Evidential Perfusion Physics-Informed Neural Networks with Residual Uncertainty Quantification

This paper introduces Evidential Perfusion Physics-Informed Neural Networks (EPPINN), a novel framework that integrates evidential deep learning with physics-informed modeling to quantify both aleatoric and epistemic uncertainties in CT perfusion imaging, thereby achieving superior accuracy and reliability in acute ischemic stroke assessment compared to existing deterministic methods.

The Gist

Imagine you are a detective trying to solve a crime in a city that is currently under a heavy fog. The crime is an ischemic stroke, where blood flow to a part of the brain has been blocked. To save the patient, you need to know exactly where the damage is (the "infarct core") and how much tissue can still be saved (the "penumbra").

To do this, doctors use a special camera called CT Perfusion (CTP). It takes a series of photos of the brain as a dye (contrast agent) flows through it. By watching how the dye moves, they can calculate how fast blood is flowing.

However, there are two big problems with this detective work:

1. The Photos are Blurry: The camera isn't perfect. The images are often noisy, and the "photos" are taken too far apart in time (like taking a picture of a race car every 3 seconds instead of every millisecond).
2. The Math is Broken: The standard way to calculate blood flow from these blurry photos is like trying to solve a puzzle where you've lost half the pieces. It's an "ill-posed" problem, meaning there are many possible answers, and the math often gets confused by the noise, leading to wild guesses.

The Old Solutions

• The Old Math (SVD): This is like using a rigid rulebook. It's fast, but if the data is noisy, the rulebook breaks, and the results are chaotic.
• The Old AI (PINNs): Scientists tried using Artificial Intelligence (Neural Networks) that "knows" the laws of physics (like how blood should flow). This is like giving the detective a map of the city's traffic laws. It helps, but the AI is deterministic. It gives you one answer and says, "This is the truth," even if it's actually just a guess based on bad data. It doesn't tell you, "Hey, I'm not sure about this part."

The New Solution: EPPINN

The authors of this paper created a new detective tool called EPPINN (Evidential Perfusion Physics-Informed Neural Network). Think of it as a Super-Detective with a "Confidence Meter."

Here is how it works, using simple analogies:

1. The "Physics Map" (Physics-Informed)

Just like the old AI, EPPINN knows the laws of physics. It knows that blood flow follows specific rules (like water flowing through a pipe). It uses these rules as a guide to keep its answers realistic.

2. The "Confidence Meter" (Evidential Learning)

This is the magic part. Instead of just giving you a single number for blood flow, EPPINN gives you two things:

• The Estimate: "I think the blood flow here is 20 units."
• The Uncertainty: "But, because the image is blurry, I'm only 60% sure. The real number could be anywhere between 10 and 30."

It does this by treating the "mistakes" in the physics rules not as errors to be fixed, but as evidence. If the AI's guess doesn't perfectly match the physics law, it doesn't just force a fit; it asks, "How much do I trust this data point?"

• If the data is noisy, the "Confidence Meter" goes up (high uncertainty).
• If the data is clear, the meter goes down (high confidence).

This is like a weather forecaster saying, "It will rain at 3 PM," versus "It will rain at 3 PM, but there's a 50% chance it might not." The second one is much more useful for making decisions!

3. The "Stabilizer" (Anti-Collapse)

Sometimes, when the math gets too tricky, the AI gets confused and decides that "Time Delay" is zero for everyone. It's like a GPS that suddenly says, "You are already at your destination," even though you just left the house. This is called "collapse."

The authors added special "guardrails" (regularization) to the AI. These guardrails force the AI to stay within realistic human limits (e.g., "Blood can't travel instantly") and prevent it from giving up and saying "zero" when things get messy.

Why Does This Matter?

In a stroke emergency, every second counts.

• Old AI: Might say, "There is a blockage here," but it's actually just noise. The doctor might treat a healthy patient, or miss a sick one because the AI was too confident in a wrong answer.
• EPPINN: Says, "There is likely a blockage here, but the image is fuzzy, so I'm not 100% sure. Please look closer."

The Results

The researchers tested this "Super-Detective" in three ways:

1. Digital Simulations: They created fake brains with known answers. EPPINN was the most accurate, especially when the "photos" were blurry or taken too far apart.
2. Public Benchmarks: It beat all other AI and math methods at finding the damaged brain tissue.
3. Real Patients: In a group of 42 real stroke patients, EPPINN correctly identified the damaged area in 41 out of 42 cases. The next best method only got 28 right.

The Bottom Line

EPPINN is like upgrading a detective from a robot that blindly follows rules to a seasoned investigator who knows the rules, understands the limitations of their evidence, and tells you exactly how much they trust their conclusion.

This makes it a safer, more reliable tool for doctors to decide who needs immediate surgery and who needs a different kind of care, potentially saving more lives during those critical first few hours of a stroke.

Technical Summary

Here is a detailed technical summary of the paper "Evidential Perfusion Physics-Informed Neural Networks with Residual Uncertainty Quantification" (EPPINN).

1. Problem Statement

Context: Acute ischemic stroke management relies on rapid differentiation between the irreversible ischemic core and salvageable penumbra using Computed Tomography Perfusion (CTP). This requires quantifying hemodynamic parameters (Cerebral Blood Flow - CBF, Cerebral Blood Volume - CBV, Mean Transit Time - MTT).
The Challenge:

• Ill-posed Inverse Problem: CTP quantification relies on deconvolution, which is highly sensitive to noise and sparse temporal sampling.
• Limitations of Current Methods:
  • Traditional SVD: Efficient but unstable under low signal-to-noise ratios (SNR) or sparse sampling.
  • Data-Driven Deep Learning: Improves robustness but often violates tracer-kinetic physical constraints, leading to physiologically implausible results and poor generalization.
  • Existing Physics-Informed Neural Networks (PINNs): Incorporate physical constraints but remain deterministic. They provide point estimates without quantifying uncertainty, which is critical for clinical decision-making. Furthermore, Bayesian PINNs or ensemble methods (which do quantify uncertainty) are computationally too expensive for rapid, per-case clinical deployment.

2. Methodology: EPPINN

The authors propose Evidential Perfusion Physics-Informed Neural Networks (EPPINN), a framework that integrates evidential deep learning with physics-informed modeling to estimate perfusion parameters while quantifying uncertainty without Bayesian sampling.

A. Core Architecture

EPPINN utilizes three coordinate-based neural networks:

1. AIF Network ($f_{AIF}$): Models the Arterial Input Function.
2. Tissue Network ($f_{tissue}$): Models the tissue concentration curve.
3. Parameter Network ($f_{par}$): Predicts perfusion parameters (CBV, MTT, transit delay $\Delta t$) and evidential parameters.

• Encoding: Spatial coordinates are encoded using a multi-resolution hash grid.
• Physics Constraint: The model enforces the tracer-kinetic convolution equation. Specifically, it minimizes the residual $r(t, x)$ derived from the box-residue function model:
  $$\frac{\partial C}{\partial t} = CBF \cdot [C_a(t-\Delta t) - C_a(t-\Delta t - MTT)]$$

B. Evidential Formulation (Uncertainty Quantification)

Instead of predicting a single residual value, EPPINN treats the physics residual as an evidential regression target.

• Distribution: The network outputs parameters $(\tilde{\alpha}, \tilde{\beta}, \tilde{\nu})$ that define a Normal–Inverse–Gamma (NIG) distribution over the residual.
• Uncertainty Decomposition: This allows for the closed-form calculation of:
  • Aleatoric Uncertainty ($\hat{\sigma}^2_{ale}$): Noise inherent in the data.
  • Epistemic Uncertainty ($\hat{\sigma}^2_{epi}$): Uncertainty due to model ignorance or lack of data.
  • Total Uncertainty: The sum of both.
• Advantage: This provides voxel-wise uncertainty estimates in a single forward pass, avoiding the computational cost of ensembles or MCMC sampling.

C. Optimization and Stability Strategies

To ensure robust convergence in noisy clinical settings, the authors introduce several stabilization mechanisms:

1. Structured Parameterization:
   • Predicts CBV and MTT directly, deriving CBF via the Central Volume Principle ($CBF = CBV / MTT$) to enforce physiological consistency.
   • Derives Time-to-Maximum (Tmax) from $\Delta t$ and MTT.
2. Progressive Annealing: The influence of the evidential loss terms is gradually increased during training. This allows the network to first fit the data curves before strictly enforcing physics consistency, preventing optimization instability.
3. Anti-Collapse Regularization: Addresses the "mode collapse" issue where transit delay ($\Delta t$) converges to zero due to weak gradients in broad AIFs.
   • Exponential Barrier: Penalizes values approaching zero.
   • Variance Penalty: Prevents uniform collapse of MTT across the image.
   • Physiological Bounds: Soft penalties enforce realistic ranges (e.g., $\Delta t < 15s$).
4. Initialization & Pre-training:
   • AIF network is pre-trained on global arterial signals.
   • Parameter network is initialized to physiologically plausible values (e.g., CBF $\approx$ 20).

3. Key Contributions

1. Uncertainty-Aware PINN: Introduced the first framework for CTP that models physics constraint deviations probabilistically using evidential learning, yielding voxel-wise aleatoric and epistemic uncertainty without additional inference overhead.
2. Stable Per-Case Optimization: Developed a training scheme combining evidential residual learning, structured parameterization, and progressive annealing, ensuring robust convergence even with sparse temporal sampling and high noise.
3. Comprehensive Validation: Validated the method on digital phantoms, the ISLES 2018 benchmark, and a retrospective clinical cohort, demonstrating superior performance over classical and PINN baselines.

4. Experimental Results

The method was evaluated on:

• Digital Phantoms: Controlled noise (PSNR 18–27 dB) and temporal sampling ($\delta t$ 1–4s).
• ISLES 2018 Benchmark: 93 cases.
• Clinical Cohort: 42 retrospective cases.

Key Findings:

• Accuracy: EPPINN achieved the lowest Normalized Mean Absolute Error (NMAE) for CBF and CBV across all noise and sampling conditions, outperforming SVD, bcSVD, boxNLR, and other PINN variants (SPPINN, ReSPPINN). Gains were most significant under sparse sampling ($\delta t = 3-4s$) and low SNR.
• Uncertainty Calibration: The empirical coverage of uncertainty intervals (68%/95%) exceeded nominal levels, indicating the model provides conservative and reliable uncertainty estimates.
• Infarct Core Detection:
  • ISLES: EPPINN achieved the highest voxel-level sensitivity (0.323) and case-level sensitivity (0.462).
  • Clinical: EPPINN correctly identified 41 out of 42 cases (97.6% case-level sensitivity), significantly outperforming SVD (28.6%) and other baselines.
• Qualitative Results: EPPINN produced spatially coherent perfusion maps with reduced high-frequency speckle noise compared to competitors, aligning better with Diffusion-Weighted Imaging (DWI) core regions.

5. Significance

This work bridges the gap between physical accuracy and clinical reliability in stroke imaging.

• Clinical Impact: By providing not just a parameter map but also a confidence map (uncertainty), EPPINN supports time-critical treatment decisions (thrombolysis/thrombectomy) by highlighting regions where the model is uncertain, potentially preventing false negatives in small or heterogeneous infarct cores.
• Efficiency: Unlike Bayesian methods, EPPINN maintains the speed of a single optimization pass, making it feasible for real-time clinical deployment.
• Robustness: The specific regularization strategies (anti-collapse, annealing) solve common failure modes in PINNs when applied to noisy, sparse medical data.