Domain adaptive Models for Multi class Brain Abnormality Classification

Abstract

Brain abnormality classification received much attention due to the rapid increase in cases. In general, brain abnormality impacts the quality of a patient s life. The early detection of the brain abnormality can reduce the serious consequences. Depending on the severity of patient s condition, a visual examination of brain is recommended by the physicians. Medical imaging is a promising technology for detailed analysis of brain and is helpful to diagnose the diseases. In this regard, brain magnetic resonance image classification becomes a prevalent task in the brain tumor diagnosis system. There are three multi-class classification problems that are popular such as tumor type classification, tumor grade classification and Alzheimer s severity level classification. newlineBrain tumor is the abnormal growth of brain cells which can be benign or malignant. The salient characteristics including location, shape, size, and texture are used to discriminate wide range of brain tumors. A separable convolution based neural network model is designed for three-class (meningioma/ glioma/ pituitary) tumor type classification. Feature extraction is performed using two separable convolution blocks, where each block consists of a separable convolution and an average pooling. The separable convolution is devised to optimize the computational cost of the model. A fully connected dense layer followed by a softmax layer is used for the three-class classification. A publicly available brain tumor dataset is considered for the evaluation. The proposed model outperforms the existing models for low resolution data. newline