🧬 biology

Interpretable Radiomic Analysis of Chest X-rays for Pneumonia Discrimination Using First-Order and GLCM Texture Features, Statistical Screening, and PCA

This study demonstrates that a transparent radiomics pipeline using first-order and GLCM texture features from segmented chest X-rays can statistically distinguish pneumonia from normal cases with moderate accuracy, suggesting its potential as an interpretable supportive tool rather than a standalone diagnostic solution.

Original authors: Muhammad Noman, Muhammad Aziz, Farooq Mukhtar

Published 2026-10-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Muhammad Noman, Muhammad Aziz, Farooq Mukhtar

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

Every year, millions of people around the world are diagnosed with respiratory illnesses, and for many, the first step in that journey is a chest X-ray. This simple, quick image is often the only tool available in crowded clinics or remote villages, offering a glimpse into the lungs without the high cost or radiation of a CT scan. However, reading these images is far from perfect. A doctor looking at an X-ray sees a flat, two-dimensional shadow of a complex three-dimensional body, where ribs and the heart can hide or mimic disease. What one doctor sees as a clear lung, another might see as a sign of infection, leading to disagreements that can reach half of all cases. To solve this, scientists are turning to a field called radiomics, which treats medical images not just as pictures to be looked at, but as vast collections of data to be measured. Instead of relying on a human eye to guess if a lung looks "cloudy" or "clear," radiomics uses computers to extract precise numbers that describe the brightness and texture of the image, hoping to find patterns that the human eye might miss.

In a recent study, researchers set out to build a transparent, easy-to-understand system that could distinguish between a healthy lung and one infected with pneumonia using only these measurable numbers. They started with a large collection of 2,000 chest X-rays from a public database, splitting them evenly between images labeled as normal and those labeled as having pneumonia. To make sure the computer was looking at the right parts of the image, they first used a specialized artificial intelligence tool to draw a digital outline around the lungs, effectively cutting out the ribs, the heart, and the rest of the body. After removing a few images where this digital outline was damaged or unclear, they were left with a clean set of 1,911 images to analyze.

The team then asked the computer to measure eleven specific characteristics of the lung tissue inside those outlines. Five of these measurements described the overall brightness and how the light and dark spots were distributed, while the other six focused on the texture, or how the pixels were arranged next to one another. They compared these numbers between the healthy lungs and the sick ones to see if there were any clear differences. The results revealed something surprising: the overall brightness of the lung did not matter. The average light level in a healthy lung was almost identical to that in a lung with pneumonia, meaning that simply looking at how bright an image is cannot tell a doctor if a patient is sick.

However, the texture of the lung told a very different story. The researchers found that the way the pixels were arranged held the key. Healthy lungs, filled with air and crisscrossed by blood vessels and airways, showed a complex, varied pattern with many different shades of gray mixed together. In contrast, lungs with pneumonia showed a much smoother, more uniform texture. This is because pneumonia fills the tiny air sacs with fluid and cells, creating a dense, cloudy appearance that flattens out the natural variations found in healthy tissue. Ten of the eleven measurements the team took showed statistically significant differences between the two groups, with the texture-based numbers showing the strongest ability to tell them apart.

To make sense of these ten different numbers, the researchers used a mathematical technique to combine them into a few broader categories. They found that nearly all the useful information could be captured in just four main groups, with the most important one describing the uniformity of the texture. This confirmed that the smoothness of the lung pattern was the single most powerful signal for detecting pneumonia in their data. When they fed these simplified numbers into three different types of computer programs designed to make predictions, the systems performed moderately well. The best program correctly identified the condition about 64 percent of the time and showed a moderate ability to distinguish between the two groups overall.

The study concludes that while these texture measurements carry real, meaningful information that can separate pneumonia from healthy lungs, they are not yet strong enough to replace a doctor's diagnosis on their own. The accuracy was too low to be used as a standalone tool, likely because the system was looking at the entire lung at once rather than focusing on specific spots of infection, and because it used a relatively simple set of measurements. The researchers suggest that this approach is best viewed as a helpful assistant that can offer an objective second opinion to a human reader, rather than a final judge. By focusing on the physical texture of the lung rather than just its brightness, this work provides a clear, explainable foundation for future tools that could one day help doctors make faster and more accurate decisions, especially in places where expert radiologists are scarce.

Technical Summary: Interpretable Radiomic Analysis of Chest X-rays for Pneumonia Discrimination

Problem Statement
Chest X-ray (CXR) remains the primary imaging modality for respiratory diseases due to its accessibility and low cost. However, clinical interpretation is inherently subjective, leading to significant inter- and intra-reader variability, with disagreement rates reported as high as 50%. Furthermore, the two-dimensional nature of CXRs causes anatomical structures (ribs, clavicles) to overlap with lung pathology, and early disease signs are often subtle. While deep learning offers high performance, its "black box" nature limits clinical trust. This study addresses the need for an objective, transparent, and interpretable quantitative framework to distinguish pneumonia from normal lungs using radiomics.

Methodology
The authors developed a reproducible radiomics pipeline using a balanced subset of 2,000 images from the NIH ChestX-ray14 dataset (1,000 "No Finding" and 1,000 "Pneumonia"). The workflow consisted of the following stages:

Preprocessing: Images were resized to 256 × 256 pixels and contrast-enhanced using Contrast-Limited Adaptive Histogram Equalization (CLAHE) to standardize acquisition variations.
Segmentation: A pre-trained four-level U-Net (initialized with public weights) was employed to segment lung fields. Images with corrupted masks (e.g., obscured boundaries) were excluded, resulting in a final cohort of 1,911 images (955 normal, 956 pneumonia).
Feature Extraction: Eleven hand-crafted features were extracted exclusively from the segmented lung regions:
- First-order statistics (5): Mean, standard deviation, skewness, kurtosis, and entropy.
- Gray-Level Co-occurrence Matrix (GLCM) features (6): Contrast, dissimilarity, homogeneity, energy, correlation, and angular second moment (ASM).
Statistical Screening & Dimensionality Reduction: Features were screened using independent-samples t-tests (with Bonferroni adjustment). Significant features were standardized via z-score and reduced using Principal Component Analysis (PCA) to capture latent structure.
Classification: Three baseline classifiers—Logistic Regression (LR), Support Vector Machine (SVM), and Random Forest (RF)—were evaluated using stratified 5-fold cross-validation. Notably, standardization and PCA were fitted on the full dataset prior to cross-validation (an unsupervised step acknowledged as a limitation).

Key Results

Feature Significance: Ten of the eleven features showed statistically significant differences between classes ( $p < 0.05$ ). The only non-significant feature was mean intensity ( $p = 0.313$ ), indicating that global brightness does not discriminate pneumonia.
Effect Sizes: The most discriminative features were GLCM-based texture metrics. Entropy, contrast, dissimilarity, homogeneity, energy, and ASM exhibited moderate effect sizes (Cohen's $d \approx 0.49–0.54$ ). First-order features like standard deviation and skewness showed smaller effects ( $d < 0.3$ ).
Physical Interpretation: Normal lungs exhibited higher entropy, contrast, and dissimilarity (indicating heterogeneous texture), while pneumonia lungs showed higher homogeneity, energy, and ASM (indicating uniform, "cloudy" consolidation).
PCA Findings: Four principal components retained 90.94% of the variance. The first component (PC1), explaining 62.02% of variance, was dominated by texture-uniformity features (homogeneity, energy, ASM vs. entropy, contrast), physically representing the shift from heterogeneous aerated lung to homogeneous consolidation.
Classification Performance: All models achieved moderate performance. The SVM performed best with an accuracy of 0.637 and an AUC of 0.687. Logistic Regression and Random Forest followed closely. The authors note that the small differences between models suggest the limitation lies in the feature set's information content rather than classifier complexity.

Key Contributions

Reproducible Pipeline: A transparent workflow combining CLAHE, deep-learning segmentation, and a compact set of 11 interpretable features.
Statistical Characterization: A rigorous quantification of feature discriminability on a balanced cohort, including effect sizes, demonstrating that texture, not intensity, drives the signal.
Interpretability: A PCA-based analysis that maps statistical features to physically meaningful concepts (e.g., texture uniformity), allowing clinicians to understand the basis of the model's decisions.
Baseline Benchmark: A clear performance baseline (AUC ~0.69) for a minimal, explainable feature set, establishing a reference point for future improvements.

Significance and Claims
The paper positions this work as a foundational, interpretable baseline rather than a standalone diagnostic solution. The authors claim that:

Texture is the Signal: The study confirms that spatial texture patterns (GLCM features) carry statistically significant information for pneumonia detection, whereas global brightness does not.
Role as a Support Tool: Given the moderate classification accuracy (AUC 0.687), the authors argue that such compact, interpretable feature sets are best utilized as supporting, explainable tools to assist radiologists, rather than as autonomous diagnostic systems.
Need for Expansion: The results highlight the necessity for richer feature sets (higher-order textures, localized features) and more robust validation to improve performance. The study explicitly avoids claiming that this specific pipeline is sufficient for clinical deployment without further development.

Limitations Acknowledged
The authors explicitly note limitations including potential label noise from NLP-derived ground truth, the use of global features rather than lesion-specific ones, reliance on a pre-trained U-Net without quantitative validation on this specific cohort, and the lack of external validation or patient-level cross-validation splits.

Technical Summary: Interpretable Radiomic Analysis of Chest X-rays for Pneumonia Discrimination

More like this