CN116385725B - Fundus image optic disk and optic cup segmentation method and device and electronic equipment - Google Patents
Fundus image optic disk and optic cup segmentation method and device and electronic equipment Download PDFInfo
- Publication number
- CN116385725B CN116385725B CN202310645552.7A CN202310645552A CN116385725B CN 116385725 B CN116385725 B CN 116385725B CN 202310645552 A CN202310645552 A CN 202310645552A CN 116385725 B CN116385725 B CN 116385725B
- Authority
- CN
- China
- Prior art keywords
- image
- fundus
- optic
- segmentation
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 230000011218 segmentation Effects 0.000 title claims abstract description 117
- 238000000034 method Methods 0.000 title claims abstract description 37
- 210000003733 optic disk Anatomy 0.000 title description 2
- 238000012549 training Methods 0.000 claims abstract description 94
- 238000003062 neural network model Methods 0.000 claims abstract description 26
- 238000012360 testing method Methods 0.000 claims abstract description 23
- 238000012545 processing Methods 0.000 claims abstract description 16
- 230000000007 visual effect Effects 0.000 claims abstract description 9
- 230000006870 function Effects 0.000 claims description 15
- 238000005070 sampling Methods 0.000 claims description 15
- 238000006243 chemical reaction Methods 0.000 claims description 8
- 230000007246 mechanism Effects 0.000 claims description 7
- 238000011176 pooling Methods 0.000 claims description 7
- 230000004927 fusion Effects 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 4
- 238000013507 mapping Methods 0.000 claims description 4
- 230000007423 decrease Effects 0.000 claims description 2
- 208000010412 Glaucoma Diseases 0.000 description 16
- 238000013527 convolutional neural network Methods 0.000 description 12
- 238000000605 extraction Methods 0.000 description 8
- 238000013135 deep learning Methods 0.000 description 7
- 238000010586 diagram Methods 0.000 description 7
- 238000012216 screening Methods 0.000 description 6
- 238000010801 machine learning Methods 0.000 description 4
- 201000004569 Blindness Diseases 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 3
- 230000007547 defect Effects 0.000 description 3
- 238000003745 diagnosis Methods 0.000 description 3
- 201000010099 disease Diseases 0.000 description 3
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 230000002427 irreversible effect Effects 0.000 description 3
- 239000000284 extract Substances 0.000 description 2
- 238000010191 image analysis Methods 0.000 description 2
- 238000007689 inspection Methods 0.000 description 2
- 230000002093 peripheral effect Effects 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 238000012546 transfer Methods 0.000 description 2
- 208000009119 Giant Axonal Neuropathy Diseases 0.000 description 1
- 208000022873 Ocular disease Diseases 0.000 description 1
- 206010047555 Visual field defect Diseases 0.000 description 1
- 206010047571 Visual impairment Diseases 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000008485 antagonism Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000033228 biological regulation Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 210000004556 brain Anatomy 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 201000003382 giant axonal neuropathy 1 Diseases 0.000 description 1
- 238000003709 image segmentation Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000001328 optic nerve Anatomy 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 210000001525 retina Anatomy 0.000 description 1
- 208000024891 symptom Diseases 0.000 description 1
- 208000029257 vision disease Diseases 0.000 description 1
- 230000004382 visual function Effects 0.000 description 1
- 230000004393 visual impairment Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
- G06N3/0455—Auto-encoder networks; Encoder-decoder networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
- G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/03—Recognition of patterns in medical or anatomical images
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Medical Informatics (AREA)
- Databases & Information Systems (AREA)
- Eye Examination Apparatus (AREA)
Abstract
The embodiment of the disclosure relates to the field of medical image processing, and provides a fundus image optic disc cup segmentation method, a fundus image optic disc cup segmentation device and electronic equipment, wherein the method comprises the following steps: acquiring a fundus image to be segmented; inputting the fundus image into a pre-trained optic cup and optic disc segmentation network model, and segmenting a optic cup area and a optic disc area in the fundus image to obtain segmentation results; the visual cup and visual disc segmentation network model is obtained by training according to the following steps: acquiring a plurality of sample fundus images carrying labels, and dividing the plurality of sample fundus images into an image training set and an image testing set; respectively training a preset U-Net network and a preset segvomer network by using an image training set to obtain a fundus image optic disc optic cup segmentation neural network model based on UNet-segvomer; and testing the fundus image optic disc and optic cup segmentation neural network model by using the image test set to obtain a trained optic cup and optic disc segmentation neural network model. The embodiment of the disclosure can improve the segmentation precision of the optic cup of the video disc.
Description
Technical Field
The disclosure relates to the technical field of medical image processing, in particular to a fundus image optic disc cup segmentation method and device and electronic equipment.
Background
Glaucoma is the first irreversible blinding ophthalmic disease worldwide that leads to visual field defects and even blindness by damaging the optic nerve. However, many patients cannot realize that they suffer from glaucoma in the early stages of the disease, and only when the disease progresses to the middle or late stages, the symptoms are perceived and medical care is taken, but at this time, the damage of visual function caused by glaucoma to the patients is irreversible. Although visual impairment due to glaucoma, such as blindness, is irreversible, we can reduce the rate of blindness by timely and effective treatment of patients at the beginning of the onset. Thus, early screening and treatment of glaucoma is critical to the patient.
In the prior art, the inspection of the optic disc on the retina by using fundus images is a commonly used preliminary screening method for glaucoma, and if a suspected glaucoma patient is found by the method, a more comprehensive ophthalmic inspection is performed on the suspected patient to confirm the diagnosis result. However, this approach is relatively subjective, requires a high level of expertise on the diagnosing physician and is relatively time consuming, which makes it impossible to use the method for large-scale screening of glaucoma in areas lacking specialized ophthalmologists and corresponding examination equipment.
With advances in technology, age, digital processors and data driven technologies have seen an exponential growth, and artificial intelligence (Artificial Intelligence) based medical screening systems have become increasingly popular. This provides a viable and cost-effective solution for automatic diagnosis of ocular diseases, including glaucoma.
Deep Learning (Deep Learning) is a subclass of artificial intelligence methods based on artificial neural networks (Learning methods inspired by human brain biological structures), belongs to one of machine Learning, and has good application prospects in the current fundus image analysis. Compared with the traditional Machine Learning (ML) method, the depth Learning has less dependence on artificial factors, automatically learns the internal relation of input data through mathematical expression, and directly extracts useful characteristics from the data, so that the depth Learning is particularly suitable for the field of medical image analysis. Diagnosis of glaucoma using deep learning requires first a quantitative assessment of glaucoma. Currently, the evaluation indexes used in the method for quantitatively evaluating glaucoma clinically include: optic disc (optic disc) diameter, optic cup to optic disc area ratio, perpendicular cup to disc ratio (cup to disc ratio, CDR). Among them, the vertical cup-to-disc ratio is widely accepted by professionals and most commonly used for glaucoma screening, which refers to the ratio of vertical cup diameter (vertical cup diameter) to vertical optic disc diameter (vertical disc diameter). In general, the greater the cup to disc ratio, the higher the probability of developing glaucoma. Thus, accurate segmentation of the optic cup and disc is important for glaucoma screening.
In early studies, segmentation of the optic disc was performed primarily based on manually extracted visual features including gradient information of images, features of stereo graphics pairs, local texture features, super-pixel classifiers, and the like. However, the manner of manually extracting the visual features is difficult to design, takes long time, is easily influenced by the fundus image shooting environment and the quality of the image, and is difficult to stably obtain the high-quality optic cup and optic disc segmentation boundary.
The deep learning technology represented by the convolutional neural network overcomes the limitation of the manual visual feature extraction mode, and achieves a satisfactory effect in the cup-optic disc segmentation task by automatically extracting abstract features from data. At present, a network adopted by a deep learning algorithm in the field of video cup and video disc segmentation mainly comprises: deep learning networks based on full convolutional networks (Fully Convolutional Networks, FCN), deep learning networks based on U-Net, and domain adaptive networks based on generating antagonism networks (GenerativeAdversarial Nets, GANs). However, the network frameworks adopted by the networks are all based on convolutional neural networks (Convolutional Neural Networks, CNN), and the convolutional operator has the problem that local receptive fields are smaller, multiple layers of stacks are required to obtain larger receptive fields, and the information amount obtained by the convolutional neural networks is reduced with the increase of the layers.
From the above, it can be seen that the conventional machine learning method still has some problems in the task of performing glaucoma prediction based on image segmentation, and cannot meet the accuracy requirement.
Disclosure of Invention
The present disclosure aims to solve at least one of the problems in the prior art, and provides a method and an apparatus for dividing a fundus image optic disc and a optic cup, and an electronic device.
In one aspect of the present disclosure, there is provided a fundus image optic disc cup segmentation method including:
acquiring a fundus image to be segmented;
inputting the fundus image into a pre-trained optic cup and optic disc segmentation network model, and segmenting a optic cup area and a optic disc area in the fundus image to obtain corresponding segmentation results;
the pre-trained optic cup and optic disc segmentation network model is obtained through training according to the following steps:
acquiring a plurality of sample fundus images carrying labels, and dividing the plurality of sample fundus images into an image training set and an image testing set;
training a preset U-Net network and a preset segfuel network respectively by utilizing the image training set to obtain a fundus image optic disc optic cup segmentation neural network model based on UNet-segfuel;
and testing the fundus image optic disc and optic disc segmentation neural network model based on UNet-segormer by using the image test set to obtain the pre-trained optic disc and optic disc segmentation neural network model.
Optionally, training the preset U-Net network and the segdormer network by using the image training set to obtain a fundus image optic disc optic cup segmentation neural network model based on UNet-segdormer, including:
performing image size adjustment processing on the sample fundus images in the image training set to obtain fundus training images carrying labels;
training the U-Net network by using the fundus training image carrying the tag to obtain a trained U-Net network;
extracting an interested region in the fundus training image by using the trained U-Net network to obtain an interested region image, wherein the interested region comprises a optic disc region and a visual cup region in the fundus training image;
performing data enhancement processing on the region-of-interest image, and training the segfuel network by using the region-of-interest image subjected to the data enhancement processing and a label corresponding to the region-of-interest image to obtain a trained segfuel network;
and forming the trained U-Net network and the trained segfuel network into the fundus image optic disc and optic cup segmentation neural network model based on the UNet-segfuel.
Optionally, the training the U-Net network by using the fundus training image with the tag to obtain a trained U-Net network includes:
the segmentation step: inputting the fundus training image carrying the tag into the U-Net network, and extracting the region of interest in the fundus training image to obtain a corresponding initial segmentation result;
updating: calculating a loss function value based on the initial segmentation result and the label corresponding to the initial segmentation result, and updating parameters of the U-Net network through back propagation;
repeating the dividing step and the updating step until the loss function value no longer decreases.
Optionally, the extracting the region of interest in the fundus training image by using the trained U-Net network, to obtain a region of interest image, includes:
inputting the fundus training image carrying the label into the trained U-Net network to obtain a video disc segmentation result;
mapping the optic disc segmentation result to the corresponding fundus training image, and cutting the fundus training image into sub-images with preset sizes based on the center of the optic disc segmentation result to obtain the region-of-interest image.
Optionally, the U-Net network comprises an encoder and a decoder, wherein:
the encoder of the U-Net network is used for extracting image characteristics and comprises 4 downsampling modules, wherein each downsampling module comprises 2 convolution layers and 1 maximum pooling layer, and the channel number of the next downsampling module is twice that of the previous downsampling module except the first downsampling module;
the decoder of the U-Net network comprises 4 sub-modules, each sub-module comprises an up-sampling convolution layer and 2 convolution layers, and the output characteristics of each up-sampling convolution layer are respectively subjected to channel splicing with the output characteristics of the corresponding down-sampling module so as to realize characteristic fusion.
Optionally, the segfuel network comprises an encoder and a decoder, wherein:
the encoder of the segfuel network comprises 4 fransformer modules, wherein each fransformer module comprises an efficient self-attention module, a hybrid forward module and an overlapped split sub-graph merging module;
the decoder of the segvormer network includes an upsampling MLP layer and an output MLP layer.
Optionally, the high-efficiency self-attention module is configured to enter an h×w image, and implement a self-attention mechanism based on the following formula (1) and the following formula (2):
（1）
（2）
wherein R represents a scaling factor; n represents H×W, H represents the height of the image, and W represents the width of the image;Krepresenting a sequence of features entered into the high-efficiency self-attention module; c represents the number of channels;representing the conversion of K into a characteristic dimension (+.>) Characteristic sequence of x (c·r); k' represents a characteristic sequence obtained after the Reshape operation is carried out on K;Linear(C·R,C) (K ') represents the feature dimension of K' through the fully connected layerC·RConversion toCThe method comprises the steps of carrying out a first treatment on the surface of the K '' represents the processing of KLinearAnd (5) a characteristic sequence obtained after operation.
Optionally, the mathematical expression of the hybrid forward module is the following formula (3):
（3）
wherein xin represents the output of the high-efficiency self-attention module;Conv3×3 denotes a convolution layer with a convolution kernel of 3×3; GELU is an activation function; MLP means MLP layer; xout is the output of the hybrid forward module.
In another aspect of the present disclosure, there is provided a fundus image optic disc cup segmentation apparatus, the segmentation apparatus including:
the acquisition module is used for acquiring fundus images to be segmented;
the segmentation module is used for inputting the fundus image into a pre-trained optic cup and optic disc segmentation network model, and segmenting a optic cup area and a optic disc area in the fundus image to obtain corresponding segmentation results;
the training module is used for training to obtain the pre-trained video cup and optic disc segmentation network model according to the following steps: acquiring a plurality of sample fundus images carrying labels, and dividing the plurality of sample fundus images into an image training set and an image testing set; training a preset U-Net network and a preset segfuel network respectively by utilizing the image training set to obtain a fundus image optic disc optic cup segmentation neural network model based on UNet-segfuel; and testing the fundus image optic disc and optic disc segmentation neural network model based on UNet-segormer by using the image test set to obtain the pre-trained optic disc and optic disc segmentation neural network model.
In another aspect of the present disclosure, there is provided an electronic device including:
at least one processor; the method comprises the steps of,
a memory communicatively coupled to the at least one processor; wherein,,
the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the fundus image optic disc cup segmentation method described above.
Compared with the prior art, the method and the device have the advantages that the visual cup and optic disc segmentation network model is built through the U-Net network and the segsormer network based on the multi-head attention mechanism, the visual cup and optic disc segmentation network model is used for segmenting the optic disc area and the optic cup area in the fundus image to be segmented, global information can be effectively obtained, the detail characteristics of the optic disc and optic cup image can be better identified, the segmentation precision of the optic disc and optic cup is improved, and the defects that the local receptive field is small, the larger receptive field can be obtained only through multi-layer stacking, and the information amount obtained by the convolutional neural network can be reduced along with the increase of the layer number in the traditional convolutional neural network convolution operator are effectively overcome.
Drawings
One or more embodiments are illustrated by way of example and not limitation in the figures of the accompanying drawings, in which like references indicate similar elements, and in which the figures do not depict a proportional limitation unless expressly stated otherwise.
Fig. 1 is a flowchart of a method for dividing a fundus image optic disc and a optic cup according to an embodiment of the present disclosure;
fig. 2 is a flowchart of training a preset U-Net network and a segvomer network respectively by using an image training set in step S220 according to another embodiment of the present disclosure;
fig. 3 is a schematic structural diagram of a UNet-segfromer-based eye fundus image optic disc optic cup segmentation neural network model according to another embodiment of the present disclosure;
fig. 4 is a flowchart of a method for dividing a fundus image optic disc and a optic cup according to another embodiment of the present disclosure;
fig. 5 is a schematic structural diagram of a fundus image optic disc and cup segmentation device according to another embodiment of the present disclosure;
fig. 6 is a schematic structural diagram of an electronic device according to another embodiment of the present disclosure.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present disclosure more apparent, the embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings. However, those of ordinary skill in the art will understand that in various embodiments of the present disclosure, numerous technical details have been set forth in order to provide a better understanding of the present disclosure. However, the technical solutions claimed in the present disclosure can be implemented without these technical details and with various changes and modifications based on the following embodiments. The following divisions of the various embodiments are for convenience of description, and should not be construed as limiting the specific implementations of the disclosure, and the various embodiments may be mutually combined and referred to without contradiction.
One embodiment of the present disclosure relates to a fundus image optic disc cup segmentation method, the flow of which is shown in fig. 1, including:
step S101, a fundus image to be segmented is acquired.
Step S102, inputting the fundus image into a pre-trained cup and disc segmentation network model, and segmenting a cup region and a disc region in the fundus image to obtain a corresponding segmentation result.
As shown in fig. 1, the pre-trained optic disc segmentation network model is obtained by training according to the following steps:
step S210, a plurality of sample fundus images carrying labels are acquired, and the plurality of sample fundus images are divided into an image training set and an image testing set.
Specifically, step S210 needs to read not only the image data of the sample fundus image but also the tag corresponding to the sample fundus image. Step S210 may transfer the read image data and the tag thereof into a list composed of the training data file path and the training data tag data, and open the corresponding sample fundus image through the opencv-python (cv 2) library, and transfer the read image data, i.e., the tag data, to the subsequent step to complete the training task.
For example, the label of the sample fundus image may include 3, background, cup, disc each, to mark the background, cup area, optic disc area in the sample fundus image, respectively, by background, cup, disc.
And S220, respectively training a preset U-Net network and a preset Segformer network by utilizing an image training set to obtain a fundus image optic disc optic cup segmentation neural network model based on UNet-Segformer.
Illustratively, as shown in FIG. 2, step S220 includes:
step S221, the sample fundus image in the image training set is subjected to image size adjustment processing, and a fundus training image carrying a label is obtained.
Specifically, step S221 may perform image size adjustment on the sample fundus images in the image training set, and adjust each sample fundus image to 640×640 pixels, so as to obtain a plurality of fundus training images carrying the label.
Step S222, training the U-Net network by using the fundus training image carrying the tag to obtain a trained U-Net network.
Illustratively, step S222 includes: the segmentation step: inputting the fundus training image carrying the tag into a U-Net network, and extracting a region of interest in the fundus training image to obtain a corresponding initial segmentation result; updating: calculating a loss function value based on the initial segmentation result and the corresponding label, and updating the parameters of the U-Net network through back propagation; the segmentation and updating steps are repeated until the loss function value is no longer decreasing.
That is, step S222 may first input the fundus training image into the region of interest extraction network, i.e. the U-Net network, perform disc optic cup region of interest extraction, then perform loss calculation with the extraction result and the corresponding label as parameters of the loss function, obtain the corresponding loss function value, update the parameters of the U-Net network through back propagation, and cycle the above operation, i.e. the step of inputting the fundus training image into the U-Net network to update the parameters of the U-Net network is performed cyclically until the loss tends to be stable, i.e. the loss function value is not reduced any more, and obtain the trained U-Net network, i.e. the U-Net region of interest extraction network is constructed.
And S223, extracting an interested region in the fundus training image by using the trained U-Net network to obtain an interested region image, wherein the interested region comprises a optic disc region and a optic cup region in the fundus training image.
Illustratively, step S223 extracts a region of interest in the fundus training image using the trained U-Net network, and obtains a region of interest image, including: inputting fundus training images carrying labels into a trained U-Net network to obtain a video disc segmentation result; mapping the video disc segmentation result to a corresponding fundus training image, and cutting the fundus training image into sub-images with preset sizes based on the center of the video disc segmentation result to obtain an interested region image.
Specifically, since the optic cup is a bright central depression of variable size that exists on the optic disc, the optic disc segmentation result obtained through the trained U-Net network is the predicted segmentation result that includes the optic disc optic cup. Mapping the segmentation result to an original image, namely a corresponding fundus training image, and cutting the fundus training image into sub-images with preset sizes, such as 512 multiplied by 512 pixel sizes, according to the center of the segmentation result to obtain an interested region image comprising a video disc region and a video cup region.
Step S224, performing data enhancement processing on the region of interest image, and training the Segfomer network by using the region of interest image and the corresponding label after the data enhancement processing to obtain a trained Segfomer network.
Specifically, in step S224, the data enhancement processing such as zooming, flipping, cropping, etc. may be performed on the region of interest image, and the specific form of the data enhancement processing is not limited in this embodiment.
And S225, forming a fundus image optic disc and optic cup segmentation neural network model based on the UNet-segmer by using the trained U-Net network and the trained segvomer network.
And step S230, testing a fundus image optic disc and optic cup segmentation neural network model based on UNet-segrormer by using an image test set to obtain a pre-trained optic cup and optic disc segmentation network model.
Specifically, in the testing process, parameters of the eye bottom image optic disc and optic cup segmentation neural network model can be adjusted according to the testing result so as to realize model optimization.
Compared with the prior art, the embodiment of the disclosure constructs the optic cup and optic disc segmentation network model by adopting the U-Net network and the segsormer network based on the multi-head attention mechanism, and segments the optic disc region and the optic cup region in the fundus image to be segmented by utilizing the optic cup and optic disc segmentation network model, so that the global information can be effectively acquired, the detail characteristics of the optic cup image of the optic disc can be better identified, the segmentation precision of the optic disc and optic cup is improved, and the defects that the local receptive field is small, the larger receptive field can be obtained by multi-layer stacking and the information amount obtained by the convolutional neural network can be reduced along with the increase of the layer number in the traditional convolutional neural network convolution operator are effectively overcome.
Illustratively, in conjunction with FIG. 3, the U-Net network includes an encoder and a decoder, wherein: the encoder of the U-Net network is used for extracting image characteristics and comprises 4 downsampling modules, wherein each downsampling module comprises 2 convolution layers and 1 maximum pooling layer so as to carry out convolution and pooling operation on images, and the channel number of the next downsampling module is twice as large as that of the previous downsampling module except the first downsampling module. That is, the number of channels doubles per sequential downsampling. Wherein, each convolution layer adopts a convolution structure with a convolution kernel of 3×3, the step size is 1, and the padding (padding) is 0. The pool size of the maximum pooling layer may be set to 2 x 2.
As shown in fig. 3, in the U-Net network, the encoder is used to extract image features layer by layer, and the structure of the encoder includes 4 downsampling modules, each downsampling module includes 2 convolution layers of 3×3 and 1 max pooling layer of 2×2. The dimension of the output characteristic diagram is reduced by half after downsampling every time, and the dimension of the channel is doubled. For example, as shown in fig. 3, the first downsampling module converts an input image with a channel number of 1 into a feature map with a channel number of 64, the second downsampling module converts the feature map with a channel number of 64 into a feature map with a channel number of 128, the third downsampling module converts the feature map with a channel number of 128 into a feature map with a channel number of 256, and the fourth downsampling module converts the feature map with a channel number of 256 into a feature map with a channel number of 512.
The decoder of the U-Net network comprises 4 sub-modules, each sub-module comprises an up-sampling convolution layer and 2 convolution layers, and the output characteristics of each up-sampling convolution layer are respectively subjected to channel splicing with the output characteristics of the corresponding down-sampling module so as to realize characteristic fusion. That is, the decoder of the U-Net network consists of one upsampled convolutional layer and feature concatenation, and 2 convolutional layers repeated 4 times. The up-sampling convolution layer is used for recovering the original resolution of the feature map, and a convolution structure with a convolution kernel of 3×3 can be adopted, and the step size is 1.
As shown in fig. 3, in the U-Net network, the decoder is used to restore image information layer by layer, and the structure of the decoder includes 4 sub-modules symmetrical to the encoder, and each sub-module includes an up-sampling convolution layer implemented by 2×2 deconvolution, i.e., an up-convolution layer and 2 convolution layers of 3×3. The output characteristic diagram is expanded by one time in size and the channel dimension is halved after passing through one sub-module. For example, as shown in fig. 3, the feature map with 1024 channels is changed into the feature map with 512 channels after passing through the first sub-module, the feature map with 512 channels is changed into the feature map with 256 channels after passing through the second sub-module, the feature map with 256 channels is changed into the feature map with 128 channels after passing through the third sub-module, the feature map with 128 channels is changed into the feature map with 64 channels after passing through the fourth sub-module, and the feature map with 64 channels is changed into the feature map with 2-dimensional channels after passing through the convolution of 1×1.
As shown in fig. 3, in the U-Net network, the encoder and decoder can be connected by two 3×3 convolutions, wherein the first 3×3 convolution converts the channel number 512 profile output by the encoder into the channel number 1024 profile. As shown in fig. 3, in the U-Net network, the output feature map of the second convolution layer in each downsampling module of the encoder is transmitted to the decoder through jump connection, and after clipping, the output feature map of the upsampling convolution layer in the sub-module corresponding to the decoder is subjected to channel splicing, so that fusion of shallow layer information and deep layer information is realized, and more semantic information is provided for the decoding process.
Illustratively, the segformat network includes an encoder and a decoder, and in conjunction with fig. 3, the encoder of the segformat network includes 4 transform modules, i.e., the transform blocks in fig. 3. Each Transformer module includes an Efficient Self-attention (efficiency-Attn) module, a Mix-forward (Mix-Feedforward Neural Network, mix-FNN) module, and an overlapping split sub-graph merging module (i.e., the overlapping split sub-graph merging layer in fig. 3).
Specifically, as shown in fig. 3, the first transducer module may output a feature map having dimensions 128×128×32, the second transducer module may output a feature map having dimensions 64×64×160, the third transducer module may output a feature map having dimensions 32×32×160, and the fourth transducer module may output a feature map having dimensions 16×16×256. The encoder of the segfuel network may also include a separate overlapping split sub-picture merging module, i.e., the overlapping split sub-picture merging layer of fig. 3, prior to the first transducer module.
The overlapped split sub-graph merging module may be composed of a convolution kernel with a step size, for example, the convolution kernel may be a 7×7 convolution with a step size of 4 and a padding of 3, and may perform feature extraction and downsampling on the input picture.
Illustratively, the high-efficiency self-attention module is used for incoming H W images, and implements a self-attention mechanism based on the following formulas (1) and (2):
（1）
（2）
wherein R represents a scaling factor. N represents h×w, H represents the height of the image, and W represents the width of the image. K represents a feature map, i.e., a feature sequence, of the input high-efficiency self-attention module. C represents the number of channels.Representing the conversion of K into a characteristic dimension (+.>) Feature sequence, which is a feature map of x (c·r). K' represents a characteristic sequence obtained by Reshape operation of K, i.e. a characteristic dimension of (++>) Characteristic diagram of x (c·r).Linear(C·R,C) (K ') represents the conversion of the characteristic dimension C.R of K' into C by the fully connected layer. K '' represents the processing of KLinearAnd (5) a characteristic sequence obtained after operation.
That is, the efficient self-attention module first converts the NxC feature map K into [ ] through reshape operation) X (C.R) feature map K ', then converting the feature dimension C.R of K' into C through a full connection layer to obtain a feature dimension (>) The feature dimension of the feature diagram K is reduced by R times, the complexity of attention operation is reduced, and a self-attention mechanism is efficiently realized.
Illustratively, the mathematical expression of the hybrid forward module is the following formula (3):
（3）
wherein xin represents the output of the high-efficiency self-attention module;Conv3×3 denotes a convolution layer with a convolution kernel of 3×3; GELU is an activation function; MLP represents an MLP (english abbreviation of multi-layer perceptron Multilayer Perceptron) layer; xout is the output of the hybrid forward module. That is, xin passes through an MLP, then a convolution of 3×3 and a GELU activation function, and then an MLP, and the obtained result is added with xin to obtain the final output xout.
Specifically, the hybrid forward module mixes a 3×3 convolution and MLP into each feed forward neural network. In particular, the hybrid forward module may also reduce the number of parameters and increase efficiency by using a deep convolution.
Illustratively, the decoder of the segsormer network includes an upsampling MLP layer and an output MLP layer.
In particular, together with fig. 3, the up-sampling MLP layer, i.e. the MLP layer in fig. 3, is used to up-sample feature maps of different resolution sizes output by the encoder of the segvormer network, so that the feature maps are the same size and projected to a fixed dimension, such as dimension 128 x 128. The output MLP layer, i.e., the MLP in fig. 3, is used to fuse the upsampled feature map to generate a final segmentation result. For example, as shown in fig. 3, the feature map of the output MLP layer, i.e., the MLP output, may be 128×128×3 dimensions.
In order to enable a person skilled in the art to better understand the above embodiments, a specific example will be described below.
As shown in fig. 4, a method for segmenting a fundus image optic disc and a optic cup includes two stages, namely a model training stage and a model practical application stage. The model training stage includes steps S110 to S130, and the model actual application stage includes steps S140 to S150. That is, the segmentation method first establishes a optic disc cup segmentation model of the fundus image based on the U-net network and the segrormer network shown in fig. 3, and then performs optic disc cup segmentation on the fundus image based on the segmentation model, thereby obtaining a optic disc region and a optic cup region in the fundus image.
Specifically, as shown in fig. 4, the model training phase includes the following steps:
step S110, fundus photo data and a tag thereof are read.
And step S120, inputting the fundus photo data and the label thereof into a U-net region of interest extraction network, training and constructing the U-net region of interest extraction network containing the optic disc cup, and outputting a disc image containing the segmented optic disc cup.
Step S130, inputting the segmented video disc sub-image into a segvomer network, training and constructing the segvomer network, and outputting a video disc region segmentation result.
Training and constructing a fundus image cup optic disc segmentation network model based on UNet-segormer: forming a fundus image cup and optic disc segmentation network model based on UNet-segfuel by the trained U-Net network and the trained segfuel network
As shown in fig. 4, the model actual application stage includes the following steps:
step S140, reading fundus photo data to be segmented.
Step S150, inputting the fundus photo data to be segmented into the fundus image optic cup optic disc segmentation network model based on UNet-segormer, and taking a segmented image of the optic cup optic disc region of the model as an output result to obtain a segmented fundus image of the optic cup optic disc region.
Another embodiment of the present disclosure relates to a fundus image optic disc cup segmentation apparatus, as shown in fig. 5, including:
an acquisition module 501 for acquiring a fundus image to be segmented;
the segmentation module 502 is configured to input a fundus image into a pre-trained optic cup and optic disc segmentation network model, and segment a optic cup region and a optic disc region in the fundus image to obtain a corresponding segmentation result;
the training module 503 is configured to train to obtain a pre-trained optic cup and optic disc segmentation network model according to the following steps: acquiring a plurality of sample fundus images carrying labels, and dividing the plurality of sample fundus images into an image training set and an image testing set; training a preset U-Net network and a preset segvomer network respectively by utilizing an image training set to obtain a fundus image optic disc optic cup segmentation neural network model based on UNet-segvomer; and testing the fundus image optic disc and optic disc segmentation neural network model based on UNet-segrormer by using an image test set to obtain a pre-trained optic disc and optic disc segmentation network model.
The specific implementation method of the fundus image optic disc and optic cup segmentation device provided in the embodiment of the present disclosure may be described with reference to the fundus image optic disc and optic cup segmentation method provided in the embodiment of the present disclosure, and will not be described herein again.
Compared with the prior art, the embodiment of the disclosure can effectively acquire global information, better identify the detail characteristics of the video disc cup image, improve the segmentation precision of the video disc cup, and effectively overcome the defects that the local receptive field of the traditional convolutional neural network convolution operator is small, a plurality of layers of the convolutional neural network convolution operator are required to be stacked to acquire a larger receptive field, and the information amount acquired by the convolutional neural network along with the increase of the layer number is reduced.
Another embodiment of the present disclosure relates to an electronic device, as shown in fig. 6, comprising:
at least one processor 601; the method comprises the steps of,
a memory 602 communicatively coupled to the at least one processor 601; wherein,,
the memory 602 stores instructions executable by the at least one processor 601, the instructions being executable by the at least one processor 601 to enable the at least one processor 601 to perform the fundus image optic disc cup segmentation method described in the above embodiments.
Where the memory and the processor are connected by a bus, the bus may comprise any number of interconnected buses and bridges, the buses connecting the various circuits of the one or more processors and the memory together. The bus may also connect various other circuits such as peripherals, voltage regulators, and power management circuits, which are well known in the art, and therefore, will not be described any further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver may be one element or may be a plurality of elements, such as a plurality of receivers and transmitters, providing a means for communicating with various other apparatus over a transmission medium. The data processed by the processor is transmitted over the wireless medium via the antenna, which further receives the data and transmits the data to the processor.
The processor is responsible for managing the bus and general processing and may also provide various functions including timing, peripheral interfaces, voltage regulation, power management, and other control functions. And memory may be used to store data used by the processor in performing operations.
It will be understood by those of ordinary skill in the art that the foregoing embodiments are specific embodiments for carrying out the present disclosure, and that various changes in form and details may be made therein without departing from the spirit and scope of the present disclosure.
Claims (7)
1. A method for segmenting a fundus image optic disc cup, the method comprising:
acquiring a fundus image to be segmented;
inputting the fundus image into a pre-trained optic cup and optic disc segmentation network model, and segmenting a optic cup area and a optic disc area in the fundus image to obtain corresponding segmentation results;
the pre-trained optic cup and optic disc segmentation network model is obtained through training according to the following steps:
acquiring a plurality of sample fundus images carrying labels, and dividing the plurality of sample fundus images into an image training set and an image testing set;
training a preset U-Net network and a preset segfuel network respectively by utilizing the image training set to obtain a fundus image optic disc optic cup segmentation neural network model based on UNet-segfuel;
testing the fundus image optic disc and optic disc segmentation neural network model based on UNet-segormer by using the image test set to obtain the pre-trained optic disc and optic disc segmentation neural network model;
the U-Net network includes an encoder and a decoder, wherein:
the encoder of the U-Net network is used for extracting image characteristics and comprises 4 downsampling modules, wherein each downsampling module comprises 2 convolution layers and 1 maximum pooling layer, and the channel number of the next downsampling module is twice that of the previous downsampling module except the first downsampling module;
the decoder of the U-Net network comprises 4 sub-modules, each sub-module comprises an up-sampling convolution layer and 2 convolution layers, and the output characteristics of each up-sampling convolution layer are respectively subjected to channel splicing with the output characteristics of the corresponding down-sampling module so as to realize characteristic fusion;
the segfuel network includes an encoder and a decoder, wherein:
the encoder of the segfuel network comprises 4 fransformer modules, wherein each fransformer module comprises an efficient self-attention module, a hybrid forward module and an overlapped split sub-graph merging module;
the decoder of the segvomer network includes an upsampling MLP layer and an output MLP layer;
the high-efficiency self-attention module is used for inputting an H×W image, and realizes a self-attention mechanism based on the following formula (1) and the following formula (2):
（1）
（2）
wherein R represents a scaling factor; n represents H×W, H represents the height of the image, and W represents the width of the image; k represents a feature sequence input into the high-efficiency self-attention module; c represents the number of channels;representing the conversion of K into a characteristic dimension (+.>) Characteristic sequence of x (c·r); />Representing a characteristic sequence obtained after the Reshape operation is carried out on K;indicating that +.>Feature dimension +.>Conversion to C; />Representation pair->Go->The feature dimension obtained after the operation is (+)>) Feature sequence of x C.
2. The method according to claim 1, wherein training the preset U-Net network and segfuel network by using the image training set to obtain a UNet-segfuel based fundus image optic disc optic cup segmentation neural network model comprises:
performing image size adjustment processing on the sample fundus images in the image training set to obtain fundus training images carrying labels;
training the U-Net network by using the fundus training image carrying the tag to obtain a trained U-Net network;
extracting an interested region in the fundus training image by using the trained U-Net network to obtain an interested region image, wherein the interested region comprises a optic disc region and a visual cup region in the fundus training image;
performing data enhancement processing on the region-of-interest image, and training the segfuel network by using the region-of-interest image subjected to the data enhancement processing and a label corresponding to the region-of-interest image to obtain a trained segfuel network;
and forming the trained U-Net network and the trained segfuel network into the fundus image optic disc and optic cup segmentation neural network model based on the UNet-segfuel.
3. The method of claim 2, wherein training the U-Net network using the fundus training image with a tag results in a trained U-Net network, comprising:
the segmentation step: inputting the fundus training image carrying the tag into the U-Net network, and extracting the region of interest in the fundus training image to obtain a corresponding initial segmentation result;
updating: calculating a loss function value based on the initial segmentation result and the label corresponding to the initial segmentation result, and updating parameters of the U-Net network through back propagation;
repeating the dividing step and the updating step until the loss function value no longer decreases.
4. A method according to claim 3, wherein extracting the region of interest in the fundus training image using the trained U-Net network to obtain a region of interest image comprises:
inputting the fundus training image carrying the label into the trained U-Net network to obtain a video disc segmentation result;
mapping the optic disc segmentation result to the corresponding fundus training image, and cutting the fundus training image into sub-images with preset sizes based on the center of the optic disc segmentation result to obtain the region-of-interest image.
5. The method according to any one of claims 1 to 4, wherein the mathematical expression of the hybrid forward module is the following formula (3):
（3）
wherein xin represents the output of the high-efficiency self-attention module;a convolution layer representing a convolution kernel of 3×3; GELU is an activation function; MLP means MLP layer; xout is the output of the hybrid forward module.
6. A fundus image optic disc cup segmentation device, the segmentation device comprising:
the acquisition module is used for acquiring fundus images to be segmented;
the segmentation module is used for inputting the fundus image into a pre-trained optic cup and optic disc segmentation network model, and segmenting a optic cup area and a optic disc area in the fundus image to obtain corresponding segmentation results;
the training module is used for training to obtain the pre-trained video cup and optic disc segmentation network model according to the following steps: acquiring a plurality of sample fundus images carrying labels, and dividing the plurality of sample fundus images into an image training set and an image testing set; training a preset U-Net network and a preset segfuel network respectively by utilizing the image training set to obtain a fundus image optic disc optic cup segmentation neural network model based on UNet-segfuel; testing the fundus image optic disc and optic disc segmentation neural network model based on UNet-segormer by using the image test set to obtain the pre-trained optic disc and optic disc segmentation neural network model;
the U-Net network includes an encoder and a decoder, wherein:
the encoder of the U-Net network is used for extracting image characteristics and comprises 4 downsampling modules, wherein each downsampling module comprises 2 convolution layers and 1 maximum pooling layer, and the channel number of the next downsampling module is twice that of the previous downsampling module except the first downsampling module;
the decoder of the U-Net network comprises 4 sub-modules, each sub-module comprises an up-sampling convolution layer and 2 convolution layers, and the output characteristics of each up-sampling convolution layer are respectively subjected to channel splicing with the output characteristics of the corresponding down-sampling module so as to realize characteristic fusion;
the segfuel network includes an encoder and a decoder, wherein:
the encoder of the segfuel network comprises 4 fransformer modules, wherein each fransformer module comprises an efficient self-attention module, a hybrid forward module and an overlapped split sub-graph merging module;
the decoder of the segvomer network includes an upsampling MLP layer and an output MLP layer;
the high-efficiency self-attention module is used for inputting an H×W image, and realizes a self-attention mechanism based on the following formula (1) and the following formula (2):
（1）
（2）
wherein R represents a scaling factor; n represents H×W, H represents the height of the image, and W represents the width of the image; k represents a feature sequence input into the high-efficiency self-attention module; c represents the number of channels;representing the conversion of K into a characteristic dimension (+.>) Characteristic sequence of x (c·r); />Representing a characteristic sequence obtained after the Reshape operation is carried out on K;indicating that +.>Feature dimension +.>Conversion to C; />Representation pair->Go->The feature dimension obtained after the operation is (+)>) Feature sequence of x C.
7. An electronic device, comprising:
at least one processor; the method comprises the steps of,
a memory communicatively coupled to the at least one processor; wherein,,
the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the fundus image optic cup segmentation method of any of claims 1-5.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310645552.7A CN116385725B (en) | 2023-06-02 | 2023-06-02 | Fundus image optic disk and optic cup segmentation method and device and electronic equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310645552.7A CN116385725B (en) | 2023-06-02 | 2023-06-02 | Fundus image optic disk and optic cup segmentation method and device and electronic equipment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116385725A CN116385725A (en) | 2023-07-04 |
| CN116385725B true CN116385725B (en) | 2023-09-08 |
Family
ID=86961912
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310645552.7A Active CN116385725B (en) | 2023-06-02 | 2023-06-02 | Fundus image optic disk and optic cup segmentation method and device and electronic equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116385725B (en) |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110992382A (en) * | 2019-12-30 | 2020-04-10 | 四川大学 | Fundus image optic cup optic disc segmentation method and system for assisting glaucoma screening |
| CN111667490A (en) * | 2020-05-07 | 2020-09-15 | 清华大学深圳国际研究生院 | Eye fundus picture cup optic disk segmentation method |
| CN112580782A (en) * | 2020-12-14 | 2021-03-30 | 华东理工大学 | Channel enhancement-based double-attention generation countermeasure network and image generation method |
| CN112598650A (en) * | 2020-12-24 | 2021-04-02 | 苏州大学 | Combined segmentation method for optic cup optic disk in fundus medical image |
| WO2021189848A1 (en) * | 2020-09-22 | 2021-09-30 | 平安科技（深圳）有限公司 | Model training method and apparatus, cup-to-disc ratio determination method and apparatus, and device and storage medium |
| CN113850810A (en) * | 2021-12-01 | 2021-12-28 | 杭州柳叶刀机器人有限公司 | Method and surgical system for correcting femur, storage medium, and electronic device |
| CN113947680A (en) * | 2021-10-12 | 2022-01-18 | 哈尔滨理工大学 | Image semantic segmentation method based on cascade multi-scale vision Transformer |
| CN114581406A (en) * | 2022-03-04 | 2022-06-03 | 桂林医学院 | Stomach pathological section cell nucleus segmentation method based on Segformer |
| CN115619814A (en) * | 2022-10-31 | 2023-01-17 | 江西师范大学 | Method and system for jointly segmenting optic disk and optic cup |
| CN115797329A (en) * | 2022-12-30 | 2023-03-14 | 中国科学院长春光学精密机械与物理研究所 | Fundus image optic disk optic cup segmentation method based on improved Unet network |
| CN116152173A (en) * | 2022-12-22 | 2023-05-23 | 上海合合信息科技股份有限公司 | Image tampering detection positioning method and device |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3583594B1 (en) * | 2017-03-29 | 2020-09-09 | Google LLC | End-to-end text-to-speech conversion |
-
2023
- 2023-06-02 CN CN202310645552.7A patent/CN116385725B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110992382A (en) * | 2019-12-30 | 2020-04-10 | 四川大学 | Fundus image optic cup optic disc segmentation method and system for assisting glaucoma screening |
| CN111667490A (en) * | 2020-05-07 | 2020-09-15 | 清华大学深圳国际研究生院 | Eye fundus picture cup optic disk segmentation method |
| WO2021189848A1 (en) * | 2020-09-22 | 2021-09-30 | 平安科技（深圳）有限公司 | Model training method and apparatus, cup-to-disc ratio determination method and apparatus, and device and storage medium |
| CN112580782A (en) * | 2020-12-14 | 2021-03-30 | 华东理工大学 | Channel enhancement-based double-attention generation countermeasure network and image generation method |
| CN112598650A (en) * | 2020-12-24 | 2021-04-02 | 苏州大学 | Combined segmentation method for optic cup optic disk in fundus medical image |
| CN113947680A (en) * | 2021-10-12 | 2022-01-18 | 哈尔滨理工大学 | Image semantic segmentation method based on cascade multi-scale vision Transformer |
| CN113850810A (en) * | 2021-12-01 | 2021-12-28 | 杭州柳叶刀机器人有限公司 | Method and surgical system for correcting femur, storage medium, and electronic device |
| CN114581406A (en) * | 2022-03-04 | 2022-06-03 | 桂林医学院 | Stomach pathological section cell nucleus segmentation method based on Segformer |
| CN115619814A (en) * | 2022-10-31 | 2023-01-17 | 江西师范大学 | Method and system for jointly segmenting optic disk and optic cup |
| CN116152173A (en) * | 2022-12-22 | 2023-05-23 | 上海合合信息科技股份有限公司 | Image tampering detection positioning method and device |
| CN115797329A (en) * | 2022-12-30 | 2023-03-14 | 中国科学院长春光学精密机械与物理研究所 | Fundus image optic disk optic cup segmentation method based on improved Unet network |
Non-Patent Citations (1)
| Title |
|---|
| A Microcontroller is All You Need: Enabling Transformer Execution on Low-Power IoT Endnodes;Alessio Burrello等;2021 IEEE International Conference on Omni-Layer Intelligent Systems (COINS);1-6 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116385725A (en) | 2023-07-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111145170B (en) | Medical image segmentation method based on deep learning | |
| CN112132817B (en) | Retina blood vessel segmentation method for fundus image based on mixed attention mechanism | |
| CN111161273B (en) | Medical ultrasonic image segmentation method based on deep learning | |
| CN110047082B (en) | Deep learning-based pancreatic neuroendocrine tumor automatic segmentation method and system | |
| CN108764342B (en) | Semantic segmentation method for optic discs and optic cups in fundus image | |
| CN111047605B (en) | Construction method and segmentation method of vertebra CT segmentation network model | |
| CN111489328B (en) | Fundus image quality evaluation method based on blood vessel segmentation and background separation | |
| WO2022121100A1 (en) | Darts network-based multi-modal medical image fusion method | |
| CN113724206B (en) | Fundus image blood vessel segmentation method and system based on self-supervision learning | |
| CN109447976A (en) | A kind of medical image cutting method and system based on artificial intelligence | |
| CN110991254B (en) | Ultrasonic image video classification prediction method and system | |
| CN115375711A (en) | Image segmentation method of global context attention network based on multi-scale fusion | |
| CN111915594A (en) | End-to-end neural network-based breast cancer focus segmentation method | |
| CN114359310A (en) | 3D ventricle nuclear magnetic resonance video segmentation optimization system based on deep learning | |
| CN114612389A (en) | Fundus image quality evaluation method and device based on multi-source multi-scale feature fusion | |
| CN112541923A (en) | Cup optic disk segmentation method based on fundus image data set migration learning | |
| CN113763406A (en) | Infant brain MRI segmentation method based on semi-supervised learning | |
| CN113269774B (en) | Parkinson disease classification and lesion region labeling method of MRI (magnetic resonance imaging) image | |
| CN113313728B (en) | Intracranial artery segmentation method and system | |
| CN116664590B (en) | Automatic segmentation method and device based on dynamic contrast enhancement magnetic resonance image | |
| CN116385725B (en) | Fundus image optic disk and optic cup segmentation method and device and electronic equipment | |
| CN113724262A (en) | CNV segmentation method in retina OCT image | |
| CN116309507A (en) | AIS focus prediction method for performing feature fusion on CTP under attention mechanism | |
| CN110033422B (en) | Fundus OCT image fusion method and device | |
| CN115272385A (en) | Automatic label generation based cooperative cross segmentation system for cerebral hemorrhage and peripheral edema |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |