Brain Tumor Classification Using ResNet-101 Based Squeeze and Excitation Deep Neural Network

Published in 2019 Second International Conference on Advanced Computational and Communication Paradigms (ICACCP), 2019

Recommended citation: Ghosal, P., Nandanwar, L., Kanchan, S., Bhadra, A., Chakraborty, J., & Nandi, D. (2019). Brain tumor classification using ResNet-101 based squeeze and excitation deep neural network. 2019 2nd International Conference on Advanced Computational and Communication Paradigms, ICACCP 2019. [Paper URL]

The brain tumor is one of the leading and most alarming cause of death with a high socio-economic impact in Occidental as well as eastern countries. Differential diagnosis and classification of tumor types (Gliomas, Meningioma, and Pituitary tumor) from MRI data are required to assist radiologists as well as to avoid the dangerous histological biopsies. In the meantime, improving the accuracy and stability of diagnosis is also one challenging task. Many methods have been proposed for this purpose till now. In this work, an automatic tool for classification of brain tumor from MRI data is presented where the image slice samples are passed into a Squeeze and Excitation ResNet model based on Convolutional Neural Network (CNN). The use of zero-centering and normalization of intensity for smooth variation of the intensity over the tissues was also investigated as a preprocessing step which together with data augmentation proved to be very effective. A relative study had been done to prove the efficacy of the proposed CNN model in free tumor database. Experimental evaluation shows that the proposed CNN archives an overall accuracy rate of 89.93% without data augmentation. Addition of data augmentation has further improved the accuracies up to 98.67%, 91.81% and 91.03% for Glioma, Meningioma and Pituitary tumor respectively with an overall accuracy of 93.83%. Promising improvement with reference to sensitivity and specificity compared with some of the state-of-the-art methods was also observed.

Download paper here