Ensemble Deep Learning Architectures for Brain Tumor Classification and Segmentation in MRI Images

Abstract

Brain tumors (BTs) are the abnormal growth of brain cells. Depending on aggressiveness, location, and impact on brain function, they are classified as either benign or malignant, originating within the brain or metastasizing from other areas of the body. Traditionally, BTs are detected through medical imaging, combined with clinical assessment and expert interpretation. The emergence of machine learning (ML) and advanced artificial intelligence (AI) techniques, particularly deep learning (DL), has changed the reliance that traditional procedures have on radiologists in early tumor detection. Diagnosing BTs involves various imaging techniques and tests to confirm the presence, type, and characteristics of the tumor. Magnetic Resonance Imaging (MRI) is a non-invasive technique that uses magnetic fields and radio waves to paint detailed images of the brain’s soft tissues, which is essential for identifying the tumor’s nature. MRI is more sensitive than computed tomography (CT); thus, it is more efficient in detecting smaller tumors. This thesis investigates the use of ensemble techniques based on deep convolutional neural network (DCNN) architectures for accurate BT classification and segmen- tation. For classification, the ensemble combines ResNet-50, DenseNet-169, and EfficientNet-vS, resulting in a performance improvement of over 0.2% by achieving an accuracy of 99.437% . For tumor segmentation, a similar ensemble strategy is employed using U-Net, U-Net++, and Attention U-Net. This approach also yields consistent, strong results across all evaluation metrics, with improvements of around 1%. All data used in this thesis were sourced from Kaggle, a widely used public platform for researchers. For the classification task, approximately 20,000 MRI scans were collected from three Kaggle datasets. For the segmentation task, a total of 4,293 MRI scans were obtained from the same platform. The results demonstrate that even general-purpose, publicly available DCNN models can detect, classify, and segment BTs with high accuracy, underscoring their potential for real-world clinical applications.