TY - JOUR T1 - <strong>[<sup>68</sup>Ga]-DOTATATE </strong><strong>PET Image Denoising using Unsupervised Deep Learning can Improve CNR in A Wide Range</strong><strong/> JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 429 LP - 429 VL - 61 IS - supplement 1 AU - JIANAN CUI AU - Kuang Gong AU - Tinsu Pan AU - Quanzheng Li Y1 - 2020/05/01 UR - http://jnm.snmjournals.org/content/61/supplement_1/429.abstract N2 - 429Objectives: [68Ga]-DOTATATE is a radiopharmaceutical which is clinically used for the detection of neuroendocrine tumors (NETs) due to its high affinity for somatostatin receptor subtype 2 (SSTR2). [68Ga]-DOTATATE PET-CT scans show an increased sensitivity and specificity of metastatic uptake which are valuable for NET detection, staging and disease management. As the uptake intensity range of [68Ga]-DOTATATE is significantly larger than traditional tracers such as [18F]-FDG, one big challenge of [68Ga]-DOTATATE PET image denoising is how to keep the lesion uptake in multiscale, especially for small lesions. Previously we have developed an anatomically-guided unsupervised deep learning method, conditional deep image prior (CDIP), for PET denoising and verified its performance through [18F]-FDG data. In this work, we further demonstrated the ability of CDIP for improving [68Ga]-DOTATATE PET image quality in multiscale. Methods: Thirty-seven patients (weight 76.2 ± 19.6kg) injected with 184.35 ± 12.47MBq of [68Ga]-DOTATATE was included in this study. The PET image at 60min post-injection was acquired with GE Discovery 690 scanner. The noisy PET image (matrix size, 128×160×320; voxel dimensions, 2.73×2.73×3 mm3) was treated as the training label and the corresponding CT image of the same size was employed as the network input. The network structure is a 6-layer 3D Unet. Constrained by the network structure and the prior high-quality CT input, the network output is a denoised PET image after training 4000 epochs. Gaussian filtering and NLM filtering guided by CT image were employed as the reference methods. The denoising results for the 37 patients using different methods are shown in Figure1 by bar plot and in Figure 2 by box plot. Figure 3 presents the denoising results of different methods in three different SUV scales (0-30, 0-15 and 0-5). Further analysis based on lesion size was conducted. The contrast-to-noise ratios (CNRs) of a big lesion (Leison1: size, 16212.5 mm3; SUVmax, 30.1) and a small lesion (Lesion2: size, 1173.6 mm3; SUVmax, 3.4) were calculated quantitatively. Results: The CDIP method can have huge improvements for all the patients with different lesion sizes (Figure 1) and improve the [68Ga]-DOTATATE PET image quality in multiscale (Figure 3). Figure 2 displays that CNR improvement ratios of CDIP are significantly higher than Gaussian and NLM (P &lt;0.0001) for all the patients. Quantitative analysis shows that for big lesion (Leison1), CDIP result has a higher CNR (107.77) than Gaussian (96.13) and NLM (97.21). For small and low uptake lesion (Lesion2), both Gaussian (5.90) and NLM (7.02) fail to get a higher CNR than the original noisy PET image (7.25), but CDIP succeeds to achieve a CNR of 8.26. This effect is verified by all the 37 patients’ data. In Figure 1, Gaussian and NLM cannot improve the CNR of the original noisy image for those who have a lesion smaller than 1500 mm3 (Patient 1-8). CDIP works well for all the lesion size. Conclusions: This study demonstrated that CDIP method can improve [68Ga]-DOTATATE PET image quality in multiscale. Furthermore, the results show that compared to Gaussian and NLM methods, CDIP can not only preserve the uptake of big lesions, but also work well for keeping the uptake of small lesions while denoising. ER -