Related Experiment Video
Updated: May 7, 2026
14:08
Automated Midline Shift and Intracranial Pressure Estimation based on Brain CT Images
Published on: April 13, 2013
43.4K
Haar-initialized parametric wavelet compression with attention-driven lightweight CNN for brain tumor classification
1Department of Electronics and Communication Engineering, Federal Institute of Science and Technology, APJ Abdul Kalam Technological University, Kerala, India.
Biomedical Physics & Engineering Express
|January 13, 2026
Summary
This study introduces a novel hybrid AI framework for accurate brain tumor identification from compressed MRI scans. The model achieves 96% accuracy, outperforming existing methods while being efficient for real-time edge deployment.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Signal Processing
Background:
- Accurate brain tumor identification from Magnetic Resonance Imaging (MRI) is crucial for effective treatment planning.
- Existing methods often struggle with compressed data, computational efficiency, and interpretability.
- Need for lightweight, interpretable models suitable for real-time, edge-based applications.
Purpose of the Study:
- To develop a lightweight hybrid framework for efficient and interpretable brain tumor identification from compressed MRI data.
- To integrate a Parametric Wavelet Transform (PWT) with a Convolutional Neural Network (CNN) and Self-Attention mechanism.
- To enhance model interpretability and validate its performance on edge devices.
Main Methods:
- A novel Parametric Wavelet Transform (PWT) layer, initialized with Haar wavelets, was developed for adaptive compression and feature extraction from MRI data.
- A lightweight CNN backbone processed frequency-domain features, enhanced by a multi-head Self-Attention mechanism for improved discriminative power.
- Grad-CAM visualizations were employed for model interpretability, highlighting tumor-relevant regions.
Main Results:
- The proposed framework achieved a classification accuracy of 96.0%, surpassing benchmark models like MobileNetV2 (93.0%) and MobileNetV3Small (95.2%).
- The model demonstrated efficiency with fewer trainable parameters (~2.8 million) and faster training times.
- Successful deployment on a Raspberry Pi 5 confirmed its suitability for real-time, point-of-care, edge-based applications.
Conclusions:
- The hybrid PWT-CNN-Attention framework offers a robust and interpretable solution for brain tumor classification from compressed MRI.
- The integration of adaptive frequency-domain compression and attention mechanisms significantly enhances diagnostic accuracy and efficiency.
- The model's performance on edge devices paves the way for advanced, accessible medical imaging diagnostics.
