A Comparative analysis of efficient CNN-based brain tumor classification models

There have been a lot of research works from the past decades and currently going on in the field of medical imaging. Brain tumor classification into various types is an interesting sub-domain of medical imaging, where many researchers are enthusiastic to serve humanity through computer- aided diagnosis (CAD). With the development of deep neural networks (DNNs), there has been a revolutionary change in the accuracy of classification and regression problems, particularly in the medical domain. Inspired by this fact, the proposed technique is based on convolutional neural network (CNN) classification strategy, which efficiently and accurately is able to classify a brain medical resonance imaging (MRI) into three classes (i.e., meningioma, glioma and pituitary tumor). In this chapter, we fine-tune two CNN models (i.e., SqueezeNet and GoogLeNet) for our specific problem of tumor classification into various grades. SqueezeNet model poses AlexNet-level accuracy with 50 times fewer parameters and fast computing capability, while GoogLeNet is one of the first architectures which introduced the inception module that helped significantly in dropping off the number of trainable weights in a network. The main target of the proposed method is to achieve higher level of accuracy alongside efficiency. The proposed method has been trained on a dataset with 3064 slices from 233 patients, having 708, 1426, 930 cases of meningioma, glioma and pituitary tumors, respectively.