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

Authors

Abstract

Early and accurate identification of Brain Tumors (BT) is one of the most challenging problems due to the complex, non-Euclidean, and irregular characteristics of brain MRI data. Graph Signal Processing (GSP) offers a robust framework for accurately depicting irregular neighborhood connectivities by modeling brain images as signals on graphs and enables the simultaneous analysis of both the spatial and spectral characteristics of the data. A key aspect of GSP is the construction of an appropriate graph, and thus the performance of the resultant graph-based representation and algorithms depends on the definition of the graph used. However, defining such graphs across diverse application domains is often complex. Ideally, the constructed graph should allow the data to exhibit smoothness or regularity over its topology. To overcome this issue, this study discusses the graph Laplacian, or graph topologies, allowing the brain MRI data to vary smoothly across the graph. We utilize this foundation by employing three forms of the graph Laplacian matrix, such as unnormalized, normalized, and random walk, to extract a discriminative Graph Laplacian Spectral (GLS) feature that accurately represents tumor-induced modifications in brain structure. Experimental evaluations of the Br35H and Kaggle-4600 MRI datasets demonstrate that an unnormalized Laplacian-based GLS feature achieves classification accuracies of 98.33% for Br35H and 98.21% for Kaggle-4600, while maintaining minimal computational cost. These classification results validate the potential of GSP and graph topology learning to improve BT detection by providing a highly effective method of modeling the inherent connectivity of brain tissue. Furthermore, the implementation of the proposed framework on the PYNQ-ZU platform has validated the suitability of our framework for efficient and real-time BT classification.

Tags

Topics