Preprocessing and Feature Extraction Evaluation for COVID-19 Detection from Chest X-Ray Images
TL;DR
The experiment revealed that preprocessing and feature extraction do not operate independently; Gaussian filtering actively enhanced the gradient signal that HOG depends on, while grayscale conversion preserved the intensity patterns that LBP encodes, demonstrating that alignment between preprocessing and feature descriptor is what separates high-performing pipelines from poor ones.
Abstract
Timely and accurate COVID-19 diagnosis remains a pressing clinical challenge, particularly in settings where RT-PCR testing is not available or delayed. Chest X-ray imaging offers a widely accessible alternative; yet a fundamental question persists in the literature: does the choice of image preprocessing determine diagnostic success more than the classifier itself? To answer this, a controlled factorial experiment was used to evaluate 108 pipeline configurations on a publicly available X-ray dataset of 251 training and 66 test images spanning three classes: COVID-19, Normal, and Viral Pneumonia. Each configuration combined one of four preprocessing techniques (Grayscale conversion, Histogram equalization, Adaptive histogram equalization, Gaussian filtering), one of three feature extractors (HOG, LBP, GLCM), one of three dimensionality reduction strategies (None, PCA, LDA), and one of three classifiers (SVM, KNN, DT). What emerged from this experiment was striking; accuracy ranged from 39.39% to 96.97% across configurations that used identical classifiers and feature extractors, with preprocessing selection alone accounting for the bulk of this 57.58 percentage point spread. Beyond this finding, the experiment revealed that preprocessing and feature extraction do not operate independently; Gaussian filtering actively enhanced the gradient signal that HOG depends on, while grayscale conversion preserved the intensity patterns that LBP encodes, demonstrating that alignment between preprocessing and feature descriptor is what separates high-performing pipelines from poor ones. The best-performing configuration—Gaussian filtering paired with HOG and SVM—reached 96.97% accuracy with macro-averaged precision, recall, and F1-score of 0.97, running entirely on standard CPU hardware with no deep learning infrastructure required. These findings reframe preprocessing from a routine preparatory step into the central design decision of any classical ML-based COVID-19 screening pipeline.