Enhancing brain tumor classification in MRI images: A deep learning-based approach for accurate diagnosis

Background

Detecting brain tumors from MRI images is crucial for early intervention, accurate diagnosis, and effective treatment planning. MRI imaging offers detailed information about the location, size, and characteristics of brain tumors which enables healthcare professionals to make decisions considering treatment options such as surgery, radiation therapy, and chemotherapy. However, this process is time-consuming and demands specialized expertise to manually assess MRI images. Presently, advancements in Computer-Aided Diagnosis (CAD), machine learning, and deep learning have enabled radiologists to pinpoint brain tumors more effectively and reliably.

Objective

Traditional machine learning techniques used in addressing this issue necessitate manually crafted features for classification purposes. Conversely, deep learning methodologies can be formulated to circumvent the need for manual feature extraction while achieving precise classification outcomes. Accordingly, we decided to propose a deep learning based model for automatic classification of brain tumors from MRI images.

Method

Two different deep learning based models were designed to detect both binary (abnormal and normal) and multiclass (glioma, meningioma, and pituitary) brain tumors. Figshare, Br35H, and Harvard Medical datasets comprising 3064, 3000, and 152 MRI images were used to train the proposed models. Initially, a deep Convolutional Neural Network (CNN) including 26 layers was applied to the Figshare dataset due to its extensive MRI image count for training purposes. While the proposed ‘Deep CNN’ architecture encountered issues of overfitting, transfer learning was utilized by individually combining fine-tuned VGG16 and Xception architectures with an adaptation of the ‘Deep CNN’ model on Br35H and Harvard Medical datasets.

Results

Experimental results indicated that the proposed Deep CNN achieved a classification accuracy of 97.27% on the Figshare dataset. Accuracies of 97.14% and 98.57% were respectively obtained using fine-tuned VGG16 and Xception on the Br35H dataset.100% accuracy was also obtained on the Harvard Medical dataset using both fine-tuned models.