A Lightweight Wavelet Group Equivariant Convolution Neural Network for Medical Image Classification

Abstract

Medical image analysis is essential in modern healthcare, demanding advanced computational techniques for accurate diagnosis and treatment. Traditional approaches often rely on human experts, but the potential for fatigue and the growing complexity of medical data have led to increased reliance on computer-assisted interventions. This paper presents a novel approach using a wavelet and group convolutional neural network for medical image classification. The proposed model integrates both spatial and frequency domain features through the use of discrete wavelet transform and group convolution, enhancing classification accuracy while remaining lightweight with only 13 thousand learnable parameters. Four key modules–input transformation block, feature importance estimator block, feature fusion block, and importance-aware fusion block- are introduced to optimize feature extraction and minimize overfitting. The robustness of the proposed model is validated across various imaging modalities, including computed tomography scans, histopathological images, and microscopic images. Experimental results demonstrate that the proposed model yields a significant improvement over existing models on lung and colon cancer 25,000, malaria cell images, and kidney stone datasets, particularly in resource-constrained environments, making it suitable for real-time deployment on low-end devices. The source code of this method is available at: https://github.com/asfakali/WG-Net