💻 bioinformatics

FilAdh: degradation-aware analysis of focal adhesions and actin stress fibers

FilAdh is a unified, degradation-aware computational framework that simultaneously segments focal adhesions and reconstructs actin stress fibers to enable accurate morphometric analysis and spatial classification, outperforming existing methods in benchmark tests while explicitly characterizing how specific imaging imperfections affect structural measurements.

Original authors: Rey-Paniagua, G., Garcia-Lopez-de-Haro, C., Hernandez-Martinez, M., Lachowski, D., Munoz-Barrutia, A.

Published 2026-10-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Rey-Paniagua, G., Garcia-Lopez-de-Haro, C., Hernandez-Martinez, M., Lachowski, D., Munoz-Barrutia, A.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Inside every living cell, a microscopic construction crew works tirelessly to maintain shape, move, and sense the world outside. This crew consists of two main teams that are physically linked but often studied in isolation. One team builds strong, sticky pads called focal adhesions that anchor the cell to its surroundings, acting like molecular grappling hooks. The other team weaves long, thin cables known as stress fibers, which are bundles of protein filaments that generate the tension needed to pull the cell forward or hold it firm. Because these two systems are mechanically connected, a change in one inevitably affects the other, yet scientists have traditionally analyzed them separately, using different tools for the sticky pads and the cables. This separation makes it difficult to see the full picture of how a cell responds to its environment, especially when the images used to study them are imperfect due to the limitations of microscopes.

A new framework called FilAdh changes this approach by treating the sticky pads and the cables as a single, integrated system. Developed by researchers at Universidad Carlos III de Madrid and the Gregorio Marañón Health Research Institute, this tool is designed to find, count, and measure both structures simultaneously within a single image. The researchers tested their system on images of three different types of human cells, including cancer cells and lung fibroblasts, which were captured using various high-powered microscopes. They compared FilAdh against several existing methods, ranging from older, rule-based software to modern artificial intelligence models. The results showed that FilAdh was exceptionally good at identifying the sticky pads, finding more of them and matching their shapes more accurately than any other method tested. For the cables, it performed just as well as the best existing artificial intelligence tools, successfully tracing the long, thin lines even when they crossed over one another.

What makes this work particularly valuable is that the researchers did not just stop at finding the structures; they also figured out how to classify the cables based on where they attach. In the two-dimensional view of a microscope, these cables can be categorized as dorsal, ventral, or peripheral depending on their position within the cell and which sticky pads they touch. Using expert-drawn maps as a guide, FilAdh correctly assigned these categories to the cables with an accuracy between 81.5% and 87.3% across the different cell types. This ability to link the cable's path to its anchor points provides a much clearer view of how the cell is organized and how it might be moving or changing shape.

However, the researchers also knew that real-world microscope images are rarely perfect. They can be grainy, blurry, or out of focus, and these flaws can trick a computer into seeing things that aren't there or missing things that are. To understand how these imperfections affect the measurements, the team deliberately degraded their test images with simulated noise, blur, and focus errors. They discovered that the two structures react differently to these problems. The sticky pads, which are small and compact, became very hard to measure accurately when the image was blurry or noisy; the software tended to either miss them entirely or merge them together into larger, incorrect blobs. The cables, being long and thin, were more resilient to these errors. While the software might miss a small segment of a cable or add a tiny false piece, the overall network remained largely intact and recognizable.

This distinction is crucial for anyone relying on these measurements. The study demonstrates that a method that looks good at simply drawing a line around a structure does not necessarily mean the numbers derived from that drawing—such as the total number of structures or their average size—are reliable. By explicitly showing how different types of image flaws distort the final counts and sizes, FilAdh offers a more honest and transparent way to study cells. It allows scientists to see not just the structures themselves, but also the specific ways in which their measurements might be biased by the quality of the image. Ultimately, this work provides a unified, robust way to observe the mechanical heart of a cell, ensuring that the story told by the data is as accurate as the microscope allows.

Technical Summary: FilAdh – Degradation-Aware Analysis of Focal Adhesions and Actin Stress Fibers

Problem Statement
Focal adhesions (FAs) and actin stress fibers (SFs) constitute a mechanically coupled system essential for cellular force transmission, migration, and mechanosensing. However, they are typically analyzed using separate workflows tailored to their distinct geometries: FAs appear as compact, often clustered patches, while SFs form thin, extended, and overlapping filamentous networks. Existing methods often analyze these structures in isolation, failing to capture their structural interdependencies. Furthermore, reliable quantitative analysis is hindered by biological complexity (e.g., nascent vs. mature adhesions, overlapping fibers) and optical limitations (e.g., photon noise, defocus, point-spread function blur). Current benchmarks often lack a unified framework that addresses both structures simultaneously while explicitly characterizing how image degradation affects specific morphometric measurements.

Methodology
The authors present FilAdh, a unified deep learning framework designed to perform FA instance segmentation, SF instance reconstruction, and FA-anchorage-based SF classification within a single pipeline. The framework consists of three specialized branches:

FA Instance Segmentation:
- Enhancement: Raw FA channel images are processed by a NAFNet-GAN generator trained to map raw inputs to an enhanced grayscale representation derived from a Horzum-based workflow (pre-threshold). This step improves signal-to-noise ratio and contrast.
- Segmentation: A Graph-Bottleneck Attention U-Net processes the enhanced image to predict an FA mask.
- Instance Separation: Touching objects are separated using marker-controlled watershed segmentation on the negative Euclidean distance transform.
SF Instance Reconstruction:
- Prediction: A Graph-Bottleneck Attention U-Net predicts both a binary filament mask and an undirected local tangent map (encoded as $\cos 2\theta, \sin 2\theta$ ).
- Reconstruction: Orientation-guided refinement separates paths with incompatible local orientations and reconnects geometrically supported segments. A minimum length threshold (1.5 µm) removes short pseudofilaments.
SF Classification:
- A classifier network takes the actin/SF channel image, the FA mask, and the SF mask as inputs.
- It assigns each filament to one of three structural classes based on operational 2D criteria: Dorsal (one FA contact, perpendicular to membrane), Ventral (two FA contacts, central, tangent to membrane), or Peripheral (two FA contacts, at cell periphery).

Key Contributions

Unified Framework: FilAdh is the first method to jointly quantify FAs and SFs, tailoring segmentation and reconstruction strategies to the specific geometry of each structure (compact patches vs. thin networks).
Degradation-Aware Analysis: The study introduces a rigorous evaluation protocol using matched synthetic degradations (Poisson-Gaussian noise, Gaussian defocus, and PSF blur) to reveal structure-specific error modes.
Comprehensive Benchmarking: The method is evaluated against a diverse set of classical (FAAS, Horzum, FilamentSensor 2.0) and deep learning (Star-convex, U-Net++, SynSeg, FAST) baselines across three cell lines (HeLa, SUIT-2, MRC-5) and multiple microscopy modalities (spinning-disk confocal, TIRF, laser-scanning confocal).
Explicit Bias Characterization: The work explicitly links segmentation errors to downstream morphometric biases (count, area, length, width), demonstrating that high segmentation scores do not guarantee unbiased biological measurements.

Results

FA Performance: FilAdh achieved the highest Dice (0.803), Intersection over Union (0.676), and Recall (0.912) on the held-out test set, significantly outperforming all benchmarked methods (paired t-tests with Holm correction). It demonstrated superior recall compared to deep learning comparators while maintaining high precision.
SF Performance: FilAdh achieved the highest clDice (0.684) and skeleton precision (0.702). While its performance was statistically comparable to other deep learning methods (SynSeg, FAST), it significantly outperformed classical methods.
Classification Accuracy: Using expert-curated masks to isolate classification from reconstruction errors, FilAdh classified SFs into dorsal, ventral, and peripheral categories with accuracies ranging from 81.5% (MRC-5) to 87.3% (HeLa).
Degradation Effects:
- FAs: Performance was highly sensitive to degradation. Gaussian defocus and PSF blur caused boundary broadening and merging, drastically reducing precision and Dice scores while inflating area and count. Poisson-Gaussian noise primarily reduced recall by obscuring weak adhesions.
- SFs: Topological metrics (clDice) remained relatively stable across degradations. However, morphometric biases persisted; for instance, Gaussian defocus increased apparent filament length and reduced counts due to fragmentation or merging, despite stable topology scores.
Efficiency: The FilAdh Graph-Bottleneck Attention U-Net (2.44M parameters) demonstrated significantly faster inference times on both GPU and CPU compared to larger benchmark networks (e.g., U-Net++ with 26M parameters).

Significance and Claims
The authors claim that FilAdh offers a practical, interpretable approach to studying the FA-actin system by preserving the distinct geometric requirements of each structure within a single workflow. The study emphasizes that "good segmentation does not necessarily mean that the measurements derived from it are equally reliable." By making degradation-dependent measurement biases explicit, FilAdh allows researchers to understand how specific imaging imperfections (noise vs. blur) differentially distort biological readouts for compact versus extended structures. The paper concludes that the design principle of matching analysis to structure geometry and evaluating methods based on error propagation into final biological measurements is a transferable strategy for other bioimage analysis problems. The authors maintain a modest stance, noting that their benchmarks are conditional on expert-curated references and that generalization to unseen acquisition systems or staining protocols requires further validation.

More like this