| CPC G06T 11/008 (2013.01) [G06N 3/0475 (2023.01); G06V 10/30 (2022.01); G06V 10/764 (2022.01); G06V 10/774 (2022.01); G06V 10/776 (2022.01); G06V 10/82 (2022.01)] | 12 Claims |

|
1. A medical image noise reduction method, comprising:
obtaining a standard-dose PET image and a constant-value image;
inputting the standard-dose PET image and the constant-value image into a decay function to obtain a respective low-dose noisy PET image and noisy constant-value image;
assembling the low-dose noisy PET image and the noisy constant-value image in a width dimension or a height dimension, and then inputting the assembly of the low-dose noisy PET image and the noisy constant-value image into a trained conjugate generative adversarial network, and outputting a denoised PET image and constant-value image output through the conjugate generative adversarial network;
wherein the conjugate generative adversarial network comprises a generator and a discriminator;
wherein the generator comprises a reflective padding layer, a convolution layer, an instance normalization layer, a nonlinear layer, a residual module, an upsampling layer, and a nonlinear layer; and
wherein the discriminator is a convolutional neural network classifier, and comprises a convolutional layer, an instance normalization layer, and a nonlinear layer;
wherein the generator comprises two parts: feature extraction and image reconstruction;
wherein in the feature extraction part, the input low-dose noisy PET image and noisy constant-value image are first processed using the padding layer, the convolutional layer, the instance normalization layer and the nonlinear layer; then four groups of feature extraction modules are used to perform feature extraction on the low-dose noisy PET image and the noisy constant-value image; then the extracted features are processed using three residual modules;
wherein in the image reconstruction part, the PET image and the constant-value image are first gradually reconstructed through four upsampling modules based on the extracted features; then the reconstructed PET image and constant-value image are processed using the padding layer, the convolution layer and the nonlinear layer, and the denoised PET image and constant-value image are output.
|