Machine learning for predicting full-count FDG PET brain images from low-count acquisitions in suspected dementia: a clinical and quantitative evaluation

Abstract Objective. Artificial intelligence methods for denoising low-count FDG PET brain images are usually evaluated using image quality metrics alone, with limited direct clinical assessment, particularly in suspected dementia. This study evaluated a machine-learning image quality transfer (IQT) method for predicting full-count FDG PET brain images from low-count acquisitions using both quantitative metrics and blinded clinical assessment. 
 Approach. Forty-one FDG PET/CT patients with suspected dementia were retrospectively included, with low-count images simulated using 5% of list-mode data. An IQT random forest model employing patch-wise regression was evaluated using image quality metrics, regional Z-scores, and blinded radiologist assessment against standard-count references.
 Main results. AI-predicted images showed an average peak signal to noise ratio (PSNR) improvement of approximately 4 dB and reduced root mean square error (RMSE) compared with low-count images. Clinically, uninterpretable scans were reduced to 0% for each reader, down from 20% and 50% respectively, with a shift from tentative to confident agreement with the reference standard reports. By contrast, the structural similarity index measure (SSIM) and regional Z-score agreement showed no improvement.
 Significance. In patients with suspected dementia, where motion and limited tolerance of long acquisitions are common, this study demonstrates for the first time the potential of an image quality transfer (IQT) method to improve the clinical usability of low-count FDG PET scans. The findings also indicate the importance of task-based clinical evaluation, since quantitative metrics alone were insufficient to capture the clinically relevant improvements observed in image interpretability and reader confidence.