Hardware-based brain tumor classification using graph Laplacian spectral features.

Area of Science:

• Medical Imaging Analysis
• Graph Signal Processing
• Computational Neuroscience

Background:

• Early and accurate Brain Tumor (BT) identification is challenging due to the complex nature of brain MRI data.
• Graph Signal Processing (GSP) offers a robust framework for analyzing irregular data by modeling images as signals on graphs.
• Graph construction is crucial for GSP performance, ideally allowing data to vary smoothly over its topology.

Purpose of the Study:

• To develop a discriminative Graph Laplacian Spectral (GLS) feature for brain MRI data.
• To investigate the effectiveness of different graph Laplacian matrices (unnormalized, normalized, random walk) for BT detection.
• To validate the proposed GSP framework for efficient and real-time BT classification.

Main Methods:

• Utilized three forms of the graph Laplacian matrix (unnormalized, normalized, random walk) to extract GLS features.
• Modeled brain MRI data as signals on graphs to analyze spatial and spectral characteristics.
• Implemented the framework on the PYNQ-ZU platform for real-time performance validation.

Main Results:

• The unnormalized Laplacian-based GLS feature achieved high classification accuracies: 98.33% on the Br35H dataset and 98.21% on the Kaggle-4600 dataset.
• The proposed method effectively represents tumor-induced modifications in brain structure.
• The framework demonstrated suitability for efficient and real-time BT classification on the PYNQ-ZU platform.

Conclusions:

• Graph Signal Processing and graph topology learning offer a powerful approach to enhance Brain Tumor detection.
• The developed GLS feature provides a highly effective method for modeling brain tissue connectivity.
• The proposed framework is suitable for real-time, efficient, and accurate Brain Tumor classification.