Low Count PSMA PET Image Reconstruction Using Noise-Aware Deep Learning Framework

Objectives: Recently, prostate specific membrane antigen (PSMA) positron emission tomography (PET) has shown great potential in detection of prostate cancer. While relatively long scanning time of standard-dose PET introduces inevitable radiation risks, deep learning-based reconstruction and enhancement methods can enable low-dose PET scanning and improve the image quality without introducing severe artifacts. In this work, we propose a noise-aware dual Res-UNet framework for standard-dose PSMA PET reconstruction using low-dose PET image. Methods and Experiment: Specifically, the generative network of noise-aware dual Res-UNet is developed to first identify an attention map indicating the location of high intensity noise in the low-dose PSMA PET images, followed by the image reconstruction part incorporating the estimated noise attention map and low-dose PET image to reconstruct the high-quality standard-dose PET image. The adaptive robust loss is applied to replace the most frequently used L1 loss for the Poisson noise distribution. To further ensure sharpness of the reconstructed image and enhance perceptual quality, structural similarity (SSIM) loss is also included. Eight subjects referred for whole-body ⁶⁸Ga-PSMA11 PET/MR scan on a PET/MR scanner (Biograph mMR, Siemens Healthcare, Erlangen, Germany) was recruited for this study. The voxel size is 2.3×2.3×5.0 mm3. The final matrix size of each PET image is 172×172×551. The standard PET exam was performed with a 3.5 min/bed setting and 6 beds in total. List-mode data from the standard exams were saved. fourfold lower dose PET images were obtained by image reconstruction using only 1/4 of the dose count from the list-mode data. To quantitatively evaluate the reconstruction performance, we used the peak signal noise ratio (PSNR) as the evaluation metric.

Results: As result images reconstructed by noise-aware dual Res-UNet in Figure 1 shows, the proposed framework better preserves the physiological and pathological structures while suppressing the noise level, and the granular sensation contained only in low-dose images are effectively reduced. Butter PSNR (+0.8dB) results indicate strong agreement with above observations. Besides, quantitative evaluation of SUVs in critical regions also shows strong consistence with the visual comparison results.

Conclusions: Experimental results showed that the noise-aware dual Res-UNet framework can be applied on real clinical application and enables low-dose PSMA PET imaging or reducing scan time of standard dose PET scans.

• Download figure
• Open in new tab
• Download powerpoint

• Download figure
• Open in new tab
• Download powerpoint