💻 computer science

Quantification of Uncertainty with Adversarial Models in Medical Image Segmentation

The paper introduces QUAM-SM, a post-hoc adversarial framework that quantifies pixel-level uncertainty in medical image segmentation by identifying fragile regions through targeted perturbations, thereby improving reliability and distinguishing between epistemic and aleatoric uncertainties for critical clinical applications.

Original authors: Hana Jebril, Thomas Pinetz, Günter Klambauer, Hrvoje Bogunović

Published 2026-06-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hana Jebril, Thomas Pinetz, Günter Klambauer, Hrvoje Bogunović

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are hiring a highly skilled robot doctor to look at medical scans (like eye photos or MRI images) and draw lines around tumors or specific organs. This robot is very fast and usually gets the shape right. But here's the problem: sometimes, the robot is too confident. It draws a line with 100% certainty, even if it's actually standing on shaky ground, right at the edge of a blurry spot where the image is unclear.

If a human doctor relies on that "100% sure" line, they might make a mistake. They might think a tumor is growing when it's just a smudge, or vice versa.

This paper introduces a new tool called QUAM-SM to fix this. Think of it as a "stress test" for the robot's confidence.

The Core Idea: The "Devil's Advocate" Robot

Instead of just asking the robot, "What do you think?" the researchers ask a second, mischievous robot (an adversarial model) to try and trick the first one.

The Setup: The main robot draws its picture (the segmentation).
The Attack: The mischievous robot tries to nudge the image just a tiny bit—like adding a speck of dust or slightly blurring an edge—to see if it can force the main robot to change its mind and draw a different line.
The Discovery: If the main robot changes its drawing easily with just a tiny nudge, that spot is marked as "Fragile." It means the robot wasn't actually sure; it was just guessing confidently. If the robot stays firm despite the nudge, that spot is marked as "Stable."

Two Types of "Not Sure"

The paper explains that there are two different reasons a robot might be shaky, and QUAM-SM is smart enough to tell them apart:

The "Messy Data" Problem (Aleatoric Uncertainty): Imagine a photo where the tumor edge is naturally blurry because the patient moved or the machine has a limit. Even a perfect human doctor would disagree on exactly where the line goes. This is data noise. The paper shows that QUAM-SM is great at spotting these blurry areas, matching the disagreements you see when multiple human experts look at the same scan.
The "Confused Brain" Problem (Epistemic Uncertainty): Imagine the robot has never seen a tumor shaped like a star before. It's not that the picture is blurry; it's that the robot's training didn't cover this specific shape. This is model ignorance. QUAM-SM also finds these spots where the robot is guessing because it lacks experience.

How It Works (The Analogy)

Think of the robot's decision-making as a tightrope walker.

Standard methods just ask the walker, "Are you steady?" and the walker says, "Yes!" even if they are wobbling.
QUAM-SM sends a gust of wind (the adversarial search) at the walker.
- If the walker wobbles or falls, QUAM-SM marks that spot on the map as High Uncertainty.
- If the walker stays perfectly still, it marks it as Low Uncertainty.

The researchers tested this on two real medical datasets: one with eye scans (optic discs) and one with prostate MRI scans. They compared QUAM-SM against other popular methods (like "Deep Ensembles" or "Monte Carlo Dropout").

The Results

The paper claims that QUAM-SM is better at the job than the others because:

It's more honest: It admits when it's unsure more accurately than the other methods.
It matches human experts: When multiple human doctors look at the same scan and disagree on the boundary, QUAM-SM's "uncertainty map" lights up exactly in those same spots.
It separates the noise: It successfully distinguishes between "the picture is blurry" (Aleatoric) and "I don't know this shape" (Epistemic).

The Bottom Line

The paper concludes that by using this "stress test" method, we can create a map that shows doctors exactly where the AI is confident and where it is vulnerable. This helps prevent the AI from tricking doctors with false confidence, making the technology safer to use in real medical workflows.

Note: The authors specifically mention that this method requires access to the training data (the "gym" where the robot learned) to work properly, but it doesn't require the robot to be retrained from scratch; it's a "post-hoc" (after-the-fact) check.

Technical Summary: QUAM-SM for Medical Image Segmentation

Problem Statement
The deployment of deep learning models in clinical workflows, particularly for medical image segmentation, requires not only high accuracy but also reliable pixel-level uncertainty quantification. Standard deep learning models often suffer from miscalibration, producing overconfident predictions that mask underlying failure modes, especially at object boundaries and in subtle pathological regions. This overconfidence can lead to incorrect clinical decisions, such as misinterpreting true disease progression as pseudo-progression. Existing uncertainty estimation methods, including Deep Ensembles, Bayesian approximations (e.g., Monte Carlo Dropout), and evidential approaches, often rely on stochastic sampling within limited regions of the posterior distribution. Consequently, they may fail to capture the full spectrum of potential prediction failures or adequately distinguish between different sources of uncertainty.

Methodology: QUAM-SM
The authors propose QUAM-SM (Quantification of Uncertainty with Adversarial Models in Medical Image Segmentation), a post-hoc framework designed to identify "adversarially fragile" pixels—regions where a model's decision is easily flipped to an incorrect class. The framework operates through the following stages:

Reference Segmentation: A fixed, pre-trained model generates a reference segmentation ( $y_{ref}$ ) for a test image, assuming access to the training data distribution to maintain consistency with the learned data manifold.
Adversarial Model Search: The core of the method involves a targeted adversarial search to generate alternative segmentation hypotheses. The algorithm optimizes a model ( $\tilde{w}$ $\tilde{w}$ ) using two complementary loss terms:
- Penalty Loss ( $L_{pen}$ ): Enforces consistency with the training distribution to ensure the adversarial models remain plausible (posterior plausibility).
- Adversarial Loss ( $L_{adv}$ ): Drives the prediction toward a specific targeted label ( $y_{mask}$ ) to explore alternative hypotheses.
  The paper investigates three strategies for generating $y_{mask}$ :
- Global Extreme Masks (GEM): Constant maps of all background or all foreground.
- Training-based Masks (TBM): Realistic anatomical configurations sampled from the training set.
- Morphology-augmented Masks (MAM): Local structural variations near boundaries created via morphological operations (dilation/erosion) on the reference prediction.
Uncertainty Quantification via Mixture Importance Sampling (MIS): The adversarial search produces a set of plausible model weights. To derive a statistically consistent estimate of the predictive distribution, the authors employ Mixture Importance Sampling. Models are weighted based on their penalty loss on the training data, ensuring that models adhering to the training distribution contribute more significantly.
Decomposition of Uncertainty: Using the weighted predictions, the framework decomposes total uncertainty into two distinct components based on the Expected Entropy (EE) framework:
- Aleatoric Uncertainty: Represents data noise and annotation variability (modeled as the expected entropy of individual predictions).
- Epistemic Uncertainty: Represents model uncertainty (modeled as the Mutual Information or disagreement between the weighted ensemble predictions).

Key Contributions

Novel Framework: The authors introduce a post-hoc uncertainty quantification framework for medical image segmentation that leverages targeted adversarial search to identify fragile, pixel-wise predictions.
Uncertainty Disentanglement: The method explicitly distinguishes between aleatoric and epistemic uncertainty, demonstrating improved consistency with inter-rater variability on multiple public datasets.
Enhanced Posterior Coverage: By extending the concept of adversarial models to the segmentation domain, the method achieves improved coverage of the posterior by highlighting pixels most susceptible to being "surely false," offering a more sensitive error map than traditional density-based methods.

Experimental Results
The QUAM-SM framework was evaluated on two public datasets with multiple expert annotations:

REFUGE (DR): Fundus photography images with seven independent annotations for optic disc and cup segmentation.
QUBIQ2021 (DQ): Prostate tumor MRI volumes annotated by six experts.

The method was compared against Monte Carlo Dropout, Deep Ensembles, Probabilistic U-Net, and the SURE evidential framework. Performance was measured using the Pearson Correlation Coefficient (PCC) and coefficient of determination ( $R^2$ ) against reference entropy maps derived from multiple annotators, as well as Soft Dice (SDice) scores for segmentation accuracy.

Quantitative Performance: QUAM-SM achieved the highest PCC and $R^2$ values across epistemic, aleatoric, and total uncertainty categories for both datasets, indicating superior alignment with reference uncertainty. It also consistently outperformed competing approaches in segmentation accuracy (SDice).
Ablation Studies: Experiments confirmed that Morphology-augmented Masks (MAM) yielded the best performance among attack strategies. Furthermore, increasing the number of adversarial iterations ( $M$ ) consistently improved performance, with optimal results observed at higher iteration counts.
Qualitative Results: Visualizations demonstrated that QUAM-SM closely matches ground-truth entropy maps, accurately capturing regions of high inter-annotator uncertainty, particularly at boundaries.

Significance and Claims
The paper claims that QUAM-SM addresses the critical issue of overconfident predictions in medical AI by providing a mechanism to actively seek out predictive instabilities. By disentangling aleatoric and epistemic uncertainty, the framework offers a more nuanced view of model reliability. The authors note that the estimated aleatoric uncertainty correlates more strongly with inter-rater variability than total uncertainty, highlighting its specific relevance for capturing annotation ambiguity. This capability supports safer clinical decision-making by distinguishing true pathological changes from artifacts and identifying regions where model decisions are most vulnerable. While the approach requires access to the training distribution, the authors assert that it provides interpretable uncertainty maps that enhance model assessment in realistic clinical benchmarks.