Robust Deep Learning for MRI-Based Alzheimer's Disease Detection: Overcoming Real-World Imaging Challenges

Mohamad Abou Ali, Mostafa Mohieldin, Fadi Dornaika, Ignacio Arganda-Carreras, Rejdi Chmouri, Ali Cherry, Ali Hage-Diab
October 2025
Cite DOI URL

Abstract

Detecting Alzheimer’s disease (AD) from MRI scans remains a significant challenge in medical imaging due to the subtle, low-contrast anatomical changes associated with neurodegeneration. Real-world scans often suffer from noise, motion artifacts, and scanner variability, all of which can severely degrade deep learning (DL) model performance. In our previous work on MRI-based 4-class brain tumor classification, we developed a robustness methodology that employed Gaussian blur and Gaussian noise as augmentation techniques to counteract the impact of various image degradations-including multiple noise types, blurring methods, and simulated patient motion. In this study, we adapt and apply the same robustness strategy to a 4-class AD classification task using a publicly available Kaggle MRI dataset. To address class imbalance, we employed a Conditional Wasserstein GAN with Gradient Penalty (WGAN-GP) to generate a balanced training set of $mathbf2, 5 6 0$ images per class, while retaining the original, unaugmented test set. A fine-tuned EfficientNetV2B0 model achieved 99 % accuracy on the augmented training data; however, when evaluated under real-world challenges, performance dropped to $mathbf4 4. 3 3 %$-significantly lower than the 69.66 % achieved in the brain tumor benchmark. This stark contrast underscores the inherent difficulty of AD classification, exacerbated by low-resolution axial-only scans and minimal interclass visual differences. Although the robustness methodology yielded partial performance recovery, the results highlight the need for higher-resolution, multi-planar imaging and domain-specific preprocessing strategies. This study advances the understanding of robustness in DL-based neuroimaging and underscores the importance of tailored augmentation pipelines for reliable Alzheimer’s diagnosis in clinical settings.

Type
Conference paper
Publication
2025 Eighth International Conference on Advances in Biomedical Engineering (ICABME)
Deep Learning Training Visualization Magnetic Resonance Imaging Training Data Generative Adversarial Networks Brain Modeling Alzheimer's Disease Artificial Intelligence MRI Alzheimer Disease MRI Brain Cancer Deep Learning (DL) Artificial Intelligence (AI) Real-World Simulation Robustness Generative Adversarial Networks (GAN) Gaussian Augmentation
Mohamad Abou Ali
Mohamad Abou Ali
Postdoctoral Researcher

I develop generalizable deep learning methods for biomedical imaging, advancing diagnostic accuracy, data augmentation, and intelligent healthcare systems.

Fadi Dornaika
Fadi Dornaika
Ikerbasque Research Professor

Ikerbasque Research Professor with expertise in computer vision, machine learning, and pattern recognition.

Ignacio Arganda-Carreras
Ignacio Arganda-Carreras
Ikerbasque Research Associate Professor

My research interests include image processing, computer vision, and deep learning for biomedical applications.