Performance Evaluation of Hybrid Supervised and Unsupervised Neural Model for Abnormal Tumor Classification in Brain MR Images

Supervised and unsupervised artificial neural networks have been successfully used for image classification in biomedical applications. Supervised neural networks yield accurate classification results though the computational speed is low. On the contrary, unsupervised neural networks are comparatively faster than supervised networks besides yielding inferior classification accuracy. In this paper, a modified hybrid neural network, namely training free counter propagation neural network (TFCPN) has been proposed for abnormal tumor classification in brain magnetic resonance (MR) images which possess the benefits of both the learning paradigms. The classes of interest are four brain tumor types namely meningioma, astrocytoma. metastase and glioma. A comprehensive feature vector is chosen to discriminate these classes. Classification of brain tumor images is generally in agreement with the expert interpretation of these images. The performance measures of the training free counter propagation network (TFCPN) are compared with the back propagation network (BPN) and the kohonen self organizing map, selected as the representative type for supervised and unsupervised neural networks. Experimental results reveal the superior nature of the hybrid neural network in terms of classification accuracy and convergence rate.