Brain tumor classification using convolutional neural networks and transfer learning

Abstract

Brain tumors are one of the top causes of mortality in both children and adults across the world. Early detection of the tumor can give the patient a new chance in life to undergo effective treatment to save them. Despite the great medical and technological advances, the current test methods for diagnosing and classifying brain tumors are prone to human error, since human-assisted manual classification can result in incorrect prognosis and diagnosis. These drawbacks highlight the need of employing a completely automated system for the detection of brain tumors. The emergence of deep learning and its successes in classification of images warranted by its performance and ability to generalize on various data, led us naturally to use it to solve this problem. This work aims to be a concise exposition of deep learning architectures applied to medical imaging, with a focus on the analysis of MRI images for the automatic classification of brain tumors for the early diagnosis purposes. We consider classification as a supervised learning problem and we address it by means of Convolutional Neural Networks (CNN). Two different CNN models are proposed for two separate classifications, with changing and tuning various hyper-parameters. Two datasets were used, the first dataset of brain MRI Images provided by Navoneel Chakrabarty and the second dataset acquired from the Kaggle platform under the name BT-multiclass. The Using the first proposed model, brain tumor detection is accomplished with 91% percent accuracy. With an accuracy of 92% percent, the second proposed model can classify brain tumors into four types: non-tumor, glioma, meningioma, and pituitary. Using transfer learning, the proposed CNN models for both classifications are then compared to other popular pre-trained CNN models such as Inception-v3, ResNet-50, and VGG-16; and satisfactory findings are obtained. Thus, the inclusion of this type of methodologies favors both the patient and the physician, making it possible to carry out more precise quantitative diagnoses