CN110223291A - A kind of training retinopathy height segmentation network method based on loss function - Google Patents
A kind of training retinopathy height segmentation network method based on loss function Download PDFInfo
- Publication number
- CN110223291A CN110223291A CN201910534317.6A CN201910534317A CN110223291A CN 110223291 A CN110223291 A CN 110223291A CN 201910534317 A CN201910534317 A CN 201910534317A CN 110223291 A CN110223291 A CN 110223291A
- Authority
- CN
- China
- Prior art keywords
- function
- segmentation
- training
- network
- sample
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- 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
- G06T2207/20084—Artificial neural networks [ANN]
-
- 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
- G06T2207/20132—Image cropping
-
- 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
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30041—Eye; Retina; Ophthalmic
-
- 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
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30096—Tumor; Lesion
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Radiology & Medical Imaging (AREA)
- Quality & Reliability (AREA)
- Eye Examination Apparatus (AREA)
Abstract
The present invention discloses a kind of training retinopathy height segmentation network method based on loss function, trains depth segmentation network, using loss function efficiently to be divided to eyeground pathological changes point.According to indicator function as a result, judging that negative sample is retained or abandons, indicator function value is 1 and retains negative sample, on the contrary then abandon negative sample.The discriminating power and learning rate of network are improved with this, wherein easily divide sample with higher probability dropping, difficulty divides sample with lower probability dropping;In the case where reservation difficulty divides sample, a large amount of samples selection time can be saved, so that network, which concentrates on hardly possible, divides sample in study.The present invention can solve the problems, such as segmentation network caused by class balanced, crossover entropy loss function, and accidentally point situation is mostly lower with learning efficiency, is efficiently divided to eyeground pathological changes point.
Description
Technical field
The invention belongs to nerual network technique field, in particular to a kind of training retinopathy height based on loss function point
Cut network method.
Background technique
Depth convolutional neural networks are as a kind of deep learning model, in image classification, target detection and Target Segmentation etc.
State-of-the-art performance is all achieved in many Computer Vision Tasks.In recent years, the semantic segmentation model based on deep learning obtains
Extensive research has been arrived, significant effect is achieved.However, as far as we know, most of existing models all concentrate on normally
On the object of size, such as animal and vehicle.The semantic segmentation of wisp is adequately studied not yet.Such as medicine neck
Retinopathy height test problems in domain.However, segmentation minor illness height is different from the object of segmentation normal size.Small object segmentation
In be constantly present class imbalance problem, this is very common in medical image, for example, the lesion point ratio in eye fundus image can be with
Down to 0.1%.The loss function made in big object segmentation can not be applicable in by this extreme imbalance in wisp segmentation, because
To be easy to for all pixels to be classified as background, and obtain meaningless 99.9% accuracy.One intuitive solution classification
The method of imbalance problem is that different weights is distributed for different classifications, and the method is known as class balanced, crossover entropy loss by us
Function.The pixel of small scale is endowed high weight, and a high proportion of pixel is endowed low weight.But this method does not consider sample
This weight, all negative samples are treated equally, and share identical weight.Therefore, it in class balanced, crossover entropy loss, bears
Sample is often mistakenly classified as positive sample, because the loss of misclassification background pixel is more much smaller than the loss of misclassification positive sample.
For above-mentioned problem, currently no effective solution has been proposed.
Summary of the invention
The present invention for the technical problems in the prior art, provides a kind of training eyeground pathological changes based on loss function
Point segmentation network method, can solve caused by class balanced, crossover entropy loss function segmentation network accidentally point situation mostly with learning efficiency compared with
Low problem efficiently divides eyeground pathological changes point.
In order to solve the above technical problems, the technical solution adopted by the present invention is that: a kind of training eyeground based on loss function
Lesion point divides network method, comprising the following steps:
Step 1: the pretreatment eyeground IDRiD data set chooses a certain number of eye fundus images as training set and survey respectively
Examination collection;With certain resolution ratio to every image down sampling;Data enhancing is carried out to training set;The super ginseng of setting segmentation network
Number;
Step 2: the weight of initialization segmentation each layer of network;
Step 3: selecting an eye fundus image at random from the training set after expansion;And it is cut out one at random from image
The region of scale cun;
Step 4: eye fundus image obtains the input of loss function layer through the processing module in over-segmentation network, i.e. segmentation is general
Rate figure p and corresponding mark y;
Step 5: according to setting, selecting a kind of discarding function;
Step 6: to each of image negative sample (background pixel);According to indicator function as a result, determining its quilt
Retain or abandons;
Step 7: weight factor β is calculated,
Wherein | Y+| be positive number of samples, i.e. segmentation sampled pixel number, | Y-| for the negative sample retained
This number, i.e. background pixel number;
Step 8: propagated forward loss is calculated,
Loss function is as follows:
Wherein β be retain negative sample number divided by the sum of itself and positive sample number, calculated using formula in step 7
It arrives;
Step 9: to each of image sample, calculating respective gradient information;
Step 10: the gradient backpropagation of loss layer updates the weighting parameter in segmentation network characterization processing module;
Step 11: if network is still not converged, or not up to maximum number of iterations, return step 3;
Step 12: after network training terminates, aneurysms segmentation being carried out to eye fundus image on test set, according to segmentation
As a result, calculating PR curve.
Preferably, carrying out data enhancing by the way of rotation, mirror image to training set in step 1.
Preferably, abandoning function will activate probability to be mapped to drop probability, and loss function is strong according to abandoning in step 5
The difference of degree and calculating cost is as follows respectively there are three types of discarding function:
It is linear to abandon function: pdrop(pj)=1.0-pj;
Square abandon function: pdrop(pj)=(1.0-pj)2;
Logarithm abandons function: pdrop(pj)=1.0+log (1.0-pj)
For the negative sample of discarding, its loss contributed is 0.
Preferably, indicator function is as follows in step 6,
Wherein, 0 r, the random number between 1;
pdrop(pj) it is to abandon function, retain negative sample for 1, it is on the contrary then abandon negative sample.
Preferably, gradient information calculation method is as follows in step 9:
Initialize the gradient of all samples;
To a certain sample i, gradient giIt is initialized as: pi-yi, wherein piTo activate probability value, yiFor mark;
For positive sample, gradient updating are as follows: gi=β × gi;
For the negative sample of reservation, gradient updating are as follows: gi=(1.0- β) × gi;
For the negative sample of discarding, gradient is set as 0.
Compared with prior art, the present invention has the beneficial effects that when the present invention solves segmentation retinopathy height
The unbalanced problem of classification, alleviate to a certain extent lesion point segmentation network kind misrecognition problem, accelerate network
Learning rate, and a variety of deep learning models can be acted on.
Detailed description of the invention
Fig. 1 is segmentation network training and test flow chart;
Fig. 2 is a kind of optional retinopathy height segmentation network diagram according to an embodiment of the present invention;
Fig. 3 is three kinds of discarding function schematic diagrames;
Fig. 4 is the PR curve comparison figure of the arterioles of fundus tumor segmentation result of distinct methods;
Fig. 5 is the PR curve comparison figure of the fundus hemorrhage point segmentation result of distinct methods;
Fig. 6 a is expert's mark figure of arterioles of fundus tumor;
Fig. 6 b is the corresponding segmentation figure of class balanced, crossover entropy loss of arterioles of fundus tumor;
Fig. 6 c is the corresponding segmentation figure of new loss function of arterioles of fundus tumor;
Fig. 7 a is expert's mark figure of fundus hemorrhage point;
Fig. 7 b is the corresponding segmentation figure of class balanced, crossover entropy loss of fundus hemorrhage point;
Fig. 7 c is the corresponding segmentation figure of new loss function of fundus hemorrhage point.
Specific embodiment
Technical solution in order to enable those skilled in the art to better understand the present invention, in the following with reference to the drawings and specific embodiments
It elaborates to the present invention.
Embodiment 1
In the relevant technologies, such as class balanced, crossover entropy loss function, the pixel of small scale is endowed high weight, a high proportion of
Pixel is endowed low weight.But this method does not consider the weight between sample, all negative samples are treated equally, and are shared
Identical weight.Therefore, in class balanced, crossover entropy loss, negative sample is often mistakenly classified as positive sample.In addition, this loss
The training effectiveness of function is not high.Therefore a kind of new loss function is proposed in order to solve the technical problem in the present embodiment
Train the lesion point to divide network so that network concentrate on it is difficult divide sample in study, improve network learning efficiency and
Anti-interference.
Before introducing the technical solution of the embodiment of the present invention, it is necessary first to the lesion used in the embodiment of the present invention
Point segmentation network is defined, and in embodiments of the present invention, divides network as shown in Figure 1 for a kind of optional lesion point, by
Three parts composition: input module, processing module and output module.Input module pre-processes training data, for example contracts
It puts, expand.Processing module includes convolution operation and pondization operation etc. in neural network.Output module divides lesion point
As a result it is saved and is visualized.
In eye fundus image aneurysms segmentation task, the segmentation network in loss function training Fig. 1, specific method are used
The following steps are included:
Step 1: the pretreatment eyeground IDRiD data set, 54 eye fundus images are as training set, and 27 images are as test
Collection.Since the size of image is too big, computer hardware is difficult to handle, and therefore, the resolution ratio of every image is down sampled to 1440
×960.Using rotation, the modes such as mirror image carry out data enhancing to training set.The hyper parameter of setting segmentation network;
Step 2: the weight of initialization segmentation each layer of network;
Step 3: selecting an eye fundus image at random from the training set after expansion.And it is cut out at random from image
800 × 800 region;
Step 4: eye fundus image obtains the input of loss function layer through the processing module in over-segmentation network, i.e. segmentation is general
Rate figure p and corresponding mark y;
Step 5: according to setting, abandoning function from linear, square discarding function and logarithm abandon in function and select one kind.Three
The schematic diagram of kind function is as shown in Figure 2;
Step 6: to each of image negative sample (background pixel), according to indicator function as a result, determining its quilt
Retaining or abandons, indicator function is as follows,
Wherein, 0 r, the random number between 1, pdrop(pj) it is to abandon function, the effect for abandoning function is will to activate probability
It is mapped to drop probability, loss function provides three kinds of discarding functions according to the difference for abandoning intensity and calculating cost, fixed respectively
Justice is as follows:
It is linear to abandon function: pdrop(pj)=1.0-pj;
Square abandon function: pdrop(pj)=(1.0-pj)2;
Logarithm abandons function: pdrop(pj)=1.0+log (1.0-pj)
For the negative sample of discarding, its loss contributed is 0;
Step 7: weight factor β is calculated,
Wherein | Y+| it is positive sample (aneurysms pixel) number, | Y-| for the negative sample (back retained
Scene element) number.
Step 8: propagated forward loss is calculated,
Loss function is as follows:
Wherein β be retain negative sample number divided by the sum of itself and positive sample number, calculated using formula in step 7
It arrives.
Step 9: to each of image sample, calculating respective gradient information;
Step 9.1: initializing the gradient of all samples.
To a certain sample i, gradient giIt is initialized as: pi-yi, wherein piTo activate probability value, yiFor mark;
Step 9.2: for positive sample, gradient updating are as follows: gi=β × gi;
Step 9.3: for the negative sample of reservation, gradient updating are as follows: gi=(1.0- β) × gi;
Step 9.4: for the negative sample of discarding, gradient is set as 0.
Step 10: the gradient backpropagation of loss layer updates the weighting parameter in segmentation network characterization processing module;
Step 11: if network is still not converged, or not up to maximum number of iterations, return step 3;
Step 12: after network training terminates, aneurysms segmentation being carried out to eye fundus image on test set, according to segmentation
As a result, calculating PR curve, as a result as shown in Figure 4.The segmentation probability graph of aneurysms is as fig. 6 c on test set.
Table 1: aneurysms segmentation effect comparing result
It is in table 1 the result shows that, loss function compared with prior art in the segmentation of aneurysms there are clear superiority,
The effect of three kinds of discarding functions is superior to class balanced, crossover entropy loss function.
Embodiment 2
In eye fundus image blutpunkte segmentation task, the segmentation network in loss function training Fig. 1, specific method packet are used
Include following steps:
Step 1: the pretreatment eyeground IDRiD data set, 54 eye fundus images are as training set, and 27 images are as test
Collection.Since the size of image is too big, computer hardware is difficult to handle, and therefore, the resolution ratio of every image is down sampled to 1440
×960.Using rotation, the modes such as mirror image carry out data enhancing to training set.The hyper parameter of setting segmentation network;
Step 2: the weight of initialization segmentation each layer of network;
Step 3: selecting an eye fundus image at random from the training set after expansion.And it is cut out at random from image
800 × 800 region;
Step 4: eye fundus image obtains the input of loss function layer through the processing module in over-segmentation network, i.e. segmentation is general
Rate figure p and corresponding mark y;
Step 5: according to setting, abandoning function from linear, square discarding function and logarithm abandon in function and select one kind.Three
The schematic diagram of kind function is as shown in Figure 3;
Step 6: to each of image negative sample (background pixel), according to indicator function as a result, determining its quilt
Retaining or abandons, indicator function is as follows,
Wherein, 0 r, the random number between 1, pdrop(pj) it is to abandon function, the effect for abandoning function is will to activate probability
It is mapped to drop probability, loss function provides three kinds of discarding functions according to the difference for abandoning intensity and calculating cost, fixed respectively
Justice is as follows:
It is linear to abandon function: pdrop(pj)=1.0-pj;
Square abandon function: pdrop(pj)=(1.0-pj)2;
Logarithm abandons function: pdrop(pj)=1.0+log (1.0-pj)
For the negative sample of discarding, its loss contributed is 0;
Step 7: weight factor β is calculated,
Wherein | Y+| it is positive sample (blutpunkte pixel) number, | Y-| for the negative sample (background retained
Pixel) number.
Step 8: propagated forward loss is calculated,
Loss function is as follows:
Wherein β be retain negative sample number divided by the sum of itself and positive sample number, calculated using formula in step 7
It arrives.
Step 9: to each of image sample, calculating respective gradient information;
Step 9.1: initializing the gradient of all samples.
To a certain sample i, gradient giIt is initialized as: pi-yi, wherein piTo activate probability value, yiFor mark;
Step 9.2: for positive sample, gradient updating are as follows: gi=β × gi;
Step 9.3: for the negative sample of reservation, gradient updating are as follows: gi=(1.0- β) × gi;
Step 9.4: for the negative sample of discarding, gradient is set as 0.
Step 10: the gradient backpropagation of loss layer updates the weighting parameter in segmentation network characterization processing module;
Step 11: if network is still not converged, or not up to maximum number of iterations, return step 3;
Step 12: after network training terminates, blutpunkte segmentation being carried out to eye fundus image on test set, is tied according to segmentation
Fruit calculates PR curve, as a result as shown in Figure 5.The segmentation probability graph of blutpunkte is as shown in Figure 7 c on test set.
Table 2: blutpunkte segmentation effect comparing result
It is in table 2 the result shows that, loss function is compared with prior art in the segmentation of blutpunkte there are clear superiority, three
The effect that kind abandons function is superior to class balanced, crossover entropy loss function.
It is described the invention in detail above by embodiment, but the content is only exemplary implementation of the invention
Example, should not be considered as limiting the scope of the invention.Protection scope of the present invention is defined by the claims.All utilizations
Technical solutions according to the invention or those skilled in the art are under the inspiration of technical solution of the present invention, in reality of the invention
In matter and protection scope, designs similar technical solution and reach above-mentioned technical effect, or to made by application range
All the changes and improvements etc. should still belong to patent of the invention and cover within protection scope.It should be noted that in order to clear
It is stated, part and protection scope of the present invention is omitted in explanation of the invention without being directly significantly associated with but this field skill
The statement of component known to art personnel and processing.
Claims (5)
1. a kind of training retinopathy height based on loss function divides network method, which comprises the following steps:
Step 1: the pretreatment eyeground IDRiD data set chooses a certain number of eye fundus images as training set and test set respectively;
With certain resolution ratio to every image down sampling;Data enhancing is carried out to training set;The hyper parameter of setting segmentation network;
Step 2: the weight of initialization segmentation each layer of network;
Step 3: selecting an eye fundus image at random from the training set after expansion;And it is cut out a scale at random from image
Very little region;
Step 4: eye fundus image obtains the input of loss function layer through the processing module in over-segmentation network, i.e. segmentation probability graph p
With corresponding mark y;
Step 5: according to setting, selecting a kind of discarding function;
Step 6: to each of image negative sample, according to indicator function as a result, determining that it is retained or abandons;
Step 7: weight factor β is calculated,
Wherein | Y+| be positive number of samples, i.e. segmentation sampled pixel number, | Y-| for the negative sample retained
Number, i.e. background pixel number;
Step 8: propagated forward loss is calculated,
Loss function is as follows:
Wherein β be retain negative sample number divided by the sum of itself and positive sample number, be calculated using formula in step 7;
Step 9: to each of image sample, calculating respective gradient information;
Step 10: the gradient backpropagation of loss layer updates the weighting parameter in segmentation network characterization processing module;
Step 11: if network is still not converged, or not up to maximum number of iterations, return step 3;
Step 12: after network training terminates, aneurysms segmentation being carried out to eye fundus image on test set, is tied according to segmentation
Fruit calculates PR curve.
2. a kind of training retinopathy height based on loss function according to claim 1 divides network method, feature
It is, in step 1, data enhancing is carried out by the way of rotation, mirror image to training set.
3. a kind of training retinopathy height based on loss function according to claim 1 divides network method, feature
It is, in step 5, abandoning function will activate probability to be mapped to drop probability, and loss function is according to discarding intensity and calculates cost
It is different there are three types of abandoning function, it is as follows respectively:
It is linear to abandon function: pdrop(pj)=1.0-pj;
Square abandon function: pdrop(pj)=(1.0-pj)2;
Logarithm abandons function: pdrop(pj)=1.0+log (1.0-pj)
For the negative sample of discarding, its loss contributed is 0.
4. a kind of training retinopathy height based on loss function according to claim 1 divides network method, feature
It is, in step 6, indicator function is as follows,
Wherein, 0 r, the random number between 1;
pdrop(pj) it is to abandon function, retain negative sample for 1, it is on the contrary then abandon negative sample.
5. a kind of training retinopathy height based on loss function according to claim 1 divides network method, feature
It is, in step 9, gradient information calculation method is as follows:
Initialize the gradient of all samples;
To a certain sample i, gradient giIt is initialized as: pi-yi, wherein piTo activate probability value, yiFor mark;
For positive sample, gradient updating are as follows: gi=β × gi;
For the negative sample of reservation, gradient updating are as follows: gi=(1.0- β) × gi;
For the negative sample of discarding, gradient is set as 0.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910534317.6A CN110223291B (en) | 2019-06-20 | 2019-06-20 | Network method for training fundus lesion point segmentation based on loss function |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910534317.6A CN110223291B (en) | 2019-06-20 | 2019-06-20 | Network method for training fundus lesion point segmentation based on loss function |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110223291A true CN110223291A (en) | 2019-09-10 |
CN110223291B CN110223291B (en) | 2021-03-19 |
Family
ID=67814254
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910534317.6A Active CN110223291B (en) | 2019-06-20 | 2019-06-20 | Network method for training fundus lesion point segmentation based on loss function |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110223291B (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111260665A (en) * | 2020-01-17 | 2020-06-09 | 北京达佳互联信息技术有限公司 | Image segmentation model training method and device |
CN111666997A (en) * | 2020-06-01 | 2020-09-15 | 安徽紫薇帝星数字科技有限公司 | Sample balancing method and target organ segmentation model construction method |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104268568A (en) * | 2014-09-17 | 2015-01-07 | 电子科技大学 | Behavior recognition method based on intelligent sub-space networks |
CN107679557A (en) * | 2017-09-19 | 2018-02-09 | 平安科技(深圳)有限公司 | Driving model training method, driver's recognition methods, device, equipment and medium |
CN108021916A (en) * | 2017-12-31 | 2018-05-11 | 南京航空航天大学 | Deep learning diabetic retinopathy sorting technique based on notice mechanism |
CN108345846A (en) * | 2018-01-29 | 2018-07-31 | 华东师范大学 | A kind of Human bodys' response method and identifying system based on convolutional neural networks |
US20190156154A1 (en) * | 2017-11-21 | 2019-05-23 | Nvidia Corporation | Training a neural network to predict superpixels using segmentation-aware affinity loss |
CN109829877A (en) * | 2018-09-20 | 2019-05-31 | 中南大学 | A kind of retinal fundus images cup disc ratio automatic evaluation method |
CN109886965A (en) * | 2019-04-09 | 2019-06-14 | 山东师范大学 | The layer of retina dividing method and system that a kind of level set and deep learning combine |
KR101953752B1 (en) * | 2018-05-31 | 2019-06-17 | 주식회사 뷰노 | Method for classifying and localizing images using deep neural network and apparatus using the same |
-
2019
- 2019-06-20 CN CN201910534317.6A patent/CN110223291B/en active Active
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104268568A (en) * | 2014-09-17 | 2015-01-07 | 电子科技大学 | Behavior recognition method based on intelligent sub-space networks |
CN107679557A (en) * | 2017-09-19 | 2018-02-09 | 平安科技(深圳)有限公司 | Driving model training method, driver's recognition methods, device, equipment and medium |
US20190156154A1 (en) * | 2017-11-21 | 2019-05-23 | Nvidia Corporation | Training a neural network to predict superpixels using segmentation-aware affinity loss |
CN108021916A (en) * | 2017-12-31 | 2018-05-11 | 南京航空航天大学 | Deep learning diabetic retinopathy sorting technique based on notice mechanism |
CN108345846A (en) * | 2018-01-29 | 2018-07-31 | 华东师范大学 | A kind of Human bodys' response method and identifying system based on convolutional neural networks |
KR101953752B1 (en) * | 2018-05-31 | 2019-06-17 | 주식회사 뷰노 | Method for classifying and localizing images using deep neural network and apparatus using the same |
CN109829877A (en) * | 2018-09-20 | 2019-05-31 | 中南大学 | A kind of retinal fundus images cup disc ratio automatic evaluation method |
CN109886965A (en) * | 2019-04-09 | 2019-06-14 | 山东师范大学 | The layer of retina dividing method and system that a kind of level set and deep learning combine |
Non-Patent Citations (2)
Title |
---|
GUOSONG等: "deeply supervised neural network with short connection for retinal vessel segmentation", 《ARXIV》 * |
褚晶辉等: "一种基于级联卷积网络的三维脑肿瘤精细分割", 《激光与光电子学进展》 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111260665A (en) * | 2020-01-17 | 2020-06-09 | 北京达佳互联信息技术有限公司 | Image segmentation model training method and device |
CN111260665B (en) * | 2020-01-17 | 2022-01-21 | 北京达佳互联信息技术有限公司 | Image segmentation model training method and device |
CN111666997A (en) * | 2020-06-01 | 2020-09-15 | 安徽紫薇帝星数字科技有限公司 | Sample balancing method and target organ segmentation model construction method |
CN111666997B (en) * | 2020-06-01 | 2023-10-27 | 安徽紫薇帝星数字科技有限公司 | Sample balancing method and target organ segmentation model construction method |
Also Published As
Publication number | Publication date |
---|---|
CN110223291B (en) | 2021-03-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109902715B (en) | Infrared dim target detection method based on context aggregation network | |
CN109272107A (en) | A method of improving the number of parameters of deep layer convolutional neural networks | |
CN104408707B (en) | Rapid digital imaging fuzzy identification and restored image quality assessment method | |
CN108961208A (en) | A kind of aggregation leucocyte segmentation number system and method | |
CN110110709A (en) | A kind of red white corpuscle differential counting method, system and equipment based on image procossing | |
CN114359738B (en) | Cross-scene robust indoor people number wireless detection method and system | |
CN110096994A (en) | A kind of small sample PolSAR image classification method based on fuzzy label semanteme priori | |
CN110569782A (en) | Target detection method based on deep learning | |
CN111275686B (en) | Method and device for generating medical image data for artificial neural network training | |
CN109102498B (en) | Method for segmenting cluster type cell nucleus in cervical smear image | |
CN104751147A (en) | Image recognition method | |
CN109344851A (en) | Image classification display methods and device, analysis instrument and storage medium | |
CN108664838A (en) | Based on the monitoring scene pedestrian detection method end to end for improving RPN depth networks | |
Ru et al. | Speedy performance estimation for neural architecture search | |
CN110349167A (en) | A kind of image instance dividing method and device | |
CN108122221A (en) | The dividing method and device of diffusion-weighted imaging image midbrain ischemic area | |
CN105718942A (en) | Hyperspectral image imbalance classification method based on mean value drifting and oversampling | |
CN109448854A (en) | A kind of construction method of pulmonary tuberculosis detection model and application | |
CN110223291A (en) | A kind of training retinopathy height segmentation network method based on loss function | |
CN108734200A (en) | Human body target visible detection method and device based on BING features | |
Wollmann et al. | DetNet: Deep neural network for particle detection in fluorescence microscopy images | |
Sun et al. | Localizing optic disc and cup for glaucoma screening via deep object detection networks | |
Shoohi et al. | DCGAN for Handling Imbalanced Malaria Dataset based on Over-Sampling Technique and using CNN. | |
Sasmal et al. | A survey on the utilization of Superpixel image for clustering based image segmentation | |
CN110610488A (en) | Classification training and detecting method and device |
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 |