公开(公告)号：US20250140383A1

公开(公告)日：2025-05-01

申请号：US18634039

申请日：2024-04-12

Applicant: Huazhong University of science and technology

Inventor： Xuming ZHANG ,  Mingwei WEN

IPC: G16H30/40 ,  G06T7/10 ,  G06T9/00 ,  G06T15/00 ,  G06V10/25 ,  G06V10/44

Abstract: Disclosed in the present invention is a temporal information enhancement-based method for 3D medical image segmentation, belonging to the field of medical image segmentation. The method provides a circle transformer module for extraction and fusion of temporal information, and uses temporal input to improve the training effect of a deep learning model, thereby effectively eliminating interference of similar features and blurred images. In a training phase, an input sample is a temporal sequence, the model training effect is enhanced by extracting temporal information, and segmentation results before and after the combination of temporal information are both constrained, so that the model is no longer temporally dependent. In comparison with a training method in which a single sample is input, the present invention can improve the accuracy of an encoder-decoder structure-based segmentation model without costs. In an application phase, only a single frame of 3D image needs to be input, and no sequence needs to be used as an input, resulting in a more flexible application mode.

公开(公告)号：US20240386527A1

公开(公告)日：2024-11-21

申请号：US18556098

申请日：2021-05-07

Applicant: HUAZHONG UNIVERSITY OF SCIENCE AND TECHNOLOGY

Inventor： Xuming ZHANG ,  Yancheng LAN

IPC: G06T5/60 ,  G06T5/70

Abstract: A method for establishing a three-dimensional ultrasound image blind denoising model and a use thereof include: adding a speckle noise to three-dimensional biological structure images of a same size and without speckle noise to obtain a training data set; establishing a three-dimensional denoising network based on an encoding-decoding structure, wherein the encoding structure is used to obtain N feature maps of a three-dimensional input image and perform a downsampling to obtain feature maps of different scales; the decoding structure is used to take a feature map obtained by the encoding structure as an input and reconstruct a three-dimensional image without speckle noise through upsampling; dividing the encoding-decoding structure into a plurality of stages by a downsampling structure and an upsampling structure; training the three-dimensional denoising network using the training data set to obtain a three-dimensional ultrasound image blind denoising model.

公开(公告)号：US20240081648A1

公开(公告)日：2024-03-14

申请号：US18554680

申请日：2021-05-10

Applicant: HUAZHONG UNIVERSITY OF SCIENCE AND TECHNOLOGY

Inventor： Xuming ZHANG ,  Tuo WANG

IPC: A61B5/00 ,  A61B5/055 ,  G06N3/084 ,  G06T7/00

CPC classification number: A61B5/004 ,  A61B5/055 ,  A61B5/7267 ,  G06N3/084 ,  G06T7/0012 ,  G06T2207/30081

Abstract: The present invention discloses a method and a system for prostate multi-modal MR image classification based on a foveated residual network, the method comprising: replacing convolution kernels of a residual network using blur kernels in a foveation operator, thereby constructing a foveated residual network; training the foveated residual network using prostate multi-modal MR images having category labels, to obtain a trained foveated residual network; and classifying, using the foveated residual network, a prostate multi-modal MR image to be classified, so as to obtain a classification result. In the present invention, a foveation operator is designed based on human visual characteristics, blur kernels of the operator are extracted and used to replace convolution kernels in a residual network, thereby constructing a foveated deep learning network which can extract features that conform to the human visual characteristics, thereby improving the classification accuracy of prostate multi-modal MR images.

公开(公告)号：US20240257356A1

公开(公告)日：2024-08-01

申请号：US18635023

申请日：2024-04-15

Applicant: HUAZHONG UNIVERSITY OF SCIENCE AND TECHNOLOGY

Inventor： Xuming ZHANG ,  Mingwei WEN ,  Quan ZHOU

IPC: G06T7/11 ,  G06T3/4046 ,  G06T7/00 ,  G06T7/143 ,  G16H30/40

CPC classification number: G06T7/11 ,  G06T3/4046 ,  G06T7/0012 ,  G06T7/143 ,  G16H30/40 ,  G06T2200/04 ,  G06T2207/20076 ,  G06T2207/20081 ,  G06T2207/20084 ,  G06T2207/20132 ,  G06T2207/30096

Abstract: The disclosure belongs to the field of image segmentation in medical image processing and discloses a three-dimensional medical image segmentation method and system based on short-term and long-term memory self-attention models, in which the method can segment a target area image in the medical image, which includes the following. (1) A training set sample is established. (2) Processing is performed on the original three-dimensional medical image to be segmented to obtain a sample to be segmented. (3) A three-dimensional medical image segmentation network based on short-term and long-term memory self-attention is established and trained. (4) The sample to be segmented is input to the network, and then a segmentation result of the target area in the sample to be segmented is output. By combining CNN and Transformer, a new model for accurate real-time segmentation of the target area (such as a tumor) in the three-dimensional medical image is obtained.

公开(公告)号：US20230316549A1

公开(公告)日：2023-10-05

申请号：US17997693

申请日：2021-04-13

Applicant: Huazhong University of Science and Technology

Inventor： Xuming ZHANG ,  Xingxing ZHU

IPC: G06T7/33 ,  G06T5/20 ,  G06V10/44 ,  G06V10/74 ,  G16H30/20

CPC classification number: G06T7/337 ,  G06T5/20 ,  G06T7/344 ,  G06V10/454 ,  G06V10/761 ,  G16H30/20 ,  G06T2207/20081

Abstract: Disclosed are a method for establishing a non-rigid multi-modal medical image registration model and an application thereof, which pertain to the field of medical image registration. The method comprises: establishing a generative adversarial network GAN_dr, wherein a generator G_dr is used to generate a deformation recovered structural representations, and a discriminator D_dr is used to determine whether the structural representations generated by G_dr has effectively recovered deformations; performing calculation with respect to structural representations of a reference image, a floating image, and an actual registered image in each sample in a medical dataset, and using a calculation result to train GAN_dr; establishing a generative adversarial network GAN_ie, wherein a generator G_ie uses the structural representations as an input to estimate a registered image, and a discriminator D_ie is used to determine whether the estimated registered image is consistent with the actual registered image; using the trained G_dr to generate the deformation recovered structural representations corresponding to each sample in the medical dataset, and training GAN_ie; and after connecting the trained G_ie to G_dr, obtaining the registration model. The present invention can achieve fast and accurate matching of medical images.

公开(公告)号：US20230196528A1

公开(公告)日：2023-06-22

申请号：US18001295

申请日：2021-01-30

Applicant: HUAZHONG UNIVERSITY OF SCIENCE AND TECHNOLOGY

Inventor： Xuming ZHANG ,  Shaozhuang YE

IPC: G06T5/50 ,  G06T3/40 ,  G06T3/60 ,  G06T5/20 ,  G06V10/771 ,  G06V10/80 ,  G16H30/40

CPC classification number: G06T5/50 ,  G06T3/40 ,  G06T3/60 ,  G06T5/20 ,  G06V10/771 ,  G06V10/806 ,  G16H30/40 ,  G06T2207/20081 ,  G06T2207/20221

Abstract: A multimodal medical image fusion method based on a DARTS network is provided. Feature extraction is performed on a multimodal medical image by using a differentiable architecture search (DARTS) network. The network performs learning by using the gradient of network weight as a loss function in a search phase. A network architecture most suitable for a current dataset is selected from different convolution operations and connections between different nodes, so that features extracted by the network have richer details. In addition, a plurality of indicators that can represent image grayscale information, correlation, detail information, structural features, and image contrast are used as a network loss function, so that the effective fusion of medical images can be implemented in an unsupervised learning way without a gold standard.

公开(公告)号：US20190130572A1

公开(公告)日：2019-05-02

申请号：US16094473

申请日：2016-10-09

Applicant: HUAZHONG UNIVERSITY OF SCIENCE AND TECHNOLOGY

Inventor： Xuming ZHANG ,  Fei ZHU ,  Jingke ZHANG ,  Jinxia REN ,  Feng ZHAO ,  Guanyu LI ,  Mingyue DING

IPC: G06T7/00 ,  G06T3/00 ,  G16H30/40 ,  G06T3/60 ,  G06T11/00

Abstract: The present invention discloses a registration method and system for a non-rigid multi-modal medical image. The registration method comprises: obtaining local descriptors of a reference image according to Zernike moments of order 0 and repetition 0 and Zernike moments of order 1 and repetition 1 of the reference image; obtaining local descriptors of a floating image according to Zernike moments of order 0 and repetition 0 and Zernike moments of order 1 and repetition 1 of the floating image; and finally obtaining a registration image according to the local descriptors of the reference image and the floating image. In the present, by using self-similarity of the multi-modal medical image and adopting the Zernike moment based local descriptor, the non-rigid multi-modal medical image registration is thus converted into the non-rigid mono-modal medical image registration, thereby greatly improving its accuracy and robustness.