Abstract
3222
Introduction: Preclinical PET imaging is widely used to quantify in-vivo biological and metabolic process at molecular level in small animal imaging. In preclinical PET, low count acquisition has numerous benefits in terms of animal logistics, maintaining integrity in longitudinal multi-tracer studies and increased throughput. Low count acquisition can be realized by either decreasing the injected dose or by shortening the acquisition time. However, both these methods lead to reduced photons, generating PET images with low signal-to-noise ratio (SNR) exhibiting poor image quality, lesion contrast, and quantification accuracy. In this work, we aim to address this tradeoff between increased image noise due to low count statistics and SNR by implementing a deep learning (DL) based framework to synthesize equivalent high count preclinical PET (HC-PET) images from low count PET (LC-PET) images and to evaluate the quantitative accuracy of the DL framework with task-specific quantification of HC-PET images.
Methods: This study was conducted in context of mice injected with [18F]-Fluorodeoxyglucose (FDG) having Patient Derived tumor Xenograft (PDX) implanted in the mammary fat pad. FDG-PET imaging was performed using Siemens Inveon MultiModality PET/CT scanner. Ten minute static images 50 minutes post injection of FDG were used to obtain PET images at different statistics levels i.e. HC-PET (corresponding to 10 minutes frame) and LC-PET (corresponding to 30 seconds frame i.e. 1/20th of HC-PET events). In designing the framework for generating HC-PET, images from LC-PET images we utilized the Residual U-Net (RU-Net) architecture (Fig 1A). The architecture consists of basic encoder-decoder block, similar to the U-Net structure, along with skip connections to preserve resolution between encoder-decoder branches. Residual learning technique was integrated by adding a residual connection between the input and output, which facilitated better feature propagation and faster convergence of the network. Among thirty-eight mice subjects used for the study, twenty six were used for training and validation of the network and the remaining were used for testing the performance of the framework using mean-squared error (MSE) as the loss function. The performance of DL-generated HC-PET images was compared to several existing denoising methods, including Non-Local Means (NLM) and Block Matching and 3D Filtering (BM3D) in terms of structural similarity index metric (SSIM), peak signal-to-noise ratio (PSNR) and normalized root mean square error (NRMSE). Task-specific quantitative analysis was performed on the lesion of interest using mean standardized uptake value (SUVMean). Bias in SUVmean was calculated with respect to HC-PET images. One way ANOVA followed by Holm-Bonferroni post-hoc analysis was performed to check for significant differences (i.e. p≤0.05) in performance between different methods.
Results: The DL framework generated HC-PET images with visually improved image quality and visual lesion contrast compared with the non-DL approaches, as depicted by representative slices in Fig. 1 B-F. RU-Net performed significantly better (p<0.05) than NLM and BM3D in all three metrics of performance (SSIM: 0.98 VS 0.91 for NLM, 0.98 VS 0.93 for BM3D; PSNR: 38.7 VS 34.9 for NLM, 38.7 VS 35.1 for BM3D; NRMSE: 0.31 VS 0.44 for NLM, 0.31 VS 0.42 for BM3D) (Fig 1 G-J). RU-Net exhibited a mean bias of 2.5% in SUVmean compared to HC-PET and SUVmean metrics were highly correlated to SUVmean metrics derived from true HC-PET (ρ=0.97, p≤0.05).
Conclusions: The proposed DL framework to generate HC-PET images performed significantly better than non-DL methods to produce quantitatively accurate results. The generated HC-PET images also exhibited enhanced qualitative metrics i.e. SSIM, PSNR and NRMSE indicating improved image quality for better lesion detection and diagnosis.