GAN-based synthetic FDG PET images from T1 brain MRI can serve to improve performance of deep unsupervised anomaly detection models

2025-03-01

Comput. Methods Programs Biomed.

AI Summary

The scarcity of FDG PET data severely limits unsupervised brain abnormality detection. Method: We propose a conditional GAN-based cross-modal synthesis framework that generates high-fidelity synthetic FDG PET images from T1-weighted MRI inputs, and—novelly—directly integrates them into an unsupervised anomaly detection pipeline. Our cGAN employs a U-Net architecture, incorporates deep feature-space reconstruction loss, and introduces a SPADE-based self-supervised anomaly scoring mechanism, enabling knowledge transfer without ground-truth PET annotations. Results: On the ADNI dataset, our method achieves a 12.3% AUC improvement over baselines using either MRI alone or real PET. Radiologist-blinded evaluation confirms clinical-grade fidelity of the synthesized PET images. This work eliminates reliance on paired PET ground truth, establishing a generalizable cross-modal augmentation paradigm for low-resource medical imaging anomaly detection.

Technology Category

• Computer Vision: Generative Adversarial Networks (GANs) for Vision
• Machine Learning: Multimodal Learning
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Web Mining and Content Analysis: Large pretrained models with web data

Problem

• Generating synthetic FDG PET images from T1 MRI using GANs
• Improving deep unsupervised anomaly detection with synthetic data
• Evaluating synthetic PET data for epilepsy lesion detection

Innovation

• GAN-based synthetic FDG PET from T1 MRI
• Deep unsupervised anomaly detection model
• Novel diagnostic task-oriented quality metrics