Evaluating augmented deep networks for MRI-based brain anomaly classification

Abstract Accurate and early detection of brain anomalies using MRI is critical for effective diagnosis and treatment planning. Recent advances in deep learning, particularly convolutional neural networks, have significantly improved the capabilities of automated medical image analysis, enabling more precise and scalable solutions in clinical settings. However, the performance of such models is highly dependent on the quality, diversity, and annotation detail of the training datasets. Building on this foundation, in this study, we investigate the performance of augmented deep neural networks for MRI-based brain anomaly classification using the BraTS 2021 dataset and our newly introduced Gazi Brains 2025 dataset, which includes MRI scans from 500 patients, a substantial portion of which are annotated for various brain anomalies. The dataset supports both binary (tumor versus normal) and multiclass (seven neurological conditions) classification tasks. We develop and evaluate a series of deep learning and transformer-based models for anomaly classification, with a particular focus on the impact of synthetic data augmentation. Using StyleGANv3 and Guided Diffusion, we generate synthetic MRI scans to enhance training data diversity and examine their effect on model performance. In addition, a ß-VAE–based generative pipeline is employed to further expand the synthetic dataset, providing controllable latent representations and additional variability through VAE sampling. Experimental results show that all three augmentation strategies-StyleGANv3, Guided Diffusion, and ß-VAE- significantly improve classification accuracy compared to baseline models trained solely on the original dataset. DenseNet achieved an accuracy value of up to 91% in binary classification when trained with augmented data, and EfficientNetV2S also achieved up to 72% in multiclass classification. Although StyleGANv3-generated images exhibited superior visual quality (low FID scores), the method was limited in sample volume. In contrast, the diffusion-based approach allowed the creation of larger synthetic datasets, though at the cost of extended training and sampling times. The ß-VAE model produced a moderate number of anatomically coherent samples with lower computational cost, offering a balanced alternative between quality and scalability.