💻 computer science

Connected vs. Isolated Pore Classification via Learned Refinement of Connected-Component Analysis

This study proposes a machine learning refinement pipeline that leverages geometric features and classifiers like XGBoost to correct false connectivity artifacts in CT-based pore analysis, successfully distinguishing isolated from interconnected pores where classical 26-connectivity methods fail under noise.

Original authors: Aditya Bansal, Ankit Singh Bahaduriya, Mayank Goswami

Published 2026-11-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Aditya Bansal, Ankit Singh Bahaduriya, Mayank Goswami

Original paper licensed under CC BY 4.0 (https://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

Inside the solid metal of a jet engine or the intricate castings of a car, tiny hollow spaces often exist. These are pores, microscopic voids that form during the manufacturing process. For engineers, the difference between a safe part and a dangerous one often comes down to whether these pores are isolated bubbles or if they have linked together to form a continuous, winding network. If the pores are isolated, the material remains strong and sealed. If they connect, they create hidden channels where stress can concentrate or fluids can leak, leading to sudden failure. To see inside these materials without breaking them, scientists use a powerful imaging technique called computed tomography, which builds a detailed three-dimensional map of the object's interior. However, this map is rarely perfect; the images are often grainy and fuzzy, filled with random static that looks like dust but is actually just electronic noise.

This noise creates a deceptive problem for the computers trying to analyze the images. When a standard computer program looks at a fuzzy image, it struggles to tell the difference between a real, solid wall of metal and a tiny speck of noise that happens to touch two separate pores. The computer might see a single noisy pixel bridging a gap and conclude that two distinct bubbles are actually one giant, connected web. This error is catastrophic because it makes a safe part look dangerous, or worse, it might miss a dangerous connection entirely. For decades, the industry has relied on a straightforward method to count these pores, but this method breaks down completely when the image quality is not perfect, leading to unreliable assessments of safety and durability.

Researchers at the Indian Institute of Technology Roorkee have developed a new approach to solve this problem, moving away from rigid rules and toward a system that learns to see like a human expert. Instead of trying to clean up the noisy image first, which can accidentally blur away the very thin connections that matter, they let the computer make a rough guess and then teach it how to correct its own mistakes. They started by generating thousands of perfect, computer-made models of porous materials. In these models, they knew exactly which pores were connected and which were separate. They then added realistic noise to these models to mimic the grainy quality of real industrial scans. This allowed them to create a training set where the answer was always known, teaching the computer to recognize the difference between a true physical bridge and a fake connection caused by static.

The process begins with a standard algorithm that breaks the noisy image into many small, separate pieces, often splitting single pores into several fragments just to be safe. This creates a messy map of potential connections. The researchers then built a system to examine every possible link between these pieces. They measured specific physical traits for each potential link, such as the width of the narrowest point between two bubbles and the brightness of the material along the path connecting them. They trained three different types of learning systems to use these measurements. One system used a set of logical rules, another used a network that understands how shapes relate to their neighbors, and a third looked directly at the raw 3D blocks of pixels. All three systems were tasked with a simple question: is this connection real, or is it just an artifact of the imaging process?

The results showed that the new learning-based systems were remarkably good at spotting the truth. When tested on the data they had learned from, these systems correctly identified whether a connection was real or fake more than ninety-four percent of the time. They successfully severed the false bridges created by noise, isolating the pores just as they exist in reality. However, when the researchers tried to measure the overall success using a standard scoring method that looks at the entire 3D shape at once, the new systems appeared to score lower than the old, flawed method. This happened because the initial step of breaking the image into small pieces sometimes split a single real pore into multiple fragments. The scoring system penalized the new methods for these internal splits, even though the new methods had correctly identified that the pores were not connected to each other. Despite this scoring quirk, a visual inspection of the 3D models revealed that the old method had created massive, unphysical webs of connected pores, while the new methods restored the true, isolated structure of the material.

The researchers found that the learning systems did not just memorize patterns; they learned the actual physics of the problem. By analyzing which measurements the computer relied on most, they confirmed that the systems focused on the width of the necks between pores and the intensity of the material, ignoring the random noise. This suggests the method is robust and based on real geometry rather than accidental artifacts. The team acknowledges that while their training data was perfect and synthetic, applying this to real-world industrial parts will require further work to handle the complex and varied types of noise found in actual factories. They also warn that in safety-critical fields like aerospace, the system must be tuned to be extremely cautious, ensuring that a dangerous connection is never mistaken for a safe, isolated bubble. Ultimately, this work demonstrates that by combining a rough initial guess with intelligent learning, it is possible to see through the noise and understand the true structure of the materials that hold our world together.

Technical Summary: Connected vs. Isolated Pore Classification via Learned Refinement of Connected-Component Analysis

Problem Statement

The structural integrity and fatigue resistance of porous materials, such as pressure-tight castings, depend fundamentally on distinguishing between isolated pores and interconnected percolating networks. The industry standard for this evaluation involves applying classical 26-connectivity connected-component (CC) labeling to thresholded volumetric data. However, this deterministic approach is critically fragile when applied to realistic Computed Tomography (CT) scans subject to Gaussian noise.

In low signal-to-noise ratio environments, noise creates "artifact webs" where a single noisy voxel can bridge two physically separate pores, causing the algorithm to falsely merge them into a single connected cluster. Conversely, missing voxels can split truly connected chains. This extreme sensitivity renders raw CC labeling ineffective for accurate topological assessment in noisy conditions.

Methodology

The study proposes a two-stage machine learning pipeline designed to refine classical thresholding errors by treating connectivity as a learned edge-classification problem.

1. Data Generation

To overcome the lack of mathematically unambiguous ground-truth connectivity in real sub-voxel CT scans, the authors generated a dataset of 250 synthetic phantom volumes ( $128^3$ voxels). These volumes contain controlled mixtures of isolated pores and deliberately connected clusters formed by intersecting spheres or thin-neck bridges. Gaussian noise was injected to simulate realistic CT degradation, creating the specific "false merge" scenarios required to train the models. The dataset was split into 70% training, 15% validation, and 15% testing, ensuring zero spatial leakage of edges.

2. Pipeline Workflow

Proposal Graph Generation: Noisy volumes undergo Otsu thresholding followed by morphological opening. A Euclidean distance transform is applied, followed by Watershed over-segmentation. This step systematically partitions noise-blurred blobs into discrete pore regions (nodes), generating a candidate adjacency graph.
Feature Engineering: For every pair of adjacent nodes proposed by the Watershed algorithm, the system extracts handcrafted geometric features:
- Euclidean distance between centroids.
- Estimated minimum neck radius (derived from the distance transform).
- Mean and 90th-percentile intensity along the 3D line path connecting centroids.
Model Training: Three distinct architectures were evaluated to classify candidate edges (connected vs. isolated):
- XGBoost: A tree-based model operating on the handcrafted tabular edge features.
- GraphSAGE: A 2-layer Graph Convolutional Network that learns node embeddings by exchanging information with neighboring pores, concatenating these with raw edge features for prediction.
- 3D-CNN: A fully convolutional network trained directly on $16 \times 16 \times 16$ voxel patches centered at candidate neck midpoints, bypassing handcrafted features to learn spatial filters directly.

Results and Evaluation

The evaluation reveals a significant contradiction between quantitative metrics and qualitative topological reality.

Edge-Level Performance: The machine learning models demonstrated robust edge-level classification. The 3D-CNN achieved the highest Validation F1 score (0.9736), followed by GraphSAGE (0.9457) and XGBoost (0.9414). These models successfully severed noise-induced bridge artifacts that the classical baseline could not distinguish.
Global Topological Performance (ARI): Surprisingly, the classical CC baseline achieved a higher Adjusted Rand Index (ARI) of 0.7681 compared to the ML-refined models (~0.64).
Visual Validation: Despite the lower ARI, visual inspection via PyVista confirmed that the classical baseline failed catastrophically, merging distinct pores into massive, unphysical webs. In contrast, the ML models successfully isolated clusters to match the physical ground truth.
Feature Importance: Permutation feature importance analysis on the XGBoost model confirmed that predictions were anchored on true physical geometry (specifically mean path intensity and minimum neck radius) rather than random noise distributions.
Training Dynamics: Both the GraphSAGE and 3D-CNN models exhibited smooth loss convergence and stable validation scores, indicating no overfitting to the synthetic noise.

Key Contributions and Significance

The paper claims several key contributions and insights:

Metric Limitation Identification: The study highlights a critical flaw in using global metrics like ARI for this specific problem. The ML models' lower ARI is attributed to "node inflation" caused by the Watershed algorithm, which shatters single physical pores into multiple sub-nodes. Missing even one internal edge in this inflated graph penalizes the global ARI, despite the model correctly identifying the external noise bridges.
Robust Topological Correction: The primary significance of the work is demonstrating that learned refinement pipelines can restore the physical structure of noisy porous media where classical 26-connectivity fails. The models successfully distinguish true physical connections from noise artifacts based on geometric and intensity features.
Physics-Aware Learning: The analysis confirms that the models learn physical laws (geometry and intensity gradients) rather than memorizing noise patterns, as evidenced by the feature importance rankings.

Limitations and Future Work

The authors explicitly note several limitations and constraints:

Domain Gap: The models are trained on synthetic data with additive Gaussian noise and idealized sphere geometries. They do not currently account for complex physical CT artifacts like beam hardening, scatter, or irregular manufacturing defects.
Safety Risks: In safety-critical applications (e.g., aerospace), a false-positive classification (labeling a leak-causing path as isolated) poses severe risks. The authors suggest that future NDE pipelines must incorporate safety-weighted loss functions to aggressively penalize such errors.
Preprocessing Trade-offs: The study argues against simple pre-filtering (e.g., Gaussian or Median filters) because such methods blur thin physical necks, destroying true connections that the learned refinement approach aims to preserve.
Real-World Application: The authors state the method will not work on real scans without domain adaptation techniques, such as CycleGAN intensity matching or real noise-profile injection.

The paper concludes that while classical methods are susceptible to noise, machine learning refinement offers a viable mechanism for topological correction, provided that evaluation metrics are adjusted to account for the specific artifacts introduced by the segmentation proposal stage.