Technical Summary: PINN-Flow

Problem Statement

The rupture risk stratification of Abdominal Aortic Aneurysms (AAA) relies on two biomechanically coupled quantities: Wall Shear Stress (WSS) derived from intraluminal blood flow and Peak Wall Stress (PWS) derived from the mechanical stress of the aortic tissue. Current deep-learning approaches address these separately using purely data-driven optical flow methods:

• Pipeline 1: Uses U-Net + RAFT on Cine-MRI to estimate fluid velocity.
• Pipeline 2: Uses U-Net + TV-L1 on B-mode ultrasound to estimate wall strain.

These methods suffer from a critical structural weakness: their loss functions enforce only photometric consistency and generic spatial smoothness, lacking any constraint based on the underlying physics of blood or tissue. Consequently, they are highly vulnerable to modality-specific artifacts (e.g., speckle decorrelation in ultrasound, phase artifacts in MRI), particularly in the near-wall boundary layer. This is the region of lowest image contrast where WSS and PWS are computed, leading to locally erratic estimates that compromise rupture risk assessment.

Methodology

The authors propose PINN-Flow, a unified Physics-Informed Neural Network (PINN) framework that jointly estimates the fluid velocity field and the wall displacement field using a shared-encoder, dual-decoder architecture.

1. Network Architecture

• Structure: A shared multi-scale encoder feeds two independent decoders: a fluid branch predicting dense velocity $(u, v)$ and a wall branch predicting wall displacement $(u_w, v_w)$.
• Feature Coupling: The shared encoder allows near-wall boundary features to inform both branches, providing soft feature-level coupling that complements the physics constraints enforced at the output level.
• Learnable Parameters: Kinematic viscosity ($\nu$) and Young's modulus ($E$) are parameterized in log-space and learned jointly with network weights.

2. Mathematical Formulation

The network is trained by minimizing a composite loss function:
$$L_{total} = \lambda_1 L_{data} + \lambda_2 L_{smooth} + \lambda_3 L_{NS} + \lambda_4 L_{Elast}$$

• Data Term ($L_{data}$): Enforces brightness constancy via a robust warping loss (Huber loss) between consecutive frames.
• Smoothness Term ($L_{smooth}$): Enforces spatial regularity on the gradients of the velocity and displacement fields.
• Navier-Stokes Term ($L_{NS}$): Applies a 2D incompressible Navier-Stokes residual to the fluid branch. It penalizes violations of mass conservation ($\nabla \cdot V = 0$) and momentum conservation. The pressure gradient term is absorbed into a reduced residual due to the lack of direct pressure measurements.
• Elasticity Term ($L_{Elast}$): Applies a linear-elastic static-equilibrium residual to the wall branch. It models the wall as a linear-elastic, small-strain, plane-stress material, penalizing violations of the equilibrium equation ($\nabla \cdot \sigma = 0$).

3. Implementation: Coordinate-Continuous Differentiation

A key technical innovation is the method for computing physics residuals. Instead of using finite differences on a discrete pixel lattice, the network output is bilinearly re-sampled onto a continuous coordinate grid via a differentiable grid sample operation.

• This transforms the discrete CNN output into a genuinely continuous, twice-differentiable function of coordinates $(x, y)$.
• Spatial and temporal derivatives required for the PDE residuals are computed via automatic differentiation over this continuous field, ensuring exact differentiability and backpropagation into network weights.

Key Contributions

1. Unified Framework: A single architecture that jointly estimates intraluminal fluid velocity and aortic wall displacement, applicable to both Cine-MRI and B-mode ultrasound.
2. Physics-Informed Loss: Implementation of a composite loss combining 2D incompressible Navier-Stokes residuals (fluid) and linear-elastic static-equilibrium residuals (wall), computed via automatic differentiation.
3. Robustness Validation: Demonstration through controlled noise-injection experiments that physics constraints lower converged loss, reduce training instability, and degrade more gracefully under noise compared to data-driven baselines.
4. Clinical Implication Analysis: Discussion of how this framework improves the estimation of WSS and PWS, the primary descriptors for AAA rupture risk.

Experimental Results

Experiments were conducted using simulated data with controlled noise levels (multiplicative speckle for ultrasound-like conditions and additive Rician noise for MRI-like conditions), as no ground-truth annotated AAA dataset with synchronized references was available.

• Convergence: PINN-Flow achieved a final training loss of 0.1053 and validation loss of 0.1112, significantly lower than the data-driven baseline (0.2369 and 0.2731, respectively).
• Stability: The standard deviation of the loss in the final 20 epochs was reduced by approximately 2.7× (0.0066 vs. 0.0180), indicating reduced oscillation and fewer local minima.
• Noise Robustness: Under increasing noise levels, PINN-Flow demonstrated superior resilience. The Normalized Root Mean Square Error (NRMSE) degradation from clean to the highest noise level was +172.6% for PINN-Flow, compared to +184.6% for RAFT and +214.7% for TV-L1.
• Performance at High Noise: At the highest simulated noise level, PINN-Flow outperformed TV-L1 by ~29% and RAFT by ~21% in mean NRMSE.

Significance and Claims

The paper claims that by constraining tracking to physically admissible velocity and displacement fields, PINN-Flow unifies previously separate AAA pipelines into a single framework. Its primary significance lies in its ability to produce more stable and artifact-resilient estimates of WSS and PWS, specifically in the low-contrast, high-noise near-wall regions where purely photometric models typically fail.

The authors emphasize that the physics prior acts as an anchor in regions where photometric information is corrupted (e.g., near the wall boundary), ensuring that the resulting fields remain divergence-free and in equilibrium. This is critical for clinical applications where the most noisy regions are also the most critical for calculating rupture risk.

Limitations

The authors explicitly acknowledge several limitations:

1. 2D Assumption: The formulation is planar; out-of-plane flow and motion are not penalized, potentially biasing estimates in regions of strong 3D curvature.
2. Pressure Term: The momentum residual omits an explicit pressure-gradient term due to the lack of direct pressure measurements.
3. Computational Cost: Second-order automatic differentiation roughly doubles to triples the backward-pass cost compared to purely data-driven networks.
4. Simulated Validation: Results are based on simulated noise models rather than a paired, ground-truth clinical dataset; thus, they establish feasibility and expected behavior rather than clinical-grade evidence.
5. Uniform Parameters: The model assumes spatially uniform viscosity and stiffness, whereas biological reality involves spatial variations.