Abstract

This paper presents an automatic classification system for pulmonary diseases in chest X-rays based on geometric normalization. The proposed method consists of three main modules. Module 1: A landmark detector: A ResNet-18 convolutional neural network with coordinate attention mechanism is trained to predict 15 landmarks defining the lung contour, achieving a mean error of 3.61 pixels (median 3.07 pixels) through an ensemble of four models with test-time augmentation. Module 2: Geometric normalizer: a set of landmarks surrounding the lung region is used to geometrically normalize each image. This normalization involves: Generalized Procrustes Analysis used once to obtain a standard lung shape, Delaunay triangulation to build a deformation mesh, and a piecewise affine transformation (warping) to map the original lung region to a standardized region. This process eliminates variations in position, scale, and orientation in the original set. Module 3: Classifier: normalized images are classified into three categories (COVID-19, Viral Pneumonia, and Normal) using a ResNet-18 classifier with transfer learning and a contrast adjustment (using the SAHS (Statistical Asymmetrical Histogram Stretching) method). The classifier was evaluated through five-fold cross-validation on the COVID-19 Radiography Database, demonstrating high stability with 98.60 ± 0.26% accuracy, and 98.00% F1-Macro, confirming the robustness of the approach. Although the classifier trained with original images reached a higher accuracy than using normalized images, Gradient-weighted Class Activation Mapping (Grad-CAM) analysis and the cropping experiment suggest that this advantage is partly driven by acquisition artifacts rather than lung pathology. In contrast, geometrically normalized images outperform their non-aligned artifact-masked/cropped counterparts: 98.60% vs. 96.24% on the COVID-19 Radiography Database and 94.67% vs. 94.17% on a balanced adult–pediatric mixed dataset including pediatric cases from the Kermany dataset, suggesting that anatomical alignment can yield a more reliable and artifact-resistant representation for pulmonary disease recognition.

IPC Classification

Keywords