Probabilistic self?learning framework for low?dose CT denoising

Purpose Despite the indispensable role of x-ray computed tomography (CT) in diagnostic medicine, associated harmful ionizing radiation dose is a major concern, as it may cause genetic diseases and cancer. Decreasing patients’ exposure can reduce hence related risks, but would inevitably induce higher quantum noise. Supervised deep learning techniques have been used to train neural networks for denoising low-dose CT (LDCT) images, success such strategies requires massive sets pixel-level paired LDCT normal-dose (NDCT) which are rarely available real clinical practice. Our purpose mitigate data scarcity problem learning-based denoising. Methods To solve this problem, we devised shift-invariant property-based network that uses only images characterize both inherent pixel correlations noise distribution, shaping into our probabilistic self-learning (PSL) framework. The AAPM Low-dose Challenge dataset was network. Both simulated datasets were employed test performance well model generalizability. compared conventional method (total variation (TV)-based), popular (noise2void (N2V)), well-known unsupervised (CycleGAN) by using qualitative visual inspection quantitative metrics including peak signal-noise-ratio (PSNR), structural similarity index (SSIM) contrast-to-noise-ratio (CNR). standard deviations (STD) selected flat regions also calculated comparison. Results PSL improve averaged PSNR/SSIM values from 27.61/0.5939 30.50/0.6797. By contrast, 31.49/0.7284 (TV), 29.43/0.6699 (N2V), 29.79/0.6992 (CycleGAN). STDs be 132.3 HU (LDCT), 25.77 19.95 75.06 (CycleGAN), 60.62 57.28 (NDCT). As low-contrast lesion detectability quantification, CNR 0.202 0.356 0.372 0.383 0.399 (PSL), 0.359 inspection, observed proposed deliver noise-suppressed detail-preserved image, while TV-based lead blocky artifact, N2V produce over-smoothed structures value biased effect, CycleGAN generate slightly noisy results with inaccurate values. We verified generalizability method, exhibited superior among various testing different distribution shifts. Conclusions A convolutional trained without datasets. Qualitatively showed achieve than all competitors, despite terms PSNR, SSIM did not always show consistently better