Room: Track 1
Purpose: Achieving diagnostic quality PET scans requires a combination of high injected activity and/or prolonged scan time to obtain a sufficient number of counts. Methods to maintain acceptable image quality in low count PET imaging are desirable as they can lessen population-level radiation doses, minimize patient motion associated with long scan times and allow for imaging radiotracer distributions after multiple half-lives.
The deep CNN uNet architecture has emerged as state of the art for denoising low count PET data. We implemented a novel architecture called dNet which replaces up/down-sampling operations in uNet with dilated convolutions, thus allowing for receptive field size growth without the need to spatially compress and upsample image representations, hypothesizing that this would lead to sharper output images.
Methods: We trained dNet and uNet on a set of 30 brain FDG-PET subjects, and tested performance on 5 FDG-PET subjects. Patients were scanned on a Siemens Biograph mMR PET/MRI scanner. Listmode data were acquired, low count images were created by randomly subsampling 10% of the registered counts. Network outputs were compared to ground truth on the basis of mean absolute percent error (MAPE), peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM).
Results: In the testing set, dNet showed significant performance improvements over uNet across all three image metrics (p<0.01 in all three instances; paired Student’s t-test). Moreover, dNet categorically outperformed uNet on a subject-by-subject basis in the testing set for all image quality metrics.
Conclusion: The novel dNet architecture outperforms uNet for low count PET denoising in an end-to-end test across a number of commonly accepted image quality and structural metrics.