Brain Tumor Classification in MRI Images Using En-CNN

Hapsari, Peni Agustin Tjahyaningtijas and Dewinda, Julianensi Rumala and Cucun, Very Angkoso and Nurul, Zainal Fanani and Joan, Santoso and Anggraini, Dwi Sensusiato and Peter, M.A van Ooijen and I Ketut, Eddy Purnama and Mauridhi, Hery Purnomo (2021) Brain Tumor Classification in MRI Images Using En-CNN. International Journal of Intelligent Engineering and Systems, 14 (4). pp. 437-451. ISSN 2185-3118

	Text (Hasil Cek Similarity) Brain Tumor Classification in MRI Images Using En-CNN.pdf - Supplemental Material Download (4MB)
	Text (Hasil Peer Review) 9. Brain Tumor Classification in MRI...pdf - Supplemental Material Restricted to Repository staff only Download (711kB)

Abstract

Brain tumors are among the most common diseases of the central nervous system and are harmful. Early diagnosis is essential for patient proper treatment. Radiologists need an automated system to identify brain tumor images successfully. The identification process is often a tedious and error-prone task. Furthermore, brain tumor binary classification is often characterized by malignant and benign because it involves multi-sequence MRI (T1, T2, T1CE, and FLAIR), making radiologist's work quite challenging. Recently, several classification methods based on deep learning are being used to classify brain tumors. Each model's performance is highly dependent on the CNN architecture used. Due to the complexity of the existing CNN architecture, hyperparameter tuning becomes a problem in its application. We propose a CNN method called en-CNN to overcome this problem. This method is based on VGG-16 that consists of seven convolutional networks, four ReLU, and four max-pooling. The proposed method is used to facilitate the hyperparameter tuning. We also proposed a new approach in which the classification of brain tumors is done directly without priorly doing the segmentation process. The new approach consists of the following stages: preprocessing, image augmentation, and applying the en-CNN method. Our new approach is also doing the classification using four MRI sequences of T1, T1CE, T2, and FLAIR. The proposed method delivers accuracy on the MRI multi-sequence BraTS 2018 dataset with an accuracy of 95.5% for T1, 95.5% for T1CE,94% for T2, and 97% for FLAIR with mini-batch size 128 and epoch 200 using ADAM optimizer. The accuracy was 4% higher than previous research in the same dataset.

Actions (login required)

View Item