Modification CNN Transfer Learning for Classification MRI Brain Tumor

Abstract

Identification, or detecting the infected part of a brain tumor on an MRI image, requires precision and takes a long time. MRI (Magnetic Resonance Imaging) is a magnetic resonance imaging technique to examine and take pictures of organs, tissues, and skeletal systems. The brain is essential because it is the center of the nervous system, which controls all human activities. Therefore, MRI of the brain has an important role, one of which is used for analysis or consideration before performing surgery. However, MRI images cannot provide optimal results when analyzed due to noise, and the bone and tumor (lumps of flesh) have the same appearance. AI (artificial intelligence), or digital image processing and computer vision, can analyze MRI images to detect or identify tumors correctly. This study proposes changes to the last layer of CNN (Convolution Neural Network) transfer learning (VGG16, InceptionV3, and ResNet-50) to identify brain tumor disease on MRI. Data were taken from Kaggle with types of glioma, meningioma, no tumor, and pituitary, with a total of 5712 training images and 1311 testing images. The proposed changes include a flattening layer and a pooling layer. The result is that replacing the flatten layer further improves accuracy, and the accuracy of the transfer learning CNNs (VGG16, InceptionV3, and ResNet-50) is 0.918, 0.762, and 0.934, respectively.

Keywords

• modification layer
• cnn transfer learning
• mri brain tumor