TY - JOUR T1 - Noise reduction for short time PET imaging with conventional and deep learning based methods<strong/> JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 239 LP - 239 VL - 60 IS - supplement 1 AU - Chen Xi AU - Yang Zhang AU - Shu Liao AU - Lifang Pang AU - Hongcheng Shi AU - Yang Lv AU - Yun Dong Y1 - 2019/05/01 UR - http://jnm.snmjournals.org/content/60/supplement_1/239.abstract N2 - 239Objectives: Because of the low sensitivity, clinical PET/CT system requires more than 10 minutes to finish a whole body PET scan. Such long time scanning may bring discomfort to the patient and introduce motion artifacts into the reconstructed image. It is urgently needed to reduce the scanning time while the image is still clinically acceptable. In this work, we proposed a deep learning based method for short-time PET imaging and compared it with conventional PET image de-noising methods. The aim of this work is to investigate the possibility to reduce the PET acquisition time by half, while the image quality is still comparable to that from standard acquisition time. Methods: Two datasets were generated for comparison. The image set A was reconstructed by using the standard protocols adopted in routine clinical practice, and the image set B was reconstructed by using half of the PET acquisition time to simulate the short-time PET imaging. A modified U-Net deep neuro-network was constructed and trained, with image dataset B as input and image dataset A as target. The network structure is displayed in figure 1. Totally 108 clinical images were used for training and other 22 images were used for testing. Also the image set B was processed with conventional image de-noising methods, i.e. Gaussian filtering and non-local mean (NLM) filtering. With image set A as ‘ground truth’, image quality assessment was performed on the images generated from image set B to see the merits of different de-noising methods. Quantitative evaluation included PSNR, NMSE and liver SNR. Additionally, a patch-to-patch comparison method was developed to address the SUV accuracy of small lesions. Images were decomposed into overlapped patches with patch size of 12mm x 12mm x 12mm, and then the mean value of each patch was calculated and compared to the ground truth. Qualitative evaluation was conducted by radiologists with at least 5 years’ experience. The radiologists reviewed the image set A and B in a blind review and were asked to rate a score 1-5 (from the worst to the best) for the lesion detectability. Results: Results of PSNR, NMSE, liver SNR and SUV accuracy are shown in figure 2. Images without any noise reduction techniques gets the lowest score (3) and the Deep learning based method gets the highest score (4.45). The scores of Gaussian filtering, NLM filtering and Ground truth are 4.09, 4.18 and 4.23, respectively. Conclusions: Gaussian filtering shows the worst de-noising performance on short-time PET imaging. The SUV accuracy of small regions drops a lot because the image is over-smoothed. NLM filtering is better than Gaussian filtering on noise reduction. It preserves the SUV accuracy of small regions before and after noise reduction. Deep learning based method outperforms Gaussian filtering and NLM filtering on both noise reduction and quantitatively accuracy. The increased accuracy of SUV may attribute to the prior information learned in the neuro-network. Deep learning based method also receives highest score from radiologists’ assessment. ER -