Self-Supervised learning framework to generate preclinical Standard-Count PET from Low-Count PET and multi-objective task-based performance evaluation

Low-Count preclinical PET (LC-PET) imaging has numerous benefits in improving animal logistics, increasing scanner throughput and preserving integrity in longitudinal studies. However, LC-PET suffers from poor signal-to-noise ratio (SNR), quantification uncertainties and segmentation difficulties. Supervised learning techniques are effective at generating Standard-Count PET (SC-PET) images from LC-PET but require large datasets for training. To address these challenges, we developed a self-supervised learning framework to generate SC-PET from LC-PET and evaluated the performance of the framework using fidelity-based, task-based segmentation and quantification metrics for different LC-PET realizations. We developed the Multi-Block Residual Network (MBR-Net) for implementing the Noise-to-Noise (N2N) self-supervised learning principle. Multiple residual units are implemented to facilitate large-scale feature compression for faster propagation and convergence across the network. The performance of N2N MBR-Net generated SC-PET images were benchmarked against supervised DL based methods including Residual UNet (RU-Net) and Dilated Network (D-Net) and non-DL denoisers like Non-Local Means (NLM) and Block Matching and 3D Filtering (BM3D). The N2N MBR-Net generated SC-PET images exhibited significantly (p<0.05) improved image quality expressed in terms of SSIM and NRMSE compared to non-DL approaches and is comparable to supervised DL methods. Tumor segmentation performance for N2N MBR-Net evaluated using Dice score exhibited similar performance to supervised DL methods and better performance than non-DL methods. In terms of SUV
MEAN
bias, N2N MBR-Net exhibited a median bias of 0.7%, 1.2% and 2.8% for 19.8M (~10% of SC-PET), 9.9M (~5% of SC-PET) and 1.6M (~0.8% of SC-PET) photon count level consistent to supervised DL methods and better than non-DL approaches. The N2N MBR-Net successfully generated quantitatively accurate SC-PET images using self-supervised learning.