Zhejiang University

Other

Huafeng Liu

Huihui Ye

Jinmin Xu

Bo Wang

Qiupeng Feng

Rui Hu

Dongsheng Ruan

Lide Mu

Ting Wang

Lin Fang

Jianhui Peng

Zijing Zhang

Fuzhen Zeng

Shining Ma

Amanjule Muhashi

Xingjian Yu

Dakun Hu

The present invention discloses a PET image reconstruction method based on Swin-Transformer regularization, the reconstruction model adopted is composed of several iterative modules, each iterative module consists of three parts: a EM iterative layer, a Swin-transformer based regularization layer, a pixel to pixel image fusion layer, the regularization layer is used to learn the prior information representing the image, comprising a convolution kernel used to extract the shallow features of the image, a Swin-Transformer layer used to extract the deep features of the image, and a convolution layer and a residual connection are used to fuse deep and shallow features. The image fusion layer fuses the results of EM iteration and regularization. The invention can reconstruct high quality PET images from Sinogram projection data, and greatly reduces the noise level of PET images while retaining the structural information.

Pet image reconstruction method based on swin-transformer regularization

The method for improving the coincidence time resolution of PET system based on STFT, the method obtains waveform data by setting a point source at a specific location, and then obtains a short term frequency domain amplitude information through short time Fourier transform (STFT), after that, carrying out a train set, a validation set, and a test set partitioning, and a residual neural network model composed of residual modules and a fully connected layer is used for training, achieving estimation of the TOF time of the PET system. The present invention utilizes Monte Carlo simulation experiments to perform STFT transform on waveform data to obtain the short term frequency domain amplitude information, then, a residual neural network model is trained by using the divided train set, and the trained model is tested on the validation set to obtain the trained model, finally, the accuracy of TOF time estimation based on STFT model and the improvement effect on the coincidence time resolution of PET system compared to traditional CFD methods are verified through the test set.

Method for improving the coincidence time resolution of PET system based on STFT

A method for predicting multi-type ECG heart rhythms based on graph convolution includes: acquiring 12-lead ECG signals from a body surface of a patient, and resampling an ECG signal of each lead to a same signal length; constructing a node mutual information pooling U-shaped graph convolution network, and extracting deep features of the ECG signals by using a feature extraction module; performing one-layer one-dimensional convolution on the deep features to obtain a graph feature matrix to be constructed; inputting the obtained undirected graph into a graph encoding module in the graph convolution network, quantitatively calculating node mutual information of the undirected graph by using the graph encoding module, and selecting a node subset with the maximum mutual information to decrease the number of nodes in the undirected graph for down-sampling; inputting the undirected graph with the decreased number of nodes into a graph decoding module.

Method for predicting multi-type electrocardiogram heart rhythms based on graph convolution

The present invention discloses a cardiac transmembrane potential reconstruction method based on a graph convolutional neural network and an iterative threshold contraction algorithm. The cardiac transmembrane potential is econstructed iteratively through the iterative threshold contraction algorithm embedded in the graph convolutional neural network, and the graph convolutional neural network is used to extract the correlation information between nodes in the non-Euclidean dat a of the cardiac transmembrane potential, while retaining the rigorous mathematical calculations of the iterative threshold contraction algorithm, a solution og the cardiac transmembrane potential is obtained after multiple iterations. Based on the reconstructed temporal and spatial sequence of the cardiac transmembrane potential, the present invention can assist in the diagnosis of common heart diseases such as premature ventricular beats, tachycardia, atrial fibrillation, myocardial ischemia and myocardial infarction in arrhythmia, and is useful for subsequent treatment and provide guidance for radiofrequency ablation surgery.

Cardiac transmembrane potential reconstruction method based on graph convolutional neural network and iterative threshold contraction algorithm

The present invention is a high-resolution magnetic resonance fingerprinting method based on radio frequency encoding, a device and storage medium, where radio frequency encoding is introduced into MRF, n time evolution signals can be collected by designing and respectively scanning a set of n radio frequency pulse signals, separate time evolution signals representing n sub-slices can be obtained after decoding and calculating according to a encoding process, and then separate quantitative images of the n sub-slices can be obtained through dictionary matching. Compared with conventional MRF, a slice resolution of the quantitative images obtained in the imaging method of the present invention is increased by n times and a signal-to-noise ratio is not reduced. Through a phantom experiment and a human experiment, accuracy of quantitative effect and an effective improvement of the resolution of the present invention are proved, and the present invention can provide higher-resolution quantitative imaging results.

High-resolution magnetic resonance fingerprinting method and device based on radio frequency encoding

The present invention discloses a PET system attenuation correction method based on a flow model. The flow model adopted is a completely reversible model, the forward and reverse mappings share the same parameters, and the structure itself is a consistency constraint. The model utilizes the spatial correlation of adjacent slices, and adopts the structure of multi-slice input and single-slice output. The model consists of multiple reversible blocks, each of which consists of a enhanced affine coupling layer and a reversible 1×1 convolutional layer, and uses several small u-nets to learn the transformation parameters of the enhanced affine coupling layer. The present invention avoids additional CT or MR scanning, saves scanning cost for the patient, and reduces the damage of CT radiation to the patient; compared with similar methods, higher quality non-attenuation-corrected PET image can be obtained.

Pet system attenuation correction method based on a flow model

The present invention discloses a magnetic resonance fingerprinting imaging method with variable number of echoes, in addition to conventional MRF coding such as changing the excitation pulse angle, the method also introduces the change of the number of echoes, so that quantitative maps of B, B+, Tand T* can be obtained in a single scan. Further, if the echo time corresponding to the in-phase, opposed-phase and in-phase of water and fat is set for three consecutive echoes, the present invention can also image water and fat, and achieve the accurate quantification of B, B+, T, T, [T*]and [T*]. Through in vivo experiments and simulations, the effectiveness of the present invention has been proved. Therefore, the present invention can provide multiple information representations for common brain diseases (glioma) and fatty diseases (such as lipoma, fatty liver, etc.), which is conducive to clinical diagnosis and treatment.

Magnetic resonance fingerprinting imaging method with variable number of echoes

The disclosure provides a modified EPI sequence for acquiring multi-shot and multi-echo images with interleaved blip-up and blip-down phase encoding; the blip-up and blip-down images are processed by topup in FSL to estimate the inhomogeneous main magnetic field Bmap that causes image distortions; the Bmap is then incorporated into the encoding matrix with a low rank constraint to form a joint reconstruction model; the joint reconstruction model is solved to obtain multiple distortion-free images; and the multiple distortion-free images are matched to dictionary to simultaneous acquire the quantitative T(=1/R) and T*(=1/R*) maps. In the phantom and in-vivo measurements, the disclosed method rapidly acquires the comparable quantitative images within one hold-breath (for 20 s) to the conventional mapping method, thus providing important practical application value for evaluation of liver damage, iron level and cancer lesion.

Method for simultaneous multiple magnetic resonance parameter mapping of liver

An attention mechanism-based low-dose dual-tracer PET reconstruction method. The method achieves low-dose dual-tracer PET image reconstruction by an attention mechanism-based convolution network model, and estimates the standard dose and separates dual-tracer PET signals in a sinogram. With the help of deep learning, a feature extraction tool, the method can reconstruct standard-dose single-tracer PET images in a PET Low-Dose Dual-Tracer Sinogram.

Attention mechanism-based low-dose dual-tracer PET reconstruction method

The present invention discloses a method for constructing an intracardiac abnormal activation point location model based on CNN and LSTM. The model can well locate specific locations of abnormal activation points of VT and obtain three-dimensional coordinates of the locations, while obtaining 12-lead body surface potential data of a patient. The method introduces an idea of deep learning into locating of the abnormal activation points of ventricular tachycardia, uses collected QRS data as an input in a training phase, as well as three-dimensional coordinates of the QRS data corresponding to mapping points as a label to train a CNN-LSTM network, utilizes Conv1D to extract features from the input data, employs LSTM for feature fusion in a time domain, and exploits fully connected layers for regression prediction of the three-dimensional coordinates to finally construct the CNN-LSTM network. The network model of the present invention achieves prediction of the locations of the abnormal activation points of VT from a perspective of data drive, thus effectively solving a problem that clinical catheter ablation time consuming and laborious.

Method for constructing intracardiac abnormal activation point location model based on CNN and LSTM

The present invention discloses a low dose Sinogram denoising and PET image reconstruction method based on teacher-student generator, the adopted network model is divided into a Sinogram denoising module and a PET image reconstruction module, the entire network needs to be processed in a training stage and a test stage. In the training stage: the present invention uses the denoising module to denoise the low dose Sinogram, and then makes the reconstruction module use the denoised Sinogram to reconstruct, in which the teacher generator is introduced in the training stage to constrain the whole, the denoising module is decoupled from the reconstruction module, and a better reconstructed image is obtained through training. In the testing stage, the present invention only needs to input low-dose Sinogram to the denoising module to obtain the denoised Sinogram, and then input the denoised Sinogram to the student generator to get the final reconstruction image.

Low dose Sinogram denoising and PET image reconstruction method based on teacher-student generator

The invention discloses a limited-angle CT reconstruction method based on Anisotropic Total Variation. According to the method, through an image reconstruction model using low dose and sparse-view-angle CT images, a fast iterative reconstruction algorithm is combined with an Anisotropic Total Variation method. The problems that in an existing limited-angle CT reconstruction method are effectively solved, such as partial boundary ambiguity, slow convergence speed and unable to accurately solve. In the process of solving the model, the slope filter is introduced in the Filtered Back-Projection to preprocess the iterative equation, and the Alternating Projection Proximal is used to solve the iterative equation, and the iteration is repeated until the termination condition is satisfied; the experimental comparison with the existing reconstruction methods shows that the invention can achieve better reconstruction effect.

Limited-angle CT reconstruction method based on anisotropic total variation

This present invention discloses a dynamic dual-tracer PET reconstruction method based on a hybrid-loss 3D CNN, which selects a corresponding 3D convolution kernel for a 3D format of dual-tracer PET data, and performs feature extraction in a stereoscopic receptive field (down-sampling) and the reconstruction (up-sampling) process, which accurately reconstructs the three-dimensional concentration distributions of two different tracers from the dynamic sinogram. The method of the invention can better reconstruct the simultaneous-injection single-acquisition dual-tracer sinogram without any model constraints. The scanning time required for dual-tracer PET can be minimized based on the method of the present invention. Using this method, the raw sinogram data of dual tracers can be reconstructed into two volumetric individual images in a short time.

Dynamic dual-tracer PET reconstruction method based on hybrid-loss 3D convolutional neural networks

A PET image reconstruction method, including: 1) injecting a PET radioactive tracer into a biological tissue, scanning by a PET device, and detecting and counting coincidence photons to obtain an original protection data matrix; 2) establishing a measurement equation model; 3) splitting the reconstruction problem into a first sub-problem and a second sub-problem; 4) solving the first sub-problem by a filtered back-projection layer, solving the second sub-problem by an improved denoising convolutional neural network, where the filtered back-projection layer and the improved denoising convolutional neural network are connected in series to form a filtered back-projection network (FBP-Net); 5) inputting original projection data into the FBP-Net, and using an image as a tag to adjust parameters of the FBP-Net to reduce an error between an output of the FBP-Net and the tag; and 6) inputting projection data to be reconstructed into the trained FBP-Net to obtain a desired reconstructed image.

Positron emission tomography image reconstruction method

The present invention discloses a DBN based separation method of a mixture of dual-tracer single-acquisition PET signals labelled with the same isotope. It predicts the two separate PET signals by establishing a complex mapping relationship between the dynamic mixed concentration distribution of the same isotope-labeled dual-tracer pairs and the two single radiotracer concentration images. Based on the compartment models and the Monte Carlo simulation, the present invention selects three sets of the same radionuclide-labeled tracer pairs as the objects and simulates the entire PET process from injection to scanning to generate enough training sets and testing sets. When inputting the testing sets into the constructed universal deep belief network trained by the training sets, the prediction results show that the two individual PET signals can been reconstructed well, which verifies the effectiveness of using the deep belief network to separate the dual-tracer PET signals labelled with the same isotope.

Growing community of inventors

Hangzhou, China

Huafeng Liu