Abstract
Fibrosis, the pathological excess of fibroblast activity, is a significant health issue that hinders the function of many organs, in some cases fatally. The severity of fibrosis-derived conditions depends on both the positioning of the affliction and the microscopic patterning of fibroblast-deposited matrix proteins. Variability in an individual’s manifestation of fibrosis is a key factor in explaining differences in symptoms and prognosis, yet this variability remains poorly understood due to a lack of experimental control. In this work, we present a computational methodology for generating patterns of fibrosis microstructure, demonstrated using histological images of four types of cardiac fibrosis. Our generator and automated tuning method prove flexible enough to capture distinct patterns, allowing for rapid generation of new realizations for high-throughput computational studies. We also demonstrate via simulation the importance of micro-scale variability, showing significant differences in electrophysiological impact even within a single class of fibrosis.
Introduction
Fibrosis is characterized by the excessive accumulation of collagen "scar" tissue, particularly in the heart. It manifests in complex microscopic patterns—interstitial, compact, diffuse, or patchy—that dictate how electrical impulses propagate. Unfortunately, current non-invasive imaging like MRI lacks the resolution to capture these patterns in living patients. Researchers typically rely on limited ex vivo histological slides, which restricts high-throughput computational modeling. There is a critical need for a method to generate infinite, realistic variations of fibrotic tissue to understand how structural variability leads to different clinical outcomes.
Methodology: Procedural Generation and Tuning
The proposed framework combines procedural texture generation with statistical optimization to create physiologically realistic patterns.
The Perlin Noise Generator
Instead of manual drawing, the system uses a weighted combination of three Perlin noise fields to simulate biological "noise":
- Base Field ($N_b$): Creates the primary fibrotic obstacles.
- Density Variation Field ($N_d$): Controls large-scale shifts in fibrosis density.
- Fibre Selection Field ($F$): Uses a modulated sinusoid to create strands following heart muscle fiber orientation.
Feature Extraction and Automated Tuning
To ensure realism, generated patterns are compared to real medical images using Fast Fourier Transforms (FFT). By fitting ellipses to the power spectrum, the researchers mathematically describe the size, orientation, and anisotropy of fibrotic clumps. Parameters are automatically tuned using Sequential Monte Carlo Approximate Bayesian Computation (SMC-ABC), finding the optimal settings to produce patterns statistically indistinguishable from real histological samples.
Key Results
High-Fidelity Pattern Recreation
The generator successfully recreated the four primary archetypes of cardiac fibrosis: Interstitial (strands between fibers), Compact (large clumps), Diffuse (scattered speckles), and Patchy (large clumps following fiber directions).
Notably, for compact and diffuse types, the authors found that omitting the "fiber selection" field ($F$) improved realism, suggesting these patterns are less constrained by underlying anatomy. The framework captures the statistical essence of a disease class rather than just replicating a specific slide, allowing for thousands of unique, physiologically plausible realizations.
Electrophysiological Sensitivity
Simulations of chronic atrial fibrillation revealed that even within a single class (e.g., "Diffuse"), different noise realizations led to vastly different outcomes. Subtle variations in arrangement—invisible to standard imaging—could trigger conduction blocks or re-entry, the precursors to cardiac arrest. This highlights that density alone is an insufficient metric for assessing clinical risk.
Discussion and Future Work
The success of Perlin noise suggests that collagen deposition follows procedural rules similar to natural textures like clouds or marble. While this study focused on 2D slices, extending the generator to 3D volumes is a logical next step. Future iterations may also incorporate curvilinear coordinate systems to map these textures onto the complex, twisting geometries of the real heart. Although pattern generation is rapid, high-resolution electrophysiological simulations remain computationally intensive, requiring further optimization for large-scale clinical application.
Summary updated