Texture Fusion and Extreme Learning Machine for Interstitial Lung Disease Classification in HRCT Images

Background: Accurate classification of interstitial lung disease (ILD) patterns in HRCT images remains challenging due to overlapping textures across disease types. We propose a computer-aided diagnosis pipeline for classifying five lung tissue classes from HRCT images: ground glass, micronodule, emphysema, fibrosis, and normal. Methods: The method applies a multi-stage texture-enhancement pre-processing step (separating coarse/fine, high/low-contrast, and directional image components) followed by extraction of nine texture, gradient and gradient-orientation descriptors (LBP, RGLBP, BDIP, bilateral, CDF, SDLGFFO, SSGSM, HOG, and gradient orientation). All nine descriptors were evaluated individually, together with one fused configuration, a combination of four of the nine (LBP, RGLBP, BDIP, and HOG), which gave the highest accuracy. Feature dimensionality is reduced using PCA guided by an eigenvalue-based estimator, and classification is performed with an extreme learning machine (ELM), with SVM (RBF kernel) used as a comparison baseline. Results: On a dataset of 5000 HRCT images (1000 per class; 500 train/500 test per class), the fused four-dimensional representation, classified with ELM (without kernel), achieved 92% accuracy—the best result among 46 configurations compared directly on the held-out test set and therefore a test-set-optimized figure rather than an independently validated estimate of generalization performance—outperforming individual descriptors and alternative dimensionality-reduction methods (LLE, Kernel LDA, Laplacian Eigenmaps, Nyström, Random Fourier). Conclusions: The proposed pipeline combines a multi-stage texture-enhancement step with multi-descriptor fusion; the enhancement step is intended to improve the separability of subtle ILD texture patterns, but its individual contribution has not been verified, as an enhanced-versus-unenhanced ablation was not performed.