Brain tumor classification from MRI scans using fine-tuned CNN models

Brain tumors are abnormal tissue masses resulting from the uncontrolled proliferation of brain and surrounding cells. Although they constitute a small percentage of cancer types, they have a high mortality rate. These tumors can be identified through magnetic resonance imaging (MRI); however, manually detecting them from these images is a challenging process requiring specialized expertise due to factors such as tumor location, significant structural differences, irregular shapes and unclear boundaries. In this disease, where rapid and accurate diagnosis is critical, the identification and classification of brain tumors is a subject of primary importance, where deep learning methods find extensive application. In this study, an open-source dataset comprises a combination of three distinct datasets—Figshare, SARTAJ, and Br35H—and a total of 7023 brain MRI images were utilized. In order to enable the convolutional neural network (CNN) models used in the study to extract an optimal number of features from the MR images, various preprocessing techniques were applied, including grayscale conversion, Gaussian blurring, thresholding, dilation, erosion, contour detection, resizing, and normalization. Additionally, data augmentation techniques were employed to prevent overfitting. Five different state-of-the-art convolutional neural network (CNN) models (ResNet50, InceptionV3, EfficientNetB0, EfficientNetB7, and InceptionResNetV2) were utilized with transfer learning, leveraging pretrained weights obtained from training on the ImageNet dataset. Fine-tuning was performed by adding layers tailored to the classification problem in this study. With the ResNet50 model, prediction accuracy rates of 99.85% on the training dataset and 98.47% on the test dataset were achieved.