Ultra-low dose PET image reconstruction system and method based on deep learning

Abstract

The invention discloses an ultra-low dose PET image reconstruction system and method based on deep learning, and the system comprises a PET and MRI registration and standardization module which carries out the registration and standardization of PET and MRI images; a displacement prediction network module which is used for fully exploring structural information between stacked two-dimensional image slices formed by disassembling the three-dimensional brain image data of the testee and predicting a displacement numerical value; a spatial variable aggregation module which captures a metabolic residual value of each slice relative to an adjacent slice through variable convolution and displacement numerical values, so that the purpose of low-dose image pre-enhancement is achieved, the execution efficiency of subsequent modules is improved, and the performance is improved; a dual-mode fusion coding module based on CNN and adaptive weight adjustment loss is used for fusing the pre-reinforced ultra-low dose PET and MRI slices; based on the generative adversarial network, the accuracy and high efficiency of the process of synthesizing the full-dose PET image are ensured, and meanwhile, the semantic understanding ability of the image is further improved.

Description

technical field [0001] The invention relates to the technical field of medical image engineering, in particular to a deep learning-based ultra-low-dose PET image reconstruction system and method. Background technique [0002] With the rapid development of medical imaging technology and artificial intelligence technology, automatic and semi-automatic computer-aided reconstruction and image quality enhancement systems have been gradually applied in dose reduction and green diagnosis and treatment scenarios in recent years, in order to reduce the radiation exposure level of patients , while improving the image signal-to-noise ratio and pathological analysis capabilities. [0003] At present, image imaging systems for pathological analysis of the human brain mainly include positron emission tomography (PET), magnetic resonance imaging (MRI), single photon emission computed tomography (SPECT), electron computed tomography (CT) and EEG. Image (EEG), in which PET has the advantage...

AI Technical Summary

Problems solved by technology

However, most similar studies have not fully explored the inter-slice spatial information variation pattern of ultra-low-dose PET image slices, but instead regard each image slice as an independent object for synthesis, resulting in the loss of the spatial relationship between slices

3. Most similar studies use common convolution (such as: 3*3 square convolution kernel), and the fixed-scale rectangular convolution kernel will obviously ignore the high-order semantic features in a large number of brain image slices (such as: brain sulcus with local distribution patterns in gyri)

4. The lower the dose requires the model to be more sensitive to the region of interest (ROIs), the ultra-low dose PET image has a large number of blurred boundaries and unavoidable noise, which requires the description of the GAN network during the synthesis process. The processor module must have an accurate judgment on the distribution mode inside the slice, and the existing mainstream synthesis frameworks lack this kind of attention mode with both full-scale and local scales.

Images