Classification of brain tumors in magnetic resonance images using a Convolutional Neural Network

Abstract

The National Cancer Institute estimates that 1.5%–1.8% of malignant tumors occur in the central nervous system (CNS), with 88% affecting the brain. In 2020, this impacted approximately 11,000 Brazilians. Brain tumors are classified as primary, originating in the brain or nearby tissues, or secondary, spreading from other organs. Diagnosis of these tumors relies on imaging techniques such as MRI and CT, with MRI preferred for its non-ionizing radiation and superior spatial resolution. Recent advances in machine learning and deep learning, particularly convolutional neural networks (CNNs), have significantly improved tumor classification and diagnosis by automating feature extraction, reducing human error, and improving accuracy. Inspired by neural networks, CNNs process images through convolutional layers, enabling the detection of patterns crucial for confident medical diagnoses. In this work, we use the Kaggle dataset, which contains four classes: tumor, meningioma, pituitary, and glioma. The images were resized to 256×256 pixels and normalized to pixel values ??between 0 and 1. The model employs a simple architecture with three convolutional layers, starting with 32 filters and doubling to 128, interspersed with max-pooling layers to reduce dimensions. The outputs are fed into a dense layer with 64 neurons, ending with a softmax output for class probabilities. We achieved 95.27% accuracy, and the model outperforms other CNNs such as EfficientNet, showing high accuracy and 100% recall for tumor-free images. Challenges remain in distinguishing meningiomas, which can be explored in future work. Its simplicity and lack of advanced preprocessing make it practical for scalable medical diagnosis.