CN113450363A - Meta-learning cell nucleus segmentation system and method based on label correction - Google Patents
Meta-learning cell nucleus segmentation system and method based on label correction Download PDFInfo
- Publication number
- CN113450363A CN113450363A CN202110651067.1A CN202110651067A CN113450363A CN 113450363 A CN113450363 A CN 113450363A CN 202110651067 A CN202110651067 A CN 202110651067A CN 113450363 A CN113450363 A CN 113450363A
- Authority
- CN
- China
- Prior art keywords
- segmentation
- correction
- network
- label
- noise
- 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.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10056—Microscopic image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20112—Image segmentation details
- G06T2207/20152—Watershed segmentation
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a meta-learning cell nucleus segmentation system and a meta-learning cell nucleus segmentation method based on label correction, wherein the system comprises a cell nucleus extraction module, a segmentation correction module and a post-processing module; the method comprises the following steps: extracting all connected domains of an original pathological picture and corresponding part noise labels of the original pathological picture, and performing pixel level mask correction; for each extracted connected domain noise label and corresponding original drawing, completing the correction of a noise label through the label correction network, and supervising the training of the segmentation network; after the correction mask of each connected domain noise label is obtained, labels of overlapped cell nuclei are segmented for all the correction masks by using a watershed algorithm with identifiers, and finally, a segmentation boundary can be obtained for each cell nucleus. The network model trained by the invention can accurately identify the boundary contour of each cell nucleus, assist pathological clinical diagnosis, improve the working efficiency of pathologists, and support tasks such as research on downstream tumor microenvironment and the like.
Description
Technical Field
The invention belongs to the technical field of medical image processing and computer vision, and particularly relates to a meta-learning cell nucleus segmentation system and method based on label correction.
Background
Pathological diagnosis is the "gold standard" for cancer review. The task of nuclear segmentation is to separate all nuclei from the background from the pathology image. Accurate segmentation of all nuclei is crucial to aid in clinical pathological diagnosis. In the traditional pathological diagnosis, a pathologist needs to manually find all cell nucleuses under a microscope visual field and judge the malignancy degree of the cell nucleuses, then the cancer grade is determined according to the morphological characteristics of the cell nucleuses, such as nuclear-to-cytoplasmic ratio, area and the like, a pathological diagnosis report is given, and the prognosis of a patient is helped to be improved. A pathological section usually contains millions of nuclei and the staining of the section is uneven, which presents a great challenge to a rapid and accurate pathological diagnosis. In the digital pathological image analysis, accurate segmentation of cell nucleus also has important support effect on downstream tasks such as genotype phenotype association, prognosis analysis and the like.
The classical machine learning method is influenced by the staining quality of pathological sections, the polymorphism of cell nuclei and the like, so that a model with good generalization performance is difficult to establish through the characteristics of artificial design, and the accurate segmentation of the cell nuclei is realized. In recent years, with the development of artificial intelligence techniques typified by deep learning, deep learning models have achieved excellent segmentation performance in a plurality of cancer types. Large-scale labeled data is fuel trained on a good deep learning model. Pathological images contain complex phenotypic information such as a high number of nuclei, dense overlap, and often appear in clusters or clusters. Labeling of medical images, such as pathology images, also typically has a certain medical knowledge bottleneck. This results in the difficulty of obtaining large-scale fine-grained fine-labeled pathological data, and further results in the data labeling often containing a certain amount of noise due to low confidence among different annotators, and the like. It is important how to train a model with good segmentation performance using data with partial labeling or with noise labeling.
There are two main types of methods currently used to solve the problem of training with partial labeling or noise labeling. The first type is that extra priori knowledge is introduced, extra constraint is artificially generated, training of a supervision model is assisted, and the problem of insufficient data annotation is relieved to a certain extent. For example, a clustering label and a Thiessen polygon surveillance Bayesian network are generated based on point sparse labeling, nuclei with high uncertainty are selected from the clustering label and then delivered to a labeling person for labeling, and the labeling burden is greatly relieved. The main problem of the method is that a pseudo label needs to be generated based on external knowledge and iterative optimization is carried out, and when more noise is contained in data, a model is difficult to converge to better performance. The second is a method that employs loss weighting. The influence of noise labeling loss on model training is reduced by learning the weight matrix. For example, the importance weight is generated in the gradient direction of each pixel loss to adjust the contribution of each pixel point to model optimization, and the influence of noise labeling during model training is weakened. The limitation of this kind of method is that only the weight of the example contribution in the learning process can be increased or decreased, and there is a problem of information bottleneck, because this method weights the importance of the loss function, if the losses of different input pairs are the same, the learned importance weight cannot be effectively distinguished.
Disclosure of Invention
The invention aims to provide a meta-learning cell nucleus segmentation system and method based on label correction to overcome the defects of the prior art.
In order to achieve the purpose, the invention adopts the following technical scheme:
a meta-learning cell nucleus segmentation system based on label correction comprises a cell nucleus extraction module, a segmentation correction module and a post-processing module;
the cell nucleus extraction module trains a segmentation model in advance according to the part marks or the noise marks to obtain an initial segmentation result, extracts each communication area in the initial segmentation result and inputs the communication area to the next module;
the segmentation correction module trains a segmentation network and a label correction network under the assistance of five hundreds of fine-grained precision labels of a total sample according to the noise labels and corresponding pictures to finish the correction of the noise cell nucleus labels; inputting a pathological picture by a segmentation network to generate a segmentation mask characteristic map, inputting the segmentation characteristic map by a label correction network and generating a correction label by a noise label to supervise the training of the segmentation network;
the post-processing module separates the overlapping nuclear masks using an identifier-based watershed algorithm based on the results output by the segmentation correction module.
A meta-learning cell nucleus segmentation method based on label correction comprises the following steps:
the method comprises the following steps:
for an original pathological picture and corresponding part noise labels thereof, training a segmentation network by using the part noise label data, preliminarily predicting the original pathological picture by the segmentation network, extracting all connected domains according to an initial prediction mask, and then correcting a pixel-level mask;
step two:
for each connected domain noise label and corresponding original image extracted in the step one, a label correction network is designed in a full convolution mode by adopting a meta-learning thought based on label correction, the correction of the noise label is completed through the label correction network under the assistance of five hundreds of labels of a total sample, and the training of a segmentation network is supervised;
step three:
and secondly, after the correction mask of each connected domain noise label is obtained, segmenting the labels of the overlapped cell nuclei by using a watershed algorithm with identifiers for all the correction masks, and finally obtaining a segmentation boundary for each cell nucleus.
The invention is further improved in that, in the second step,
with D ═ x, y }mRepresenting a small number of clean data samples, D '═ x, y' }MRepresenting data samples containing noise labels; where M, M denotes the number of noise and clean samples, y' denotes the noise sample label, the segmentation network is parameterized as a function with a parameter W, y ═ fW(x) The tag correction network is formalized as a function with a parameter θ, yc=gθ(h, (x), y'); wherein h (x) represents a feature map of the output of the segmented network, ycA tag indicating a tag corrected by the tag correction network; training a segmentation network and a label correction network as a bidirectional optimization process; by minimizing the equation (1) when the tag correction network parameter θ is fixedObtaining an optimal W parameter by the objective function;
by means of a metadata set containing hundreds of five clean samples of total samples and determined optimal parameters of a segmentation networkOptimizing the formula (2) objective function to obtain an optimal theta value;
the invention further improves the method that the values of W and theta are alternately updated by adopting a bidirectional optimization method in the actual training process.
A further improvement of the invention is that the values of W and θ are updated alternately in an iterative manner during one cycle in order to obtain optimal split network and tag corrected network parameters.
The invention is further improved in that a cycle optimization algorithm process comprises the following steps:
1. initializing split network parameters W(0)And tag correction network parameter θ(0);
2. In the process of the t-th iteration, the parameters of the segmentation network are temporarily updated according to the formula (3), and the objective function formula (1) is minimized through one-time gradient update, wherein the calculation formula of the loss at the t moment is not shown in the formula (4);
3. updating the label correction network parameter theta by minimizing the objective function (2), wherein the updating process is as shown in formula (5), and the loss calculation formula at the moment of t +1 is as shown in formula (6);
4. finally, the updating of the segmentation network parameters W is completed by minimizing the segmentation network objective function (1); wherein, the loss calculation formula at the time of t +1 is shown as (8);
compared with the prior art, the invention has at least the following beneficial technical effects:
1. a new meta-learning framework is provided based on the idea of label correction, a label correction network maps a data characteristic graph and a noise label into a correction label, and under the pixel level granularity, the influence of noise labeling is effectively inhibited, and the segmentation network training is assisted.
2. The load of fine-grained fine marking data is greatly relieved, a marker only needs to add partial marking, and the marked boundary does not need to be too accurate, and the segmentation effect of the supervision scene can be achieved by using the proposed method.
3. The bottleneck and threshold of pathological section labeling are reduced, and a labeling person can finish the task of data labeling without having professional medical knowledge background like a pathologist.
4. The network model trained by the invention can accurately identify the boundary contour of each cell nucleus, assist pathological clinical diagnosis, improve the working efficiency of pathologists, and support tasks such as research on downstream tumor microenvironment and the like.
Drawings
Fig. 1 is a block diagram of a partitioning system, which includes three modules: the device comprises a cell nucleus extraction module, a segmentation correction module and a post-processing module.
Fig. 2 is a specific structure and optimization process of the segmentation correction module, which mainly includes two parts, namely a segmentation network and a label correction network.
Fig. 3 is a structure of a segmented network, including five convolutional layers and residual connection.
FIG. 4 is a structure of a label correction network, including two convolutional layers and one activation function layer.
Detailed Description
Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art. It should be noted that the embodiments and features of the embodiments may be combined with each other without conflict. The present invention will be described in detail below with reference to the embodiments with reference to the attached drawings.
As shown in FIG. 1, the invention provides a meta-learning cell nucleus segmentation system based on label correction, which comprises a cell nucleus extraction module, a segmentation correction module and a post-processing module.
The cell nucleus extraction module trains a segmentation model in advance according to the part marks or the noise marks to obtain initial segmentation results, extracts each connected region in the initial segmentation results, and inputs the extracted connected region to the next module; the segmentation correction module trains a segmentation network and a label correction network under the assistance of five hundreds of fine-grained precision labels of a total sample according to the noise labels and corresponding pictures to finish the correction of the noise cell nucleus labels; the segmentation network inputs pathological pictures to generate a segmentation mask feature map, the segmentation network structure is shown in fig. 3 and mainly comprises five layers of convolution structures, and feature fusion is carried out between different layers through residual connection. The label correction network inputs the segmentation characteristic diagram and the noise label to generate the training of the correction label supervision segmentation network, and the structure of the label correction network is shown in figure 4 and mainly comprises two convolution layers and a final activation function layer;
the post-processing module separates the overlapping nuclear masks using an identifier-based watershed algorithm based on the results output by the segmentation correction module.
The invention provides a meta-learning cell nucleus segmentation method based on label correction, which comprises the following specific implementation steps of:
the method comprises the following steps:
for an original pathological picture and corresponding part noise labels thereof, training a segmentation network by using the part noise label data, preliminarily predicting the original pathological picture by the network, extracting all connected domains according to an initial prediction mask, and then putting the extracted connected domains into a segmentation correction module for pixel-level mask correction;
step two:
and (3) for the noise label and the corresponding original image of each connected domain extracted in the step one, designing a label correction network in a full convolution mode by adopting a meta-learning thought based on label correction, completing the correction of the noise label through the label correction network under the assistance of five hundreds of labels of a total sample, and supervising the training of the segmentation network.
With D ═ x, y }mRepresenting a small number of clean data samples, D '═ x, y' }MRepresenting data samples containing noise labels; where M, M denotes the number of noise and clean samples, y' denotes the noise sample label, the segmentation network is parameterized as a function with a parameter W, y ═ fW(x) The tag correction network is formalized as a function with a parameter θ, yc=gθ(h, (x), y'); wherein h (x) represents a feature map of the output of the segmented network, ycA tag indicating a tag corrected by the tag correction network; training a segmentation network and a label correction network as a bidirectional optimization process; when the tag correction network parameter theta is fixedThe optimal W parameter is obtained by minimizing the objective function of equation (1).
By means of a metadata set containing hundreds of five clean samples of total samples and determined optimal parameters of a segmentation networkAnd (3) optimizing the formula (2) objective function to obtain an optimal theta value.
In the actual training process, the values of W and theta are alternately updated by adopting a bidirectional optimization method.
In order to obtain the optimal parameters of the segmentation network and the label correction network, the values of W and theta are alternately updated in an iterative mode in the process of one cycle. The one-time loop optimization algorithm process mainly comprises the following steps:
1. initializing split network parameters W(0)And tag correction network parameter θ(0)。
2. During the t-th iteration, the parameters of the segmentation network are temporarily updated according to formula (3), and the objective function formula (1) is minimized through one gradient update. The formula for calculating the loss at time t is shown in (4).
3. The tag corrected network parameter θ can be updated by minimizing the objective function (2), the updating process being as in equation (5). The loss calculation formula at time t +1 is shown in (6).
4. Finally, the updating of the split network parameters W is done by minimizing the split network objective function (1).
The loss calculation formula at time t +1 is shown in (8).
The once circulation whole optimization process is shown in fig. 2, firstly, inputting an image into a current segmentation network, then calculating a logic stethogram of a prediction result, secondly, inputting a part of noise mask and the logic stethogram of the prediction result into a label correction network to obtain a correction mask, thirdly, calculating the loss of the logic stethogram and the correction mask, then calculating the loss gradient of parameters related to the segmentation network, thirdly, updating the parameters of the segmentation network under the condition of keeping the gradient map, fifthly, inputting a pair of masks and pictures marked with fine granularity into a new segmentation network, calculating the loss, sixthly, calculating the gradient of the loss, and updating the parameters of the label correction network.
Step three:
and step two, after the correction masks of all the noise labels are obtained, a watershed algorithm with identifiers is adopted for all the labels, labels of overlapped cell nuclei are segmented, and finally a better segmentation boundary can be obtained for each cell nucleus.
Examples
This example combines the methods presented above to segment cancer cells in pathological sections of renal clear cell carcinoma. The method comprises the following steps:
(1) and generating a data set. And cutting the image blocks and the mask blocks with the sizes of 64-64 according to the segmentation mask of the pathological section to be used as a construction data set. The number of training sets is 2000, the number of test sets is 1500, and the number of metadata sets is 50.
(2) A noise mask is generated. In the embodiment, two noise generation modes are adopted, wherein the first mode is partial gold standard marking, a 40% proportion gold standard mask is randomly selected for reservation, and the rest marking masks are deleted; and the second method is partial weak labeling, and on the basis of partial gold standard labeling, the gold standard is subjected to random one-to-three pixel expansion operation or expanded into a box.
(3) Implementation details. In this embodiment, a pytorech frame is adopted for implementation, a Resnet-32 structure is adopted as a feature extractor for the partition network, and a U-shaped structure is adopted for the network structure. The label correction network uses a layer of 3 by 3 convolution and a layer of 1 by 1 convolution. In training, the learning rate of the segmented network is set to 1 × 10-3After 300 training generations, dropping 0.1, the learning rate of the label correction network is set to 1x10-4Adam is used as an optimizer for both networks.
(4) And outputting the result. After training is finished, the original image is input into the segmentation network, the segmentation result of the cell nucleus can be output, the label correction network only assists in the training of the segmentation model in the training process, the experimental result is shown in table 1, and the method even achieves the performance of a supervision scene under the scene of swelling noise and part of gold standard labeling. Under three noise scenes, compared with the method for training by directly using noise data, the performance of the method is greatly improved.
Table 1 shows the results of experiments performed in the examples using the proposed segmentation method.
Although the invention has been described in detail hereinabove with respect to a general description and specific embodiments thereof, it will be apparent to those skilled in the art that modifications or improvements may be made thereto based on the invention. Accordingly, such modifications and improvements are intended to be within the scope of the invention as claimed.
Claims (6)
1. A meta-learning cell nucleus segmentation system based on label correction is characterized by comprising a cell nucleus extraction module, a segmentation correction module and a post-processing module;
the cell nucleus extraction module trains a segmentation model in advance according to the part marks or the noise marks to obtain an initial segmentation result, extracts each communication area in the initial segmentation result and inputs the communication area to the next module;
the segmentation correction module trains a segmentation network and a label correction network under the assistance of five hundreds of fine-grained precision labels of a total sample according to the noise labels and corresponding pictures to finish the correction of the noise cell nucleus labels; inputting a pathological picture by a segmentation network to generate a segmentation mask characteristic map, inputting the segmentation characteristic map by a label correction network and generating a correction label by a noise label to supervise the training of the segmentation network;
the post-processing module separates the overlapping nuclear masks using an identifier-based watershed algorithm based on the results output by the segmentation correction module.
2. A meta-learning cell nucleus segmentation method based on label correction is characterized by comprising the following steps:
the method comprises the following steps:
for an original pathological picture and corresponding part noise labels thereof, training a segmentation network by using the part noise label data, preliminarily predicting the original pathological picture by the segmentation network, extracting all connected domains according to an initial prediction mask, and then correcting a pixel-level mask;
step two:
for each connected domain noise label and corresponding original image extracted in the step one, a label correction network is designed in a full convolution mode by adopting a meta-learning thought based on label correction, the correction of the noise label is completed through the label correction network under the assistance of five hundreds of labels of a total sample, and the training of a segmentation network is supervised;
step three:
and secondly, after the correction mask of each connected domain noise label is obtained, segmenting the labels of the overlapped cell nuclei by using a watershed algorithm with identifiers for all the correction masks, and finally obtaining a segmentation boundary for each cell nucleus.
3. The method as claimed in claim 2, wherein in step two,
with D ═ x, y }mRepresenting a small number of clean data samples, D '═ x, y' }MRepresenting data samples containing noise labels; where M, M denotes the number of noise and clean samples, y' denotes the noise sample label, the segmentation network is parameterized as a function with a parameter W, y ═ fW(x) The tag correction network is formalized as a function with a parameter θ, yc=gθ(h, (x), y'); wherein h (x) represents a feature map of the output of the segmented network, ycA tag indicating a tag corrected by the tag correction network; training a segmentation network and a label correction network as a bidirectional optimization process; when the label correction network parameter theta is fixed, obtaining an optimal W parameter by minimizing the objective function of the formula (1);
by means of a metadata set containing hundreds of five clean samples of total samples and determined optimal parameters of a segmentation networkOptimizing the formula (2) objective function to obtain an optimal theta value;
4. the method as claimed in claim 3, wherein the values of W and θ are updated alternately during the actual training process by using a bi-directional optimization method.
5. The method as claimed in claim 4, wherein in order to obtain the optimal segmentation network and label correction network parameters, the values of W and θ are updated alternately in an iterative manner during one cycle.
6. The method as claimed in claim 5, wherein the one-pass optimization algorithm comprises the following steps:
1. initializing split network parameters W(0)And tag correction network parameter θ(0);
2. In the process of the t-th iteration, the parameters of the segmentation network are temporarily updated according to the formula (3), and the objective function formula (1) is minimized through one-time gradient update, wherein the calculation formula of the loss at the t moment is not shown in the formula (4);
3. updating the label correction network parameter theta by minimizing the objective function (2), wherein the updating process is as shown in formula (5), and the loss calculation formula at the moment of t +1 is as shown in formula (6);
4. finally, the updating of the segmentation network parameters W is completed by minimizing the segmentation network objective function (1); wherein, the loss calculation formula at the time of t +1 is shown as (8);
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110651067.1A CN113450363B (en) | 2021-06-10 | 2021-06-10 | Meta-learning cell nucleus segmentation system and method based on label correction |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110651067.1A CN113450363B (en) | 2021-06-10 | 2021-06-10 | Meta-learning cell nucleus segmentation system and method based on label correction |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113450363A true CN113450363A (en) | 2021-09-28 |
CN113450363B CN113450363B (en) | 2023-05-02 |
Family
ID=77811177
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110651067.1A Active CN113450363B (en) | 2021-06-10 | 2021-06-10 | Meta-learning cell nucleus segmentation system and method based on label correction |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113450363B (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113962999A (en) * | 2021-10-19 | 2022-01-21 | 浙江大学 | Noise label segmentation method based on Gaussian mixture model and label correction model |
CN114511552A (en) * | 2022-02-24 | 2022-05-17 | 中南大学湘雅医院 | Thyroid ultrasound nodule fuzzy boundary-oriented segmentation method based on meta-learning |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150169985A1 (en) * | 2013-12-12 | 2015-06-18 | University Of Pittsburgh - Of The Commonwealth System Of Higher Education | Automated epithelial nuclei segmentation for computational disease detection algorithms |
CN109493330A (en) * | 2018-11-06 | 2019-03-19 | 电子科技大学 | A kind of nucleus example dividing method based on multi-task learning |
US20190147215A1 (en) * | 2017-11-16 | 2019-05-16 | General Electric Company | System and method for single channel whole cell segmentation |
CN111091571A (en) * | 2019-12-12 | 2020-05-01 | 珠海圣美生物诊断技术有限公司 | Nucleus segmentation method and device, electronic equipment and computer-readable storage medium |
CN112102323A (en) * | 2020-09-17 | 2020-12-18 | 陕西师范大学 | Adherent nucleus segmentation method based on generation of countermeasure network and Caps-Unet network |
CN112132843A (en) * | 2020-09-30 | 2020-12-25 | 福建师范大学 | Hematoxylin-eosin staining pathological image segmentation method based on unsupervised deep learning |
-
2021
- 2021-06-10 CN CN202110651067.1A patent/CN113450363B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150169985A1 (en) * | 2013-12-12 | 2015-06-18 | University Of Pittsburgh - Of The Commonwealth System Of Higher Education | Automated epithelial nuclei segmentation for computational disease detection algorithms |
US20190147215A1 (en) * | 2017-11-16 | 2019-05-16 | General Electric Company | System and method for single channel whole cell segmentation |
CN109493330A (en) * | 2018-11-06 | 2019-03-19 | 电子科技大学 | A kind of nucleus example dividing method based on multi-task learning |
CN111091571A (en) * | 2019-12-12 | 2020-05-01 | 珠海圣美生物诊断技术有限公司 | Nucleus segmentation method and device, electronic equipment and computer-readable storage medium |
CN112102323A (en) * | 2020-09-17 | 2020-12-18 | 陕西师范大学 | Adherent nucleus segmentation method based on generation of countermeasure network and Caps-Unet network |
CN112132843A (en) * | 2020-09-30 | 2020-12-25 | 福建师范大学 | Hematoxylin-eosin staining pathological image segmentation method based on unsupervised deep learning |
Non-Patent Citations (3)
Title |
---|
GUOQING ZHENG等: "META LABEL CORRECTION FOR LEARNING WITH WEAK SUPERVISION", 《ARXIV-MACHINE LEARNING》 * |
JIXIN WANG等: "Meta Corrupted Pixels Mining for Medical Image Segmentation", 《ARXIV-COMPUTER VISION AND PATTERN RECOGNITION》 * |
宁静艳: "基于深度学习的细胞核图像分割方法研究", 《医药卫生科技辑》 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113962999A (en) * | 2021-10-19 | 2022-01-21 | 浙江大学 | Noise label segmentation method based on Gaussian mixture model and label correction model |
CN114511552A (en) * | 2022-02-24 | 2022-05-17 | 中南大学湘雅医院 | Thyroid ultrasound nodule fuzzy boundary-oriented segmentation method based on meta-learning |
Also Published As
Publication number | Publication date |
---|---|
CN113450363B (en) | 2023-05-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107256558A (en) | The cervical cell image automatic segmentation method and system of a kind of unsupervised formula | |
CN110706214B (en) | Three-dimensional U-Net brain tumor segmentation method fusing condition randomness and residual error | |
CN113450363B (en) | Meta-learning cell nucleus segmentation system and method based on label correction | |
CN113111916B (en) | Medical image semantic segmentation method and system based on weak supervision | |
CN109102498B (en) | Method for segmenting cluster type cell nucleus in cervical smear image | |
CN111784564B (en) | Automatic image matting method and system | |
CN111882620A (en) | Road drivable area segmentation method based on multi-scale information | |
CN110008992B (en) | Deep learning method for prostate cancer auxiliary diagnosis | |
CN110349170B (en) | Full-connection CRF cascade FCN and K mean brain tumor segmentation algorithm | |
CN114972254A (en) | Cervical cell image segmentation method based on convolutional neural network | |
CN113160229A (en) | Pancreas segmentation method and device based on hierarchical supervision cascade pyramid network | |
CN115601330A (en) | Colonic polyp segmentation method based on multi-scale space reverse attention mechanism | |
CN112837338A (en) | Semi-supervised medical image segmentation method based on generation countermeasure network | |
CN116363149A (en) | Medical image segmentation method based on U-Net improvement | |
CN113538422B (en) | Pathological image automatic classification method based on dyeing intensity matrix | |
CN112508860B (en) | Artificial intelligence interpretation method and system for positive check of immunohistochemical image | |
CN117437423A (en) | Weak supervision medical image segmentation method and device based on SAM collaborative learning and cross-layer feature aggregation enhancement | |
CN111612803B (en) | Vehicle image semantic segmentation method based on image definition | |
CN113222903A (en) | Full-section histopathology image analysis method and system | |
CN117649657A (en) | Bone marrow cell detection system based on improved Mask R-CNN | |
CN116386803A (en) | Cytopathology report generation method based on graph | |
Shi et al. | Meta mask correction for nuclei segmentation in histopathological image | |
CN113409321B (en) | Cell nucleus image segmentation method based on pixel classification and distance regression | |
CN113888551A (en) | Liver tumor image segmentation method based on dense connection network of high-low layer feature fusion | |
Cheng et al. | Deep adversarial image synthesis for nuclei segmentation of histopathology image |
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 |