Improving ¹⁸F-FDG PET Quantification Through a Spatial Normalization Method.

Area of Science:

• Neuroimaging
• Radiochemistry
• Artificial Intelligence

Background:

• ¹⁸F-FDG PET imaging is crucial for diagnosing brain diseases like tumors, epilepsy, dementia, and Parkinson's.
• Accurate quantification of ¹⁸F-FDG PET requires matched 3D T₁ MRI for anatomical detail.
• Current methods necessitate co-registered MRI, increasing complexity and resource demands.

Purpose of the Study:

• To develop and validate a deep learning-based transfer learning approach for spatial normalization of ¹⁸F-FDG PET brain images.
• To eliminate the requirement for 3D MRI scans in the spatial normalization process.
• To enhance the efficiency and accuracy of quantitative analysis in brain PET imaging.

Main Methods:

• A deep neural network, pretrained on amyloid PET, was fine-tuned using 103 ¹⁸F-FDG PET and MRI datasets.
• The model was tested on 65 internal and 78 external datasets for spatial normalization performance.
• Comparison with Statistical Parametric Mapping (SPM) using FreeSurfer-derived segmentation and SUV ratio calculations.

Main Results:

• The proposed transfer learning method demonstrated superior spatial normalization compared to SPM, with better image matching.
• Higher correlation and intraclass correlation coefficients for SUV ratios were observed across brain regions in both internal and external datasets.
• The method showed robust performance even with diverse datasets from different ethnicities and varying PET scanners/reconstruction algorithms.

Conclusions:

• Transfer learning effectively adapts deep neural networks for ¹⁸F-FDG PET spatial normalization without MRI.
• The approach is resource-efficient and offers improved performance, requiring fewer datasets than traditional deep learning training.
• This technique broadens the applicability of deep learning for brain PET spatial normalization in clinical and research settings.