Abstract
Early diagnosis of Alzheimer's Disease (AD) is critical but challenging using standard Magnetic Resonance Imaging (MRI). While Positron Emission Tomography (PET) provides superior functional insights, its cost and accessibility limit widespread use. This paper presents GANDALF, an end-to-end framework that utilizes Generative Adversarial Networks (GANs) to synthesize PET scans from MRI data. By incorporating a discriminator-adaptive loss fine-tuning strategy, GANDALF ensures that synthesized images are not only visually realistic but also clinically relevant for multi-class disease staging.
Introduction
Positron Emission Tomography (PET) is the clinical "gold standard" for Alzheimer’s diagnosis, capable of visualizing amyloid plaques and metabolic changes before structural damage occurs. However, PET's high cost, radiation exposure, and logistical requirements make it less accessible than MRI. While MRI is safer and more common, structural changes like brain atrophy are often too subtle for early-stage diagnosis.
Previous research has explored cross-modal synthesis using GANs to bridge this gap. However, these methods often decouple image generation from the diagnostic task. GANDALF addresses this by integrating synthesis and classification into a single training objective, ensuring the generated PET features directly support clinical decision-making.
Methodology: The GANDALF Framework
GANDALF is based on the pix2pix conditional GAN (cGAN) architecture, enhanced with several domain-specific modifications.
Architecture
The framework employs a modified U-Net generator designed to handle the resolution disparity between high-resolution MRI and lower-resolution PET scans. The discriminator serves a dual purpose: it distinguishes between real and synthetic PET scans while simultaneously acting as a multi-class classifier to identify the stage of cognitive impairment.
Discriminator-Adaptive Loss Fine-tuning
To address the inherent instability of GAN training, GANDALF introduces a dynamic loss adjustment strategy. The model utilizes weighted loss components ($\lambda_{GAN}$, $\lambda_{CLS}$, $\lambda_{L1}$) that shift focus over time. Initially, the model prioritizes visual realism. As training matures, the focus shifts toward classification accuracy. A "restore and adjust" mechanism monitors loss oscillations; if the training diverges, the system reverts to the previous best checkpoint and recalibrates the loss weights to maintain stability.
Experimental Setup
The study utilized data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), comprising 1,525 image triplets (T1-MRI, AV45-PET, and FDG-PET) from 1,033 subjects. The dataset spans the clinical spectrum: Cognitive Normal (CN), Early Mild Cognitive Impairment (EMCI), Late Mild Cognitive Impairment (LMCI), and Alzheimer’s Disease (AD). The data was partitioned into 70% training, 10% validation, and 20% testing sets.
Results and Discussion
GANDALF was evaluated against state-of-the-art benchmarks across binary and multi-class classification tasks.
Multi-Class Performance
The primary strength of GANDALF lies in its ability to handle complex multi-class staging. In the three-class task (AD/MCI/CN), GANDALF achieved an accuracy of 78.7%, a significant improvement over previous methods which ranged from 61.1% to 71.3%.
In the highly challenging four-class task (AD/LMCI/EMCI/CN), GANDALF achieved 37.0% accuracy. While numerically lower, this represents a pioneering result for four-way classification using only MRI inputs, outperforming the pix2pix + CNN baseline (33.0%).
Binary Classification
For binary classification (AD vs. CN), GANDALF reached 85.2% accuracy. While some pure CNN approaches achieved higher results (94.1%) in this specific task, the authors note that the "imagined" PET features are most valuable for the subtle distinctions required in early-stage multi-class diagnosis rather than the more obvious structural differences in late-stage AD.
Conclusion
GANDALF demonstrates that viewing GAN-based synthesis as a sophisticated feature extraction process can significantly enhance the diagnostic utility of standard MRI scans. By stabilizing training through adaptive loss fine-tuning and aligning the generation process with clinical goals, the framework provides a promising path toward more accessible and accurate Alzheimer's staging. Future work will focus on further hyperparameter optimization and clinical validation of the synthesized biomarkers.
Summary updated