Machine-Learning Model for Differentiating Round Pneumonia and Primary Lung Cancer Using CT-Based Radiomic Analysis

Medicine (Baltimore) 2025 AI 6 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Page 1
When Pneumonia Looks Like Cancer: A Machine Learning Solution

The Diagnostic Dilemma Round pneumonia is a rare form of lobar pneumonia in adults where infection is confined to one lung lobe by structural connective tissue anomalies, appearing as a well-circumscribed round mass on CT - a presentation that can be radiologically indistinguishable from primary lung cancer.

Why Differentiation Matters Misclassifying lung cancer as pneumonia delays treatment, while misclassifying pneumonia as cancer leads to unnecessary invasive biopsies with risks of pneumothorax and other complications. An accurate non-invasive tool would improve both patient safety and diagnostic efficiency.

The Radiomic Approach This study applied radiomics - quantitative feature extraction from CT images - combined with seven machine learning classifiers to identify textural, shape, and intensity features that distinguish round pneumonia from malignant lung masses.

Key Results Naive Bayes, support vector machine, and random forest models all achieved perfect classification on the full dataset (AUC=1.000). After feature selection down to five key features, Naive Bayes maintained AUC=1.000, accuracy=0.979, sensitivity=0.958, and specificity=1.000.

TL;DR: Machine learning models using CT-derived radiomic features achieve near-perfect accuracy in distinguishing round pneumonia from primary lung cancer, potentially eliminating unnecessary biopsies in this challenging diagnostic scenario.
Pages 2-3
Study Design: Extracting 107 CT Features from 48 Carefully Selected Patients

Patient Selection The study included 24 patients with round pneumonia (confirmed by clinical and radiological improvement after antibiotic therapy) and 24 with histopathologically confirmed NSCLC (12 adenocarcinoma, 7 bronchial carcinoma, 5 squamous cell carcinoma) from a 2020-2025 hospital cohort.

CT Protocol and Segmentation All images were obtained on a 128-slice CT scanner with standardized parameters. Lesion boundaries were manually drawn by two radiologists using 3D Slicer software, and inter-observer consistency was ensured through joint delineation.

Feature Extraction A total of 107 radiomic features were extracted per case, spanning first-order statistics, shape and size descriptors, and texture matrices including gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GLRLM), gray-level size zone matrix (GLSZM), and neighborhood gray tone difference matrix (NGTDM).

Feature Selection An information-gain algorithm was applied to rank all 107 features by their discriminative power and select the five most informative features, reducing model complexity while maintaining or improving performance.

TL;DR: The study extracted 107 quantitative CT features from 48 patients and used information-gain feature selection to identify the five most discriminative features for distinguishing round pneumonia from lung cancer.
Pages 3-4
Seven Classifiers Evaluated: From Naive Bayes to Neural Networks

Model Comparison Framework Seven classifiers were evaluated: Naive Bayes, support vector machine (SVM), Random Forest, Decision Tree, Neural Network, Logistic Regression, and k-nearest neighbors (k-NN). Performance was assessed using AUC, classification accuracy, sensitivity, and specificity.

Top Performers Naive Bayes, SVM, and Random Forest achieved perfect AUC=1.000 on the full 107-feature dataset, indicating that these models can completely separate the two conditions when given rich radiomic feature input.

After Feature Selection After reducing to five features using the information-gain algorithm, Naive Bayes maintained AUC=1.000 with 97.9% accuracy, 95.8% sensitivity, and 100% specificity - demonstrating that a simple, interpretable model with minimal features can perform as well as complex multi-feature models.

Why Naive Bayes Excels Here The Naive Bayes classifier's strong performance with few features suggests that the selected radiomic features provide near-linearly separable discrimination between the two conditions, a favorable property for clinical deployment since simpler models are easier to validate and implement.

TL;DR: Multiple classifiers achieved perfect or near-perfect discrimination, with Naive Bayes maintaining AUC=1.000 even after feature selection down to just five radiomic features - making it both highly accurate and clinically practical.
Page 2
What Radiomics Actually Measures: Beyond What the Eye Can See

Definition Radiomics converts medical images into structured, quantitative data by systematically extracting hundreds of features that describe lesion shape, size, texture, and intensity patterns. These features capture information too subtle or numerous for radiologists to assess manually.

An Intuitive Analogy The authors compare radiomics to photographing a fruit and algorithmically profiling its size, shape, surface texture, and color. Just as this profile can distinguish a Granny Smith from another apple variety, radiomic profiles can distinguish benign from malignant lesions based on quantitative image data.

Radiogenomics Linkage The fundamental goal of radiomics is to link imaging phenotypes to underlying biological behavior and genetic characteristics - a concept called radiogenomics. Features visible on CT may reflect tumor microenvironment, vascularization, and molecular pathology.

The Dual Challenge in Thoracic Radiomics Two core challenges exist: first, accurately extracting phenotypic features from complex thoracic images; second, identifying from thousands of features which ones correlate with underlying genotype and clinical behavior in a generalizable way.

TL;DR: Radiomics quantifies CT image information invisible to human inspection, linking image texture and shape to biological characteristics - enabling machine learning models to differentiate conditions that appear visually similar to radiologists.
Pages 1, 4
How This Model Could Change Clinical Practice

Avoiding Unnecessary Biopsy The current standard for ambiguous lung masses involves biopsy, which carries a risk of pneumothorax (up to 20-30% in some series). A non-invasive radiomic model with near-perfect specificity would allow clinicians to confidently diagnose round pneumonia and initiate antibiotic therapy without biopsy.

Trial of Antibiotics with Imaging Reassessment In clinical practice, round pneumonia often resolves with antibiotics. This model could serve as a first-line triage tool to stratify patients into high-confidence pneumonia (treat with antibiotics and reimage) versus high-confidence malignancy (proceed to tissue diagnosis).

Reducing Intratumoral Heterogeneity Issues Single biopsy samples may not represent the full tumor due to intratumoral heterogeneity. A whole-lesion radiomic analysis integrates information across the entire mass, capturing heterogeneity that a single biopsy site would miss.

Resource-Limited Settings This approach adds diagnostic intelligence on top of existing CT infrastructure without requiring additional specialized equipment or laboratory tests, making it particularly valuable in settings where biopsy expertise or multidisciplinary team resources are limited.

TL;DR: A CT-based radiomic classifier could spare patients with round pneumonia from unnecessary biopsy while ensuring timely diagnosis for true malignancies, improving both patient safety and healthcare resource efficiency.
Pages 4-5
Limitations and Next Steps for Radiomic Differentiation

Small Sample Size With only 48 patients (24 per group), the study's perfect performance metrics may reflect overfitting or limited generalizability. Larger multicenter datasets are essential to confirm that these radiomic features truly generalize across different scanners, imaging protocols, and patient demographics.

Manual Segmentation Variability Lesion boundaries were drawn manually by radiologists, introducing inter-observer variability even with dual-reader review. Automated or semi-automated segmentation tools would improve reproducibility in real-world deployment.

External Validation Needed The study was performed at a single institution. Prospective external validation in different hospital settings with different CT scanners and patient populations is required before clinical implementation.

Expanding to Other Mimics Round pneumonia is one of several lung conditions that mimic malignancy (others include pulmonary carcinoid, hamartoma, and fungal infection). Future work should develop broader classifiers that can distinguish among multiple benign and malignant diagnoses in a multiclass framework.

TL;DR: While results are promising, small sample size and single-center design require larger prospective multicenter validation, and future models should expand to differentiate lung cancer from the full spectrum of benign mimics beyond round pneumonia alone.
Citation: Open Access, 2025. Available at: PMC12440435.