An automated diagnostic support system for chest X-ray classification based on chromatic intensity

Abstract

Automated diagnostic support is crucial for improving healthcare efficiency and enabling rapid patient screening. While deep learning models often deliver state-of-the-art performance, their computational demands and reliance on large annotated datasets motivate exploring lightweight statistical alternatives that can complement or precede deep architectures in resource-constrained settings. This work introduces a chromatic-based classification system for chest X-rays that computes statistical descriptors from pixel intensity distributions and applies both traditional supervised learning techniques and a convolutional neural network to achieve accurate predictions. Working with radiographic images from the NIH ChestX-ray14 collection, our system categorizes scans into three distinct groups: high radiolucency, intermediate pattern, and high opacity, grounded in the Beer–Lambert attenuation law and the Fleischner Society terminology. Labels were assigned by mapping eight clinically validated pathologies from the dataset to their corresponding attenuation category. A class-balanced subset of 9,258 images was constructed to ensure equal representation across categories. Comparative evaluation shows that VGG-16 (macro F1-score of 0.639) clearly outperforms SVM (macro F1: 0.364) and Random Forest (macro F1: 0.390) under identical experimental conditions, indicating that pixel-level convolutional feature learning captures spatially localized pathological patterns that global chromatic descriptors cannot, at the cost of reduced interpretability relative to the four handcrafted radiomic features. Although each category corresponds to possible conditions, this work supports decision-making as a reproducible, computationally inexpensive triage tool that complements radiological evaluation by making it faster, rather than eliminating the radiologist’s opinion.