Parametric Wavelet Subband Analysis for Tumor-Specific Feature Localization and Classification in MRI Brain Images

This paper presents an adaptive frequency-domain-based deep learning framework for tumor classification of brain MRI images using a novel Parametric Wavelet Transform combined with a Convolutional Neural Network (CNN) and attention mechanism. Unlike conventional fixed wavelet decompositions, the proposed parametric wavelet layer employs trainable convolutional kernels that learn optimal frequency filters directly from the data, enhancing tumor-relevant feature extraction. The CNN backbone processes these decomposed features, while a self-attention module is integrated to emphasize critical regions and contextual relationships across the feature maps. Extensive experiments were conducted on a four-class brain tumor dataset comprising pituitary, glioma, meningioma tumor, and notumor categories. The proposed model obtained a classification accuracy of 96%, showing superior performance compared to baseline architectures. The Grad-CAM heat maps highlighted the model's attention toward tumor-relevant regions, reinforcing both its interpretability and diagnostic reliability. These results suggest that the integration of parametric wavelet-based preprocessing offers meaningful enhancement to frequency-domain-driven deep learning methods for medical image analysis.