エッジデバイスにおける脳腫瘍分類のための注意駆動型軽量CNNとHaar初期化パラメトリックウェーブレット圧縮

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. In this study, a Parametric Wavelet Transform (PWT) layer was proposed for efficient compression and adaptive feature extraction from brain MRI images. Initialized with Haar wavelet filters, the PWT layer is trainable, enabling the model to learn optimal frequency decompositions directly from the data while preserving critical diagnostic features. MRI images are preprocessed through this PWT layer to selectively extract and stack the approximation (cA) and diagonal detail (cD) subbands, effectively reducing spatial redundancy and enhancing the representation of diagnostically salient structures for downstream classification. A custom lightweight CNN backbone extracts local features from the frequency-domain representations. At the same time, the integrated self-attention module learns relevant patterns and improves the discriminative power across wavelet-transformed inputs. Model interpretability is addressed using Grad-CAM visualizations, which highlight tumor-relevant regions, thereby improving transparency and clinical trust. Based on the experimental findings, the proposed framework achieves a classification accuracy of 96.0%, outperforming benchmark architectures such as MobileNetV2 (93.0%) and MobileNetV3Small (95.2%) while maintaining fewer trainable parameters (~2.8 million) and achieving faster training time. An ablation study was conducted to evaluate the individual contributions of the PWT compression, CNN backbone, and self-attention module, confirming the additive benefit of each component in achieving optimal performance. The model is successfully deployed on a Raspberry Pi 5, confirming its suitability for real-time, point-of-care, edge-based medical imaging applications. Overall, this work introduces a novel combination of adaptive frequency-domain compression, attention-driven refinement, and efficient embedded deployment for robust and interpretable brain tumor classification. .