Haar-initialized parametric wavelet compression with attention-driven lightweight CNN for brain tumor classification on edge devices.

This paper presents a lightweight hybrid framework that integrates a Haar-initialized Parametric Wavelet Transform (PWT) with a Convolutional Neural Network (CNN) enhanced by a multi-head Self-Attention mechanism for efficient and interpretable tumor identification from compressed Magnetic Resonance Imaging (MRI) brain image data. A Parametric Wavelet Transform (PWT) layer, initialized with Haar wavelet filters, performs compression and adaptive feature extraction from brain MRI images, enabling the model to learn optimal frequency decompositions while preserving diagnostic features. MRI images are preprocessed through this PWT layer to selectively extract and stack the approximation and diagonal detail subbands, reducing spatial redundancy and enhancing the representation of diagnostically salient structures. A custom lightweight CNN backbone extracts local features from frequency-domain representations. The integrated self-attention module captures salient features and enhances the discriminative power across wavelet-transformed inputs. Grad-CAM visualizations focussed on explaining the model's predictions and attended to tumor relevant regions. The primary contribution of the proposed model focuses on the overall performance with a classification accuracy of 95.88%, which is higher than the benchmark models of MobileNetV2 (93.1%) and MobileNetV3Small (94.80%) while preserving less trainable parameters and memory footprint. An ablation study confirms the individual contributions towards the overall model performance of PWT compression, the CNN backbone, and the self-attention module. Deploying the model on a Raspberry Pi 5 highlights the potential for real-time, point-of-care, edge-based medical imaging. This work is a pioneering integrated approach incorporating adaptive frequency-domain compression alongside attention-based refinement to produce interpretable and robust designs for embedded implementations of brain tumor classification.