Performance Evaluation of ResNet Model for Noise Reduction According to Gaussian Noise Electromagnetic Radiation Imaging

초록

Reducing noise distribution, such as Gaussian noise, is crucial for improving the diagnostic accuracy of digital positron emission tomography (PET) images. PET imaging detects gamma rays, a form of high-energy electromagnetic radiation, emitted from radiopharmaceuticals administered to patients. Recently, deep convolutional neural network (CNN) has been widely utilized in medical imaging. We evaluated the performance of a residual network (ResNet) framework in reducing noise in PET phantom images. The ResNet architecture comprises a 3 x 3 convolution layer, batch normalization, ReLU activation, 16 residual blocks, and a fully connected layer. For training, 1785 PET images from the QIN PET Segmentation Challenge were used as ground truth (GT), with Gaussian noise added at 0.01 and 0.08 standard deviation levels as inputs. Furthermore, the dataset was divided into training, validation, and test sets at a ratio of 8:1:1. Performance metrics, including contrast-to-noise ratio (CNR), coefficient of variation (COY), peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), intensity profiles, and normalized noise power spectrum (NNPS), were employed. The quantitative analyses such as CNR, COY, intensity profile, and NNPS indicated improved image quality by reducing noise distribution, and the output images were similar to the GT images in terms of similarity evaluation, such as PSNR and SSIM. In conclusion, the deep learning-based noise reduction effect enhances PET image quality by improving the signal-to-noise ratio and preserving essential anatomical and functional details in electromagnetic radiation-based medical imaging.

키워드