Image optimization for low-dose ¹⁸F-FDG breast PET/MRI using deep learning: a pilot study

Positron emission tomography with magnetic resonance imaging (PET/MRI) provides noninvasive molecular characterization of breast cancer and has the potential to improve diagnostic accuracy, staging, treatment response assessment, and guide personalized care. Reducing the radiation dose from 2-deoxy-2-[¹⁸F]fluoro-D-glucose (¹⁸F-FDG) to a level similar to digital mammography while maintaining image quality may facilitate clinical utilization. This study was performed to evaluate diagnostic image quality and lesion conspicuity of low-dose ¹⁸F-FDG breast PET/MRI using denoising and to evaluate the effect of denoising on radiotracer uptake semi-quantification.

This pilot study was a secondary analysis of a single-institution prospective study of ¹⁸F-FDG breast PET/MRI for 23 women with primary invasive breast cancer. Random undersampling of the PET data from a 30-min simultaneous prone ¹⁸F-FDG breast PET/MRI was used to produce simulated low-dose (SLD) images approximating 90% reduced injected activity (37 MBq). A deep learning-based, denoising (DN) algorithm was then applied. SLD and DN datasets were evaluated by three readers for image quality and lesion conspicuity and were compared to full-dose images to assess clinical acceptability. Image noise was quantified by liver SUV_{standard deviation}. ¹⁸F-FDG uptake (SUV_max and SUV_mean) was measured in tumor and normal breast tissue and compared between datasets. Wilcoxon signed rank test, repeated measures analysis of variance, and Bland-Altman analyses were performed.

Compared to SLD, DN datasets scored better for artifacts, perceived signal-to-noise ratio, image sharpness/resolution, and overall image quality (p < 0.001). Compared to full-dose images, DN scored better than SLD for diagnostic image quality (p < 0.001) and lesion conspicuity (p = 0.002). For diagnostic image quality, DN was equivalent or slightly inferior to full-dose images in 89.9% (62/69) of reads. Readers preferred DN (76.8%; 53/69) to SLD (4.3%; 3/69), and both were preferred equally in 10.1% (7/69) of reads. Compared to SLD, quantitative image noise was less in DN images (p < 0.0001). Agreement was excellent between all datasets for tumor SUV_max and SUV_mean.

Acceptable image quality may be achieved with low-dose ¹⁸F-FDG breast PET/MRI using denoising, which warrants further prospective validation.