🤖 AI

GRC-ProbNet: Uncertainty-aware Feature Extraction for Cardiovascular Disease Classification

This paper introduces GRC-ProbNet, an uncertainty-aware framework that leverages deep ensembles to generate multiple segmentation masks and extract uncertainty features, thereby significantly improving cardiovascular disease classification performance compared to the deterministic baseline GRC-Net.

Original authors: Yash Shah, Omar Todd, Philipp Seeböck, Georg Langs, Ben Glocker, Raghav Mehta

Published 2026-07-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yash Shah, Omar Todd, Philipp Seeböck, Georg Langs, Ben Glocker, Raghav Mehta

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 your heart is a complex, squishy city inside your chest, and doctors use special 3D X-ray cameras (called CT scans) to take pictures of it. Sometimes, these pictures are a bit blurry or tricky to read, making it hard to tell if the city is healthy or if it's got a "traffic jam" (cardiovascular disease).

For a long time, computer programs tried to solve this by acting like a single, super-confident detective. They'd look at the scan, draw a sharp line around the heart parts, and say, "This is the left ventricle, and it's definitely healthy!" But here's the problem: even real human experts sometimes disagree on exactly where the lines should go because the heart is so wiggly and ambiguous. The old computer detective didn't know it was guessing; it just acted like it knew everything for sure.

Enter GRC-ProbNet, a new, super-smart team of detectives. Instead of sending just one detective to look at the scan, this system sends five slightly different detectives (a "deep ensemble") to the same picture. Each one has a tiny bit of a different perspective because they were trained with a different random starting point.

The "Five Detectives" Experiment

When these five detectives look at the same heart scan, they might all agree on the big picture, but they might squabble a little bit about the fuzzy edges.

The Old Way (GRC-Net): The system just took the average of their drawings and ignored the squabbling. It treated the result as a perfect, solid fact.
The New Way (GRC-ProbNet): The system pays close attention to where the detectives disagree. It creates a "worry map." If all five detectives agree the edge is sharp, the worry level is low. If they can't agree on where the edge is, the worry level is high.

The researchers tested this on two sets of heart data: MM-WHS (a set of healthy hearts used to train the detectives) and ASOCA (a set of 40 scans, half healthy and half sick, used to test if the system could spot the disease).

The Big Surprise: "Worry" Doesn't Always Mean "Wrong"

Here is the most fascinating part of the story. The researchers asked a tricky question: Does the "worry map" that best shows where the detectives made a mistake also help them find the disease?

They measured how well different types of "worry" matched up with actual errors. They found that Total Entropy (a fancy math way of saying "total confusion") was the best at spotting where the detectives drew the lines in the wrong place. It had a correlation score of 0.521 with the errors.

However, when they tried to use that "Total Confusion" map to help the computer diagnose the disease, it didn't work as well as they hoped.

Instead, a different kind of "worry" called KL Divergence (which measures how much the detectives disagree with each other specifically) turned out to be the secret weapon for finding the disease.

The Result: When the system used the "KL Divergence" map, it boosted the disease detection score (AUROC) to 92.92%.
The Baseline: The old, single-detective system only reached 91.25%.

Why Did This Happen?

The paper suggests a clever reason for this mix-up. The "Total Confusion" (Entropy) is high whenever the heart edges are just naturally blurry—which happens in both healthy and sick hearts. So, it's great at saying, "Hey, this edge is hard to draw!" but not so great at saying, "This heart is sick!"

On the other hand, the "Disagreement" (KL Divergence) seems to spike specifically when the heart is sick. The researchers suspect that sick heart structures might sit right on the edge of what the computer thinks is "normal," causing the five detectives to argue more about them than they do for healthy hearts. So, even though this type of worry wasn't the best at spotting drawing errors, it was the best at spotting the disease.

The "One Number" vs. "Many Numbers" Rule

The team also tried a simple trick: instead of giving the disease detector a whole map of where the heart parts were worried, they just gave it one single number representing the total worry for the whole image.

The Verdict: That didn't work well.
The Winner: Giving the detector a separate worry score for each of the 7 heart structures (like the Left Ventricle, Aorta, etc.) was the key. This "per-structure" approach preserved the location of the trouble, allowing the system to say, "The Left Ventricle is confused, and that's a sign of trouble," rather than just "The whole image is a bit fuzzy."

The Takeaway

This paper doesn't claim to have solved heart disease forever. It suggests that in medical AI, uncertainty is a feature, not a bug, but you have to be careful how you use it.

The main lesson is that the "worry" that tells you a computer is making a mistake isn't always the same "worry" that tells you a patient is sick. By using a team of five AI detectives and listening to their specific disagreements about each part of the heart, the new GRC-ProbNet system improved disease detection from 91.25% to 92.92%. It's a small but meaningful step, showing that admitting "I'm not 100% sure" can actually make a computer smarter at saving lives.

Technical Summary: GRC-ProbNet – Uncertainty-aware Feature Extraction for Cardiovascular Disease Classification

Problem Statement
Cardiovascular disease (CVD) remains a leading cause of global mortality, necessitating efficient diagnostic tools. While deep learning has advanced the automated detection of CVD from cardiac computed tomography (CT) scans, existing pipelines often rely on deterministic segmentation masks. These deterministic outputs fail to capture the inherent ambiguity and inter-rater variation present in cardiac anatomy. Consequently, standard pipelines lack the ability to propagate meaningful uncertainty to downstream classification tasks, limiting their efficacy in representing the confidence of intermediate anatomical interpretations.

Methodology
The authors propose GRC-ProbNet, an extension of the existing hybrid GRC-Net pipeline. Unlike its predecessor, which relies on a single deterministic segmentation, GRC-ProbNet integrates an uncertainty-aware framework into a modular, four-stage pipeline:

Ensemble Segmentation: Instead of a single model, the system employs a deep ensemble of five independently fine-tuned Anatomix models (differing only by random seed) to generate multiple plausible segmentation masks for a given input CT.
Registration and Atlas Construction: These multiple masks are processed through a frozen Atlas Image-and-Spatial Transformer Network (Atlas-ISTN). This stage registers the segmentations to a population-derived healthy atlas, generating deformation fields and warped probability maps.
Uncertainty Quantification: The system calculates voxel-level uncertainty maps based on the disagreement among the ensemble members. It decomposes predictive uncertainty into three specific representations:
- Total Predictive Entropy ( $H$ ): Captures both aleatoric (data noise) and epistemic (model knowledge) uncertainty.
- Epistemic Uncertainty (Jensen-Shannon Divergence): Derived from the Kullback–Leibler (KL) divergence between individual seed predictions and the consensus mean, measuring model disagreement.
- Aleatoric Uncertainty: Calculated by subtracting the epistemic component from the total entropy, isolating inherent data ambiguity.
Feature Extraction and Classification: The pipeline extracts three types of features for each cardiac structure:
- Radiomic Features: 107 features per structure (intensity statistics, shape descriptors, texture) derived from CT intensities within the segmentation masks.
- Geometric Features: 3 features per structure derived from deformation fields relative to the healthy atlas.
- Uncertainty Features: The uncertainty maps are aggregated into image-level features using two strategies: Global (averaging across all voxels) or Per-Structure (averaging within specific anatomical regions).
  These features are concatenated and fed into a Multilayer Perceptron (MLP) for final CVD classification.

Key Contributions

Framework Introduction: The paper introduces GRC-ProbNet, a hybrid pipeline that propagates explicit uncertainty representations (entropy, KL divergence, and aleatoric maps) through a radiomic and geometric feature extraction stage.
Systematic Evaluation: The authors systematically compare different uncertainty types (total, epistemic, aleatoric) and encoding strategies (global vs. per-structure) against a deterministic baseline.
Novel Finding on Uncertainty Utility: The study demonstrates that the uncertainty metric most correlated with segmentation error is not necessarily the one that yields the best downstream classification performance.

Results
Experiments were conducted on the MM-WHS dataset for segmentation/uncertainty evaluation and the ASOCA dataset for downstream CVD classification.

Uncertainty vs. Segmentation Error: On the MM-WHS dataset, Total Entropy showed the strongest correlation with segmentation error (Normalized Cross-Correlation of 0.521), followed closely by Aleatoric uncertainty. KL Divergence (Epistemic) showed the weakest correlation (0.408).
Classification Performance: On the ASOCA dataset, the uncertainty-augmented models generally outperformed the deterministic baseline (GRC-Net, AUROC 91.25%).
- Encoding Strategy: Per-structure uncertainty encodings consistently outperformed global scalar encodings.
- Best Performance: The KL Divergence with Per-Structure encoding (K.S.) achieved the highest AUROC of 92.92%, slightly outperforming the Entropy-based variants (E.S. at 92.71%).
Key Insight: While Entropy best predicted where the segmentation was wrong, KL Divergence provided the strongest signal for disease classification. The authors suggest that Entropy is dominated by boundary ambiguity common to both healthy and diseased cases, whereas KL divergence captures model disagreement that may be more specific to pathological anatomies lying near decision boundaries.

Significance and Claims
The paper claims that GRC-ProbNet substantively improves CVD classification performance (AUROC 92.92% vs. 91.25%) by leveraging uncertainty features. The primary significance lies in the finding that how uncertainty is encoded is as critical as including it. Specifically:

Structure-specific encoding preserves necessary spatial information that global summaries lose.
Uncertainty metrics serve different purposes: A metric that best reflects segmentation quality (error localization) is not automatically the best predictor for downstream clinical tasks.
Pipeline Design: This approach is particularly relevant for hybrid medical imaging pipelines where explicit feature extraction occurs between deep learning and non-deep learning stages, allowing for the explicit representation of uncertainty before final classification rather than absorbing it into an end-to-end model.

The authors conclude that future work should involve larger cohorts, the integration of multiple expert annotators to model inter-observer variability, and further investigation into uncertainty propagation through the registration stage.