Shallow Convolution Neural Network Architecture for Malignancy Identification from Brain Images

Abstract

A brain tumor is characterised by abnormal cell growth in the human brain that can be non-cancerous or malignant (cancerous). Early detection of this malignancy can help to cure it timely and reduce the mortality rate. Technological advancements and the emergence of machine learning and deep learning techniques have aided radiologists in the diagnosis of tumors without the use of invasive methods. The Convolutional Neural Network (CNN) is a popular deep learning architecture that contributes significantly to automating computer vision tasks that otherwise need human intelligence. This paper presents a shallow CNN architecture for the automatic classification of brain images as healthy or malignant. The Grid Search method is employed for architecture design as well as to configure shallow CNN with optimal hyper-parameters. The proposed CNN model is much simpler and shallower as compared to existing pre-trained CNN models, requiring fewer computational resources. It also provides accuracy comparable to the pioneer methods for malignancy identification on two public datasets of brain images without using segmentation and hand-crafted feature engineering.