Image SNR Enhancement for a Short Axial FOV Brain PET System Using Generative Deep Learning

The signal-to-noise ratio (SNR) of positron emission tomography (PET) images is determined by several factors including the geometry of the scanner. Low system sensitivity caused by a short axial field of view (FOV) results in a low reconstructed image SNR that can complicate clinical decision-making. Therefore, a longer FOV is highly desirable (e.g., a total body geometry). However, this raises the scanner’s cost by increasing the volume of crystals, number of detectors, and readout electronics. We have developed a deep-learning framework to enhance the image quality of data acquired from a prototype brain-dedicated PET insert system for PET/MRI with an axial FOV of just 2.8 cm. We employed a retrospective analysis on 18F-fluorodeoxyglucose PET scans of 28 patients with either Glioblastoma (n=9) or Alzheimer’s disease (n=19) acquired on a commercial PET/MRI scanner with 60 cm diameter and 25 cm axial FOV. From this data we reconstructed low statistics PET images mimicking that acquired from the 2.8 cm axial FOV brain PET prototype using the 25 cm axial FOV commercial system dataset using a "fault-tolerant reconstruction" algorithm, which allowed us to constrain the count statistics from a set of detectors in a single ring of the latter system to match the geometry of the former system. A conditional generative adversarial network (cGAN) was trained and tested using the simulated short axial FOV images as input, with the paired 25 cm axial FOV image data as the target. We performed 5-fold cross-validation and compared the deep learning (DL)-enhanced images to the target images using 4 metrics: peak-signal-to-noise-ratio (PSNR), root mean squared error (RMSE), mean absolute error (MAE), and structural similarity index (SSIM). The DL-enhanced PET images from the 2.8 cm axial FOV system had a median PSNR of 39.09 (interquartile range (IQR): 32.80–45.32), a median SSIM of 0.98 (IQR: 0.97–0.99), a median RMSE of 0.07 (IQR: 0.04–0.09), and a median MAE of 0.004 (IQR: 0.000–0.009). We also assessed the pretrained cGAN model’s performance in a zero-shot denoising task using patient data collected with our first generation PETcoil system. The ability of the cGAN model to enhance the quality of PET images acquired with a short axial FOV suggests a potential method to provide high-quality, high-accuracy images comparable to those of large axial FOV systems.