncertainties for different types of data degradation. While the uncertainty predictions reflect the presence and severity of artefacts, the network provides more robust and generalisable segmentation predictions given the quality of the data. We show that models trained with artefact augmentation provide informative measures of uncertainty on both simulated artefacts and problematic real-world images identified by human raters, both qualitatively and quantitatively in the form of error bars on volume measurements. Relating artefact uncertainty to segmentation Dice scores, we observe that our uncertainty predictions provide a better estimate of MRI quality from the point of view of the task (gray matter segmentation) compared to commonly used metrics of quality including signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR), hence providing a real-time quality metric indicative of segmentation quality.

Motivation & Objective

• To address the limitations of subjective, time-intensive visual quality control in MRI by automating quality assessment based on task-specific performance.
• To develop a method that estimates uncertainty not just globally, but decoupled by distinct artefact types (e.g., motion, RF spikes) to enable targeted intervention.
• To improve segmentation robustness and generalization by training with synthetic k-space artefacts, simulating real-world data degradation.
• To validate that uncertainty predictions correlate better with segmentation performance (Dice score) than conventional image quality metrics like SNR and CNR.

Proposed method

• A student-teacher CNN architecture is used, where the teacher network produces uncertainty estimates via a heteroscedastic noise model.
• Training data is augmented with realistic k-space artefacts (e.g., motion, RF spikes) to simulate diverse image degradation without requiring paired clean/artefact data.
• The model decouples uncertainty into independent components per artefact type by conditioning the noise variance on artefact-specific embeddings.
• Uncertainty is estimated per voxel using a Bayesian neural network, with variance predicted per artefact type to reflect task-specific risk.
• The method is self-supervised: no ground-truth uncertainty is required; uncertainty is learned from the model’s predictive variance during training.
• A patch-based inference strategy is applied to handle 3D MRI volumes, with boundary artefacts mitigated via overlapping patches and averaging.

Experimental results

Research questions

• RQ1 Can a deep learning model estimate segmentation uncertainty in a way that reflects the true impact of MRI artefacts on downstream task performance?
• RQ2 Does decoupling uncertainty by artefact type improve the interpretability and utility of uncertainty predictions for clinical and research workflows?
• RQ3 Can uncertainty predictions derived from k-space artefact augmentation outperform standard image quality metrics (SNR, CNR) in predicting segmentation accuracy?
• RQ4 How robust are the uncertainty estimates on real-world scans with clinically problematic artefacts not seen during training?
• RQ5 To what extent can the model generalize to unseen artefact types or real-world data without fine-tuning?

Key findings

• The proposed model produces uncertainty estimates that correlate more strongly with segmentation Dice scores than SNR or CNR, demonstrating superior predictive power for task-specific quality.
• Decoupled uncertainty predictions successfully localize and distinguish between different artefact types (e.g., motion vs. RF spikes), even when they co-occur.
• On real-world scans identified as problematic by human raters, the model’s uncertainty maps clearly highlight regions affected by artefacts, confirming clinical relevance.
• The model generalizes from simulated artefacts to real scans, with uncertainty predictions reflecting actual segmentation errors in volume measurements.
• The method enables real-time quality estimation during inference, with uncertainty quantified as error bars on segmentation volumes.
• Despite limitations in distinguishing visually similar artefacts (e.g., blur vs. motion), the model shows transferability across domains and scalability to multi-class segmentation tasks.