TY - JOUR T1 - Self-supervised Bone Scan Denoising JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 37 LP - 37 VL - 62 IS - supplement 1 AU - Si Young Yie AU - Seung Kwan Kang AU - Donghwi Hwang AU - Hongyoon Choi AU - Jae Sung Lee Y1 - 2021/05/01 UR - http://jnm.snmjournals.org/content/62/supplement_1/37.abstract N2 - 37Objectives: Bone scan is a common clinical practice performed to evaluate skeletal lesions. It is widely used because of its low cost, but suffers from the low sensitivity and high noise level of gamma camera. To reduce the noise levels, it may be helpful to increase radiotracer dose or scan time. However, increasing radiotracer dose results in high radiation exposure to the patient. Also, the longer scan time increases the likelihood of patient movement and reduces the throughput of the gamma camera. Vendor-supplied filters, such as half-time filters, are commonly used to reduce noise levels in gamma camera. However, the half-time filters applied to low-count images (<50% of full count) sometimes result in significant blocky artifact. Noise2Noise (N2N) is a self-supervised training model that doesn’t require large number of corresponding noisy and clean image pairs to train a denoising network. The aim of this study is to explore the feasibility of self-supervised denoising (N2N) in bone scans by comparing its performance with the conventional Noise2Clean (N2C) approach. Methods: For training and evaluating the denoising networks, 99mTc-MDP or DPD bone scan data of 133 patients were acquired (27 males and 106 females, age = 55.3±13.0 years, acquisition time = 854±124 sec, 30 for training and 103 for testing) using a GE Discovery 670 scanner. From the list-mode data, we generated 10 data bins with 10% time of the full scan duration. The matrix and pixel sizes of the images were 1024 × 256 and 2.21 × 2.21 mm2, respectively. We used two-dimensional U-Net to train and test denoising models: N2N and N2C. The dimensions of input and output data were 256 × 256. In network training, the input is a 10% time image, the N2N target is an independent 10% time image, and the N2C target is the full-count image. For each training method, we trained a one-channel input (1CH) network and a two-channel input (2CH) network. The 1CH network uses either one of the anterior or posterior side of the bone scan as an input. On the other hand, the 2CH network uses both sides of the bone scan as a single input. The Mean Square Error was used as the loss function to optimize the network. To evaluate the performance of denoising network, we calculated the Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM) and High Frequency Error Norm (HFEN) with respect to the full-count bone scan image. Results: The N2N denoising network showed equivalent performance to N2C network. PSNR difference between N2N and N2C denoising methods was within 0.2dB (Input: 30.5±6.40, 1CH N2N: 38.1±5.14, 1CH N2C: 38.2±4.89, 2CH N2N: 38.4±4.89, 2CH N2C: 38.5±4.77 in dB). SSIM difference between the denoising methods was within 0.03 (Input: 0.753±0.092, 1CH N2N: 0.893±0.060, 1CH N2C: 0.915±0.046, 2CH N2N: 0.905±0.053, 2CH N2C: 0.916±0.046), and HFEN difference was within 0.02 (Input: 0.796±0.179, 1CH N2N: 0.447±0.062, 1CH N2C: 0.446±0.052, 2CH N2N: 0.455±0.052, 2CH N2C: 0.466±0.052). Overall, the 1CH network outperformed the 2CH network with regard to HFEN. Conclusion: Self-supervised denoising methods were useful for reducing the noise in the low-count bone scan images. N2N method which requires only paired low-count data for network learning showed remarkable denoising performance. The self-supervised denoising method will be useful for reducing bone scan time or radiation dose. ER -