Computerized diagnosis of brain tumor using graph based CNN classification.
Authors
Affiliations (3)
Affiliations (3)
- Dept. of EEE, Arunachala College of Engineering for Women, Vellichanthai, Nagercoil, Kanyakumari District, Tamil Nadu, 629203, India. Electronic address: [email protected].
- Dept. of ECE, Arunachala College of Engineering for Women, Vellichanthai, Nagercoil, Kanyakumari District, Tamil Nadu, 629203, India.
- Dept. of EEE, Arunachala College of Engineering for Women, Vellichanthai, Nagercoil, Kanyakumari District, Tamil Nadu, 629203, India.
Abstract
Magnetic resonance imaging (MRI) is essential in the accurate diagnosis of brain tumors so that sound treatment planning can be done, however, clinical evaluation frequently depends on the subjective interpretation of the expert, which is time consuming. Current computer-aided diagnosis strategies that rely entirely on traditional Convolutional Neural Networks (CNNs) mainly acquire local spatial information and pixel grids and could not specifically represent structural associations and heterogeneity inside the tumor regions. Equally, most graph-based methods fail to directly use segmentation-based tumor structure within a unified learning structure. To overcome such limitations, in this study, a four-stage classification framework is proposed based on MRI images of the publicly accessible Kaggle dataset on the topic of Brain Tumor. The suggested method conducts adaptive bilateral filtering of the noise and edge enhancement and then conducts semantic segregation to isolate the tumorareas. Structural boundary information is coded as the Histogram of Oriented Gradients (HOG) descriptors are obtained in the areas that have been segmented. These region-level representations are then learned as nodes of a tumor-aware graph where the relationship between spatial and feature similarities is represented as edges and are labeled using a hybrid Graph Neural Networkconvolutional Neural Network (GNN-CNN) model, which combines learning similarities between relational graphs with convolutional features generation. The proposed framework includes inter-region dependencies, in addition to local texture cues, which increase the discrimination between benign and malignant tumors. Experimental performance on the Kaggle dataset shows high performance with 98.5% accuracy, 98.5% precision, 97.8% recall, and 97.8% F1-score, which indicates the efficiency and performance of the proposed algorithm in the classification of brain tumors in an automated manner.