38results about How to "Improve segmentation performance" patented technology

The invention relates to the technical field of image processing, in particular to a fetus head full-automatic segmentation method based on three-dimensional ultrasound, which comprises the followingsteps of firstly, carrying out data enhancement on a three-dimensional ultrasound volume data set of a fetus head to obtain an enhanced data set; inputting the enhanced data set into a full convolutional neural network, and training a model in an end-to-end volume mapping manner to realize pre-segmentation of the data set; and finally, carrying out iterative optimization processing on a pre-segmentation result by adopting a cascade full convolutional neural network based on an automatic context to obtain a final segmentation result. The invention aims to overcome a plurality of defects in fetal head measurement by the existing two-dimensional ultrasound, so that the subsequent diagnosis efficiency of doctors is improved, and more other prenatal studies are promoted.

The invention relates to application of a deep learning algorithm in the field of medical image analysis, in particular to a retinal vessel segmentation algorithm fusing W-net and a conditional generative adversarial network. According to the method, the problems of low segmentation sensitivity and insufficient micro blood vessel segmentation are well solved, great progress is made in the aspectsof the parameter utilization rate, the information circulation and the feature analysis ability of the network, complete segmentation of the main blood vessel and fine segmentation of the micro bloodvessel are facilitated, the blood vessel intersection is not prone to breakage, and a focus and an optic disc are not prone to being segmented into the blood vessel by mistake. According to the invention, a plurality of network models are fused under the condition of low complexity; the whole segmentation performance on a DRIVE data set is excellent, the sensitivity and accuracy of the method are87.18% and 96.95% respectively, the ROC curve value reaches 98.42%, the method can be used for computer-aided diagnosis in the medical field, and rapid and automatic retinal vessel segmentation is achieved.

The invention discloses a multi-modal brain glioma image segmentation method of an adaptive attention gate. The method comprises the steps of S1, preprocessing the MRI glioma image data to obtain a data sample set; S2, building a segmentation network model of the adaptive attention gate with multi-level features, and training the segmentation model; S3, performing lesion segmentation prediction byusing the trained segmentation network model. According to the present invention, the attention gate can automatically learn a group of weights to represent the importance of each layer of features in the body features, the non-pathological noise of the superficial layer features is adaptively suppresses, and the lesion details are added to the deep layer features; a tumor segmentation network adopts a mixed loss function combining a binary cross entropy loss function and a dice loss function to train a model. According to the method, the semantic information and the detail features of different levels of a deep full convolutional network are fully mined, and under the action of the attention gate, the best glioma segmentation effect is achieved by fusing the features of different levels.

The embodiment of the invention provides an image segmentation annotation method and device, and the method comprises the steps: intercepting an image block from a to-be-segmented image, and enablingthe image block to comprise a to-be-segmented target object; segmenting the image block by using a trained segmentation model to obtain a segmentation result of the target object; wherein the segmentation model is used for predicting segmentation results of different types of objects and endowing the segmentation results of different types of objects with the same true value; and receiving an annotation instruction, and setting the true value of the segmentation result of the target object as the true value specified by the annotation instruction. The embodiment of the invention provides a semi-automatic segmentation annotation tool, and annotation time and labor cost can be reduced under the condition that the same precision as manual annotation is guaranteed.

The invention discloses an SAR image change detection method based on hyper pixel saliency analysis, for mainly solving the problem of low detection precision of the traditional technology. The scheme is as follows: 1) inputting images of the same scenes at two different moments to obtain a log ratio graph of the images; performing guide filtering on the log ratio graph to obtain a filtering graph; 2) performing hyper pixel segmentation on the filtering graph to obtain segmented graphs; 3) calculating the segmented graphs by using a saliency method to obtain saliency graphs; performing threshold processing on the saliency graphs, and combing the two images at different moments to obtain a differential graph; and 4) clustering the differential graph by using a fuzzy local information C mean clustering method to obtain a change graph. According to the SAR image change detection method, hyper pixel segmentation and guide filtering are imported, thereby not only reducing the noise of SAR images, but also effectively improving the accuracy of the SAR images by the obtained saliency graphs, and the SAR image change detection method can be used for detecting change areas.

The invention discloses a cerebrovascular image segmentation method based on semi-supervised learning, which adjusts the weight of unlabeled data consistency loss through a region connectivity model, and reduces the unreliability and noise effect of a teacher model, on the basis of the regional connectivity, a student model can more reliably learn from the teacher model step by step. According to the method, the difficulty that network training needs a large amount of manual annotation is solved, the unannotated data is effectively utilized, and the segmentation performance is well improved.

The invention discloses a low-field-intensity MR stomach image segmentation method based on transfer learning image enhancement. The method mainly solves the problems that a low-field-intensity 3DMR image is large in noise, more in artifacts and less in image data. According to the scheme, the method comprises the following steps: acquiring high-field-intensity and low-field-intensity 3DMR stomachimage data sets and corresponding label data sets, and preprocessing the image data; inputting the preprocessed data into a cyclic generative adversarial network to obtain an enhanced pseudo-low field intensity 3DMR stomach image set; constructing and training a 3D Res3Unet segmentation network; and inputting the pseudo low-field-intensity 3DMR stomach image set into a segmentation network, completing fine adjustment of segmentation network parameters, forming a 3D Res3Unet segmentation network after transfer learning, and inputting test data into the network to obtain a segmentation result.According to the method, the segmentation of the low-field-intensity 3DMR stomach image is realized, and the image segmentation precision is effectively improved.

The invention discloses a self-adaptive network suitable for high-reflection bright spot segmentation in a retinal optical coherence tomography image. The self-adaptive network comprises a feature coding module, a self-adaptive SA module and a feature decoding module, the feature coding module comprises a feature extraction unit and a dual residual DR module, the dual residual DR module comprisestwo residual blocks, the adaptive SA module comprises a feature input end, a deformable convolution layer, matrix multiplication and pixel-level summation, and the feature decoding module reconstructshigh-level features generated by the adaptive SA module, gradually performs feature splicing with local information guided by the dual residual DR module through 2 * 2 deconvolution layer deconvolution, and uses a result obtained through 1 * 1 convolution layer convolution as output of the feature decoding module. The method can simplify the learning process of the whole network and enhance gradient propagation while enhancing feature extraction, and can adapt to segmentation targets of different sizes.

The invention discloses a pointer type instrument automatic reading method based on radial gray scale. The method comprises the following steps: constructing and training a dial scale segmentation convolutional neural network; calculating the circle center and the scale radius of the instrument; positioning and identifying dial scale numbers; respectively counting the sum of gray values of pixel points in the radius direction of the scale segmentation image, the scale reading image and the pointer threshold segmentation image; and calculating the indexing value of each section of scale sectionby section according to a statistical result to finish automatic reading. The invention can meet the reading requirements of various pointer instruments, and can achieve the high-precision and high-robustness automatic instrument reading method based on the strategy of radial gray scale statistics.

The invention relates to a real scene image segmentation method based on contrast self-supervised learning. The method comprises the following steps: designing a contrast self-supervised learning segmentation model and a loss function; the comparison self-supervised learning segmentation model comprises an upper branch and a lower branch, and each branch comprises an encoder, a decoder, a feature projection module and a predictor which are connected in sequence; an input image is randomly cut into two image patches with an overlapping area, and the two image patches are processed by comparing two branches of the self-supervised learning segmentation model; the designed loss function comprises image-level loss and pixel-level context alignment loss, and the image-level loss is mainly used for calculating loss between an output feature map of the feature projection module and a prediction map obtained by the predictor; pixel-level context alignment loss mainly constructs a spatial context relationship between image pixels by maximizing the similarity of two image patch overlapping regions and minimizing the interference degree of a non-overlapping region on feature extraction.

The invention provides a cross-domain adaptive semantic segmentation method and device based on data perturbation; the method comprises the steps: obtaining to-be-processed data and semantic segmentation features after data perturbation is added; determining a loss function based on the to-be-processed data and the semantic segmentation features; acquiring a cross-domain adaptive semantic segmentation model through an error back propagation algorithm training model based on the loss function. According to the invention, disturbance is randomly added to a large amount of label-free data in a target domain, and it is guaranteed that the image subjected to disturbance processing can keep semantic consistency; therefore, the problem of field inconsistency between a source domain and a target domain is solved from two perspectives of data disturbance and a cross-domain prototype classifier; in addition, a targeted design is made for a small amount of supervision problems with higher practical application value in practical application, and excellent segmentation performance is obtained under an adversarial-based learning framework; thereby migrating the knowledge of the existing labeled sample into the new data model.

The invention discloses a medical image segmentation method and device, computer equipment and a storage medium. The method comprises the following steps: inputting a sample image into a PVT feature encoder for global semantic feature extraction to obtain a low-level feature and a plurality of high-level features; performing convolution processing on the low-level features to obtain a boundary prediction map; inputting the plurality of advanced features into a feature pyramid network for multiple times of up-sampling and feature fusion to obtain a plurality of corresponding network features; inputting each network feature into a foreground and background prediction module to obtain a foreground prediction map and a background prediction map; splicing the foreground prediction images to obtain a global foreground prediction image, and splicing the background prediction images to obtain a global background prediction image; and carrying out loss calculation by using the loss function, carrying out back propagation, and updating network parameters to obtain a medical image segmentation model. According to the invention, boundary information is used to guide feature expression, and a correction mechanism of foreground and background prediction difference is used to realize more accurate segmentation.

The invention belongs to the technical field of lung lesion image segmentation, and particularly relates to a lung lesion image segmentation method based on CovSegNet, and the method comprises the following steps: data collection, data preprocessing, model construction, model storage, and model evaluation. The data acquisition is used for acquiring various data sets from pulmonary infection, performing data annotation on images in the acquired data set, and constructing a data set required by model training; the data preprocessing is used for data division, normalization and image scaling, and data expansion is carried out; the model construction is based on a CovSegNet segmentation network model, training data are input, and a parameter model is constructed; the model saves the model after the loss function is not reduced any more; the model evaluation is used for evaluating the stored model through a plurality of evaluation indexes and knowing the related performance of the model.

The invention belongs to the technical field of network data processing, and discloses an improved DME edema area neural network segmentation model construction method, which comprises the steps of performing OCT image denoising preprocessing; based on an improved DeepLab neural network, achieving coarse segmentation of a DME edema area, and designing a DeepLab neural network structure by using hole convolution and a spatial pyramid pooling module; introducing a full-connection conditional random field to optimize a DME edema region boundary; and evaluating the segmentation model precision byutilizing the evaluation indexes of the model segmentation performance. According to the method, the contrast of the image can be improved, the edge texture information of the lesion part is reservedwhile the noise is removed, and a good image data foundation is laid for accurate identification and segmentation of the edema area; good lesion part segmentation performance can be obtained, the feeling visual field is enlarged, the segmentation performance is enhanced, and the segmentation speed of the OCT image is improved.

The invention provides a semantic segmentation method for a tree structure in a three-dimensional tomography image, and relates to the technical field of medical image processing, and the method comprises the steps: obtaining a to-be-tested three-dimensional tomography image; preprocessing the image, obtaining a preprocessed image, and preprocessing comprises the steps of unifying the image resolution, cutting the image to be in a unified size and normalizing the gray value of the image; and inputting the preprocessed image into the tree-structure semantic segmentation network to obtain a semantic segmentation prediction result corresponding to the image. According to the method, the multi-task full convolutional network and the graph convolutional network are used for completing feature extraction together, then a semantic segmentation result of a tree structure is obtained, spatial context information is modeled, structural priori knowledge is explicitly introduced, and good segmentation performance can be achieved.

The present invention discloses an image segmentation algorithm based on an ALR (Automatic Local Ratio)-CV (Chan Vese) model and edge transition. The algorithm comprises the following steps of: 1) obtaining image edge information; 2) calculating and obtaining a distance map of a binary system edge mask according to the binary system edge mask of the edge information; 3) constructing an ETM (Edge Conversion Map) (x, y) according to the distance map; 4) minimizing an energy functional ECV of a CV model, obtaining an equation (2) of a level set function, removing regular terms in the equation (2)to obtain an equation (3), dividing the sides of the equal sign of the equation (3) by an energy item weight coefficient [Lambda]1, performing normalization in the execution process, obtaining an equation (6), replacing a general ratio of the energy item weight coefficient in the original CV model with a local ratio of the energy item weight coefficient, and obtaining an ALR-CV model; and 5) employing the ALR-CV model to segment the edge conversion map ETM (x, y). The automatic local ratio is introduced to automatically perform parameter regulation of the CV model to enhance segmentation performances of the CV model, so that curve evolution cannot be forced to always stay at a fixed position, and part of boundaries can be flexibly neglected.

The invention provides a bronchus segmentation method for a lung CT image. The method comprises the following steps: (a) acquiring the lung CT image and labeling a bronchus; (b) preprocessing the lung CT image, including lung parenchyma extraction and data normalization, and calculating a lung boundary distance map corresponding to the image; (c) inputting the preprocessed image, the voxel coordinates and the distance map to the lung boundary into a 3D Unet network model for end-to-end training to obtain a 3D Unet training model, and in the training process, adopting a Laplace filter to enhance the boundary region of the image and calculating boundary enhancement loss (LBE) and Dice loss; and (d) performing bronchial segmentation based on the obtained 3D Unet training model. According to the invention, the voxel coordinates of the CT image and the distance from the voxel coordinates to the lung boundary serve as additional semantic information to be input into the bronchial segmentation model, boundary enhancement loss is introduced in the model training process, the segmentation performance of the tracheal boundary area is improved to the maximum extent, and leakage and breakage of tracheal segmentation are reduced.

The invention discloses a neural network, a magnetic resonance brain tissue segmentation method and device based on the neural network, computing equipment and a storage medium. The neural network canbe a convolutional neural network, comprises at least one three-dimensional (3D) asymmetric convolution module or a convolutional neural network, and comprises at least one encoder, a bottleneck layer and a decoder which are coupled in sequence, and the at least one 3D asymmetric convolution module is located in at least one of the encoder, the bottleneck layer and the decoder; the 3D asymmetricconvolution module comprises a 3D standard convolution layer and three 3D asymmetric convolution layers corresponding to the 3D standard convolution layer, and the four formed convolution layers are used for carrying out convolution operation on data input into the 3D asymmetric convolution module, sum operation is carried out on convolution operation results and then the convolution operation results are output. By using the method, the brain tissue segmentation precision of the brain MRI image data can be improved.

The invention relates to the field of artificial intelligence, in particular to a skin injury picture segmentation method based on a deep network, and the method comprises the steps: carrying out a segmentation task by using training data instead of manually extracting skin picture features, and learning deep convolutional features suitable for the segmentation task by using the training data; according to the method, preprocessing is very simple, and only normalization of picture pixel values is carried out; besides, compared with a preprocessing mode that TDLS and Jafari use a guide filter,the problem that illumination and contrast change greatly is solved, training data are enriched in a data enhancement mode, and a model learns optimal feature representation by himself / herself to carry out segmentation; according to the method, the index of the true positive rate exceeds that of an existing method, the operation time of the method on a GPU and a CPU is far shorter than that of anexisting model, and real-time skin image segmentation can be achieved; in addition, a fully-connected conditional random field is used as a post-processing method, low-level texture color features canbe effectively utilized, and segmentation of an edge area is sharpened.

The invention discloses a retinal vessel segmentation method and device based on multi-scale matched filtering and particle swarm optimization. The method comprises the following steps: firstly, processing image brightness by using an MSR algorithm to improve the phenomenon of uneven illumination of an eye fundus image; secondly, applying multi-scale Gaussian matched filtering to enhance the contrast of a target blood vessel and a background; and finally, segmenting the image by using a particle swarm optimization OTSU multi-threshold. When a DRIVE data set is segmented, the accuracy rate, sensitivity and specificity of the DRIVE data set are 95.72%, 77.98% and 97.29% respectively. A segmentation result proves that the method is relatively high in accuracy and can be used for segmenting more small blood vessels.

The invention discloses a context attention and fusion network suitable for joint segmentation of multiple types of retinal effusion, which comprises a feature coding module, a context contraction coding CSE module, a context pyramid guidance CPG module and a feature decoding module, and is characterized in that the context contraction coding CSE module is embedded in the feature coding module; the context pyramid guidance CPG module is arranged between the feature coding module and the feature decoding module, and the context contraction coding CSE module is in jump connection with the feature decoding module; according to the retina OCT image, each level feature is selectively aggregated through the feature coding module and the context contraction coding CSE module, multi-scale context information is obtained through the context pyramid guiding CPG module and input into the feature decoding module, and the feature decoding module outputs a segmentation result. According to the context attention and fusion network suitable for joint segmentation of multiple types of retina effusion, the problems that in the prior art, feature aggregation is free of selectivity, and the multi-scale information extraction capacity is low are solved.

The invention discloses a children brain MRI demyelination focus positioning method based on deep semi-supervised segmentation, and the method comprises the steps: firstly constructing a data set, and then constructing a segmentation model which comprises a Student network and a Teacher network; carrying out Student network training by utilizing the constructed data set, and carrying out parameter drift according to the Student network parameters after back propagation to obtain Teamer network parameters; according to the method, a pre-processing mode is designed in a targeted manner, so that the method has good compensation for possible problems in the imaging process. By means of image enhancement, image plus noise and semi-supervision, the model can obtain good segmentation performance only through a small number of annotated images, good robustness is obtained for noise and artifacts, and the problems that the annotation process is large in difficulty, inaccurate and time-consuming and labor-consuming are solved.

The invention discloses a logistics package separation method based on Transform, and the method comprises the following steps: sending an image into an improved Transform semantic segmentation model, dividing the received image into a plurality of image blocks, and transmitting the image blocks into a hierarchical encoder, the encoder outputs multi-level image features with different resolutions by using overlapping feature merging operation and a feedforward neural network in combination with a self-attention mechanism; carrying out feature splicing and fusion by using a lightweight decoder based on a multi-layer perceptron, and predicting package segmentation mask information of the image; performing image morphological post-processing on the mask information, extracting edge information of all the packages, acquiring the distribution condition of the current packages, performing statistics on the distribution condition of the packages on the conveyor belt, and acquiring the package in the forefront of the conveyor belt as a target package; and the target parcel information is used as updating input of a Kalman filtering target tracking link, so that single-piece sorting of the logistics parcels is realized.

The invention discloses a blood vessel image segmentation method, device and equipment based on U-Net. The method comprises the following steps: acquiring a blood vessel segmentation data set; preprocessing the blood vessel segmentation data set; performing image block cutting operation on the preprocessed blood vessel segmentation image to obtain sample data; according to the sample data, building a blood vessel image segmentation network through a Pytorch deep learning framework; and performing blood vessel image segmentation according to the blood vessel image segmentation network, and evaluating a blood vessel image segmentation result. A convolution block in a tube image segmentation network is replaced by a multi-scale feature aggregation block; the first input of the multi-scale feature aggregation block is a multi-scale high-level feature, and the second input of the multi-scale feature aggregation block is a multi-scale low-level feature; according to the blood vessel image segmentation network, the multi-scale high-level features and the multi-scale low-level features in the multi-scale feature aggregation block are fused through the MS-CAM module, the segmentation performance can be improved, and the blood vessel image segmentation network can be widely applied to the technical field of artificial intelligence.

A method for providing an integrated feature map by using an ensemble of a plurality of outputs from a convolutional neural network (CNN) is provided. The method includes steps of: a CNN device (a) receiving an input image and applying a plurality of modification functions to the input image to thereby generate a plurality of modified input images; (b) applying convolution operations to each of the modified input images to thereby obtain each of modified feature maps corresponding to each of the modified input images; (c) applying each of reverse transform functions, corresponding to each of the modification functions, to each of the corresponding modified feature maps, to thereby generate each of reverse transform feature maps corresponding to each of the modified feature maps; and (d) integrating at least part of the reverse transform feature maps to thereby obtain an integrated feature map.

The invention discloses an ear CT image vestibule segmentation method mixing 2D and 3D convolutional neural networks. The method comprises three steps of constructing a data set, designing a 2DCNN segmentation network based on a plurality of depth feature fusion strategies, and designing a 3D DDenseUNet segmentation network. The 2D network adopts an encoder-decoder structure as a backbone network to extract vestibular features of the ear CT image; the method comprises the following steps: firstly, constructing a vestibule, then integrating DenseNet-BC and U-Net network architectures, constructing a 3DDenseUNet network, fusing low-level spatial information and high-level semantic information, and finally realizing precise segmentation of the vestibule. The segmentation network designed for the vestibular structure can obtain segmentation performance better than that of a general segmentation method, and the working efficiency and quality of medical staff in the radiology department are improved. The ear key structure can be accurately and automatically segmented, a doctor is helped to complete a large amount of repeated work, and the burden of the doctor is effectively relieved.

The invention discloses an aerial image segmentation method based on a hierarchical context network. The method comprises the steps of firstly designing and constructing a pixel point-pixel point sub-network, then designing and constructing a pixel point-object sub-network, and forming a hierarchical context network according to the constructed pixel point-pixel point sub-network and pixel point-object sub-network, obtaining hierarchical context information, and then completing the segmentation operation of the aerial image by using the obtained hierarchical context information. According to the method, hierarchical context information of two granularities of semantics and details is constructed, so that the category of a target object is better helped to be judged, the spatial detail information of the target object is described, category feature representation is directly learned from an image by using an unsupervised clustering method, the classification relevance implied by feature representation is utilized, convolutional features are further helped to construct hierarchical context information, and the finally proposed hierarchical context network obtains the optimal segmentation performance on two public competition data sets and GF-2 satellite data.

The invention relates to the technical field of image processing, in particular to an image segmentation method fusing a superpixel block and an overall nested edge, which comprises the following steps: S1, defining the number of seed points; s2, marking an edge probability graph E; s3, initializing seed points, and generating pre-segmentation; s4, generating an initial label L0 by using the spatial distance of the FSLIC and the color energy of the SEEDS; s5, according to the relation between E and L0, the edge of the super pixel is defined as a strong boundary or a weak boundary; s6, based on the weak boundary, according to the color energy, dividing the super-pixel edge again; and S7, outputting a sugarcane aphid image segmentation result. According to the invention, the probability edge graph and the fusion clustering strategy are introduced into the super-pixel segmentation method, and the advantages of the super-pixel method, such as high processing speed and capability of coping with the variability of natural illumination, are brought into full play; the superpixel segmentation method is efficient, reliable and excellent in segmentation performance.

The invention discloses a method for segmenting a myopia macular lesion area in a retina OCT image based on an improved U-shaped network, and the method comprises the steps: building a network structure which comprises an encoder module, a feature aggregation pooling module FAPM, a decoder module and a deep supervision module, wherein the feature aggregation pooling module FAPM is arranged in an encoder channel of the encoder module and used for down-sampling and aggregating global context information and local information, and the deep supervision module is arranged in each layer, except a bottom layer, of the decoder and used for guiding a network to generate a feature map with higher representativeness, in an experiment, the output of the network structure is verified by adopting a joint segmentation task of RBCC damage and myopia traction streak lesion areas in an optical coherence tomography OCT image. According to the method for segmenting the myopia macular lesion area in the retina OCT image based on the improved U-shaped network, automatic segmentation of RBCC damage and myopia traction lines in the retina OCT image is achieved, and the small target segmentation performance is improved.