Beyond CNNS: A Multi- Scale Spectral-Structural Feature Extraction Approach for Interpretable Chest X-Ray Classification

Abstract

Medical image classification, especially the analysis of chest X-ray images, is a complex issue due to high intra-class
variability, low contrast areas, and subtle pathological features. Although convolutional neural networks (CNNs) achieve state-of-the-art performance, they require large amounts of labeled data, have limitations such as high computational cost and low interpretability. To overcome these limitations, we have proposed a multi-scale matrix-based feature extraction framework in this paper. The proposed approach generates a hierarchical multi-resolution matrix representation by 2 × 2 average pooling, which can have various scales from 512 × 512 to 8 × 8. At each scale, you get features like matrix rank, log determinant, trace, and log-scale eigenvalues. These features comprehensively reflect the structural and spectral properties of the image. The features obtained at all scales are combined into a feature vector of size 1101 for grayscale or 3303 for RGB images. This feature vector is given as input to a dense multi-layer perceptron (MLP) classifier which learns the nonlinear decision boundaries and provides an accurate classification. The proposed method significantly reduces the computational complexity while eliminating the need for convolution-based feature learning. The experiment results demonstrate that this approach shows good generalization ability even in low data conditions. This research contributes to the development of hierarchical matrix-based feature extraction method, utilization of spectral eigen features and design of a low-resource, and efficient classification framework.