WO2020252665A1 - 用于分割医学图像中的重叠细胞质的方法及系统 - Google Patents

用于分割医学图像中的重叠细胞质的方法及系统 Download PDF

Info

Publication number
WO2020252665A1
WO2020252665A1 PCT/CN2019/091761 CN2019091761W WO2020252665A1 WO 2020252665 A1 WO2020252665 A1 WO 2020252665A1 CN 2019091761 W CN2019091761 W CN 2019091761W WO 2020252665 A1 WO2020252665 A1 WO 2020252665A1
Authority
WO
WIPO (PCT)
Prior art keywords
shape
evolution
cytoplasm
hypothesis
alignment
Prior art date
Application number
PCT/CN2019/091761
Other languages
English (en)
French (fr)
Inventor
宋有义
秦璟
Original Assignee
香港理工大学
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 香港理工大学 filed Critical 香港理工大学
Priority to US17/596,382 priority Critical patent/US20220237782A1/en
Priority to CN201980000985.9A priority patent/CN112470166A/zh
Priority to PCT/CN2019/091761 priority patent/WO2020252665A1/zh
Publication of WO2020252665A1 publication Critical patent/WO2020252665A1/zh

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
    • G06T7/11—Region-based segmentation
    • 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
    • G06T7/136—Segmentation; Edge detection involving thresholding
    • 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
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H30/00—ICT specially adapted for the handling or processing of medical images
    • G16H30/40—ICT specially adapted for the handling or processing of medical images for processing medical images, e.g. editing
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/70—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients
    • 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
    • G06T2207/20084—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
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30024—Cell structures in vitro; Tissue sections in vitro

Definitions

  • the present invention relates to a method and system for segmenting overlapping cytoplasm in medical images, and more specifically, to a method and system for segmenting overlapping cytoplasm in medical images for cervical cancer screening.
  • Cervical cancer ranks fourth in the mortality rate of malignant tumors in women, and is also the fourth cancer in incidence.
  • High-quality cervical cancer screening can greatly reduce the incidence and mortality of cervical cancer.
  • cervical cancer screening refers to examining the abnormality of each cervical cell sampled from the cervix and placed on a glass slide under a microscope to assess whether there are cervical cancer cells.
  • segmentation of the overlapping cytoplasm of the cells in the cervical image is a key task, because in order to check the abnormality of the cells, the characteristics of the cell level (for example, the shape of the cells, Size, and the area ratio of cytoplasm to nucleus) are clinically important.
  • the intensity (or color) information in the overlapping area is usually confusing and even misleading, this intensity (or The lack of color information makes this task very challenging.
  • the traditional method of segmenting the overlapping cytoplasm in medical images is achieved by leveraging the intensity information between the cytoplasm of each cell in the clump or combining spatial information. This goal is generally achieved by extending typical segmentation models, such as threshold segmentation, watershed segmentation and image cutting. These methods theoretically assume that the intensity information is sufficient to identify the occluded boundary part. However, this assumption is flawed. In fact, the intensity information of the overlapping regions is usually confusing and even misleading.
  • the shape prior-based method shows a good segmentation performance, which inserts additional shape information into the segmentation method.
  • These methods either use simple shape estimation (for example, the cytoplasm has an elliptical or star shape), or match shape examples from a limited set of shapes collected in advance to model the prior shape. Then, using the intensity information, the segmentation result needs to be as similar as possible to the modeled prior shape to segment the overlapping cytoplasm. It is usually realized by an active contour model or a level set model, where the prior shape is designed as a regular term in the energy function, and it is assumed that the minimum (or maximum) value of the energy function is obtained by the segmentation function that produces the best segmentation result.
  • these methods only use the local prior shape (that is, the prior information is only the shape of a single cytoplasm) to evolve the shape of the cytoplasm, without considering the shape relationship between all cytoplasm and clumps, which leads to the segmentation of these methods The results are usually inconsistent with the clump evidence, and the segmented cytoplasmic boundary deviates from the ideal boundary of the clump.
  • these methods require that the final shape is as similar to the modeled prior shape as possible, these methods do not impose shape constraints on the final shape. In fact, these methods try to balance the parameters between the strength evidence and the local prior shape. Find the right compromise, so these methods can produce enormous segmentation results when the strength evidence contradicts the local prior shape.
  • a technical problem to be solved by the present invention is to provide a method and system that can more accurately and effectively segment the overlapping cytoplasm in medical images based on the prior shape.
  • the present invention relates to a method for segmenting overlapping cytoplasm in a medical image, which includes: establishing a cytoplasmic shape hypothesis set; and selecting a shape hypothesis for each cytoplasm from the established cytoplasmic shape hypothesis set to perform constrained To segment the overlapping cytoplasm in medical images, wherein the constrained multi-shape evolution includes: segmenting a clump area composed of multiple overlapping cytoplasms to provide clump evidence; The shape alignment of the mass of the shape hypothesis; and the shape evolution to determine a better shape hypothesis for each cytoplasm.
  • the method for segmenting overlapping cytoplasm in a medical image further includes a step of updating the learned shape instance importance of the cytoplasm shape instance for the established cytoplasm shape hypothesis set.
  • the shape alignment and the shape evolution are performed iteratively, the output of the shape alignment is used as the input of the shape evolution, and the output of the shape evolution is used as the input of the shape alignment, and the shape alignment evaluates whether a new shape evolution needs to be started Once it is determined that no new shape evolution is needed, the current shape hypothesis after shape alignment is the segmentation result of the overlapping cytoplasm.
  • the cytoplasmic shape hypothesis set is established by the following formula:
  • ⁇ represents the average shape of the collected shape instances
  • Mx i represents the linear combination of the eigenvectors of the covariance matrix of the collected shapes, where each column of the matrix M is an eigenvector, and x i is the weight vector of the linear combination
  • S i represents the shape hypothesis labeled i.
  • the shape alignment includes filling in the shape hypothesis selected from the cytoplasmic shape hypothesis set to obtain a binary image, and the following formula is used to obtain the rotation angle and the scaled size of the binary image and the corresponding cytoplasmic alignment. :
  • B c represents the image of the segmented blob area
  • B i represents the aligned result image
  • B i is a binary image with the same size as B c
  • B i is required to be inside B c .
  • the shape evolution includes: setting an objective function, and the objective function is:
  • x represents the set x k represents the k-th evolution x; N represents the number of cytoplasm in the clump area; Is the result of alignment The generated binary image; (x, y) represents the coordinates of the pixels in the image; and confirm that the X k+1 having a lower value than x k .
  • the determination enables X k+1, which has a lower value than x k includes: for the matrix p obtained from the objective function, the formula is obtained using Taylor's theorem:
  • ⁇ is a scalar between the interval (0,1), so as to obtain the area with x k as the center and
  • m k represents the position at x k Approximate by the minimum value of m k
  • the minimum value of and solve the formula by trust region method:
  • the step of learning the importance of shape instances includes: randomly selecting a group of shape instances, and calculating the average shape of the group of shape instances according to the following formula:
  • the t first feature vector matrix M c constituting the matrix M (e 1 e 2 ... e t), the corresponding eigenvalues ⁇ 1 ⁇ 2 ⁇ ... ⁇ t .
  • the step of learning the importance of the shape instance is recalculated to update the shape hypothesis set; until the segmentation result is no longer available The update is stopped when the predetermined threshold is reduced or reached.
  • the present invention also relates to a system for segmenting overlapping cytoplasm in medical images, which includes: a shape hypothesis set module for establishing a cytoplasmic shape hypothesis set; and a multi-shape evolution module for creating The cytoplasmic shape hypothesis focuses on selecting shape hypotheses for each cytoplasm to perform constrained multi-shape evolution to segment overlapping cytoplasms in medical images, wherein the multi-shape evolution module is configured to: The clump area is segmented to provide clump evidence; shape alignment used to assess the quality of the selected shape hypothesis; and shape evolution to determine a better shape hypothesis for each cytoplasm.
  • the system for segmenting overlapping cytoplasm in a medical image further includes a learning shape instance importance module that updates the cytoplasm shape instance for the established cytoplasm shape hypothesis set.
  • the learning shape instance importance module is configured to randomly select a group of shape instances, and calculate the average shape of the group of shape instances according to the following formula:
  • the t first feature vector matrix M c constituting the matrix M (e 1 e 2 ... e t), the corresponding eigenvalues ⁇ 1 ⁇ 2 ⁇ ... ⁇ t .
  • the segmentation result obtained by the constrained multi-shape evolution module is greater than a predetermined threshold, it is recalculated by the learning shape instance importance module to update the shape hypothesis set; until the segmentation result is no longer available The update is stopped when the predetermined threshold is reduced or reached.
  • the method and system of the present invention model a priori shape through an infinite set of shape hypotheses, and at the same time combine the local prior shape and the overall prior shape with intensity information for evolution, and constrain the result shape in each evolution to The shape assumptions are concentrated.
  • the method and system of the present invention can better identify the occluded boundary portion, thereby better segmenting the overlapping cytoplasm, thereby providing medical diagnosis Provides more precise shape characteristics.
  • the infinite shape hypothesis set established by the present invention can better describe all possible shapes of the cytoplasm, thereby more efficiently segmenting overlapping cytoplasms with different shapes;
  • the constrained multi-shape evolution algorithm of the present invention considers all cytoplasm and The shape relationship of the entire clump is used to combine the local prior shape and the overall prior shape and intensity information for evolution, so as to obtain more information for segmentation;
  • the present invention uses the shape statistics calculation of each shape instance Importance, and thus invisible shapes, can be well approximated by the shape hypotheses in the shape hypothesis set;
  • the multi-shape evolution step of the present invention embeds the learning step to obtain useful information more effectively. Therefore, compared with the prior art, the method and system for segmenting overlapping cytoplasm in medical images of the present invention can obtain more accurate results
  • Figures 1a-1d show schematic diagrams of segmenting a clump with overlapping cytoplasm according to the method of the present invention
  • FIG. 2 shows a flowchart of the method of the present invention
  • Figure 3 shows a flow chart of the constrained multi-shape evolution steps of the present invention
  • Figure 4 shows a schematic diagram of the overlapping cytoplasmic segmentation system of the present invention.
  • Figure 5 shows a comparison image of segmentation results obtained according to the method of the present invention and the method of the prior art
  • the present invention divides all overlapping cytoplasm in a clump by evolving the cytoplasm shape guided by the modeled local prior shape and the overall prior shape, and simultaneously evolving the mutual shape constraints of the cytoplasm, so that the segmentation based on the prior shape of the present invention
  • the method of overlapping cytoplasm in medical images can accurately and efficiently obtain the cytoplasmic segmentation results, thereby improving the accuracy and efficiency of cervical cancer screening.
  • a shape hypothesis set with infinite cytoplasmic shape hypotheses is established; in the multi-shape evolution step, in addition to considering the local prior shape, the present invention has evolved The shape of the cytoplasm, and then the algorithm is used to make the segmentation result consistent with the clump evidence, so as to obtain the overall prior shape; in addition, in the multi-shape evolution step, the final shape obtained in the evolution process is required to be in the shape assumption set, thereby reducing the existing technology In order to make the established shape hypothesis better restore any invisible cytoplasmic shape, the present invention also adds the step of learning the importance of shape instances in the shape statistics calculation.
  • FIGS. 1a-1d show schematic diagrams of using the method of the present invention to segment a clump with overlapping cytoplasm.
  • Figure 1a shows the input blob image
  • Figure 1b shows the initially segmented blob area
  • Figure 1c shows the deviation area 11 in the random evolution and the area 12 covered by all aligned shapes
  • Figure 1d shows the segmentation result obtained by the method of the present invention.
  • 2 shows a flow chart of the method of the present invention.
  • the method includes three steps: a step 201 of establishing a shape hypothesis set, a step 202 of constrained multi-shape evolution, and a step 203 of learning the importance of shape instances.
  • the following content describes in detail the three steps of the above-mentioned building shape hypothesis set step 201, constrained multi-shape evolution step 202, and learning shape instance importance step 203.
  • the shape of the cytoplasm of the cell is parameterized.
  • a boundary point in the form of a vector s is used to describe the shape of the cytoplasm of each cell.
  • the k-th item s stores the distance value of the boundary point whose angle value is equal to k in the polar coordinate system, and the origin of the polar coordinate system is located at the centroid of the cell nucleus. It should be noted that each cell is composed of cytoplasm and nucleus.
  • the centroid of the nucleus rather than the centroid of the cytoplasm is selected as the origin of the polar coordinate system for the consideration of feasibility, because when the cytoplasm of each cell overlaps ( See Figure 1a), it is easier to detect the centroid of the nucleus than to detect the centroid of the cytoplasm.
  • the shape hypothesis set is established using the statistical shape information of the cytoplasm.
  • the shape hypothesis set is expressed as the following formula (1):
  • Mx i represents the linear combination of the eigenvectors of the covariance matrix of the collected shapes (where each column of the matrix M is an eigenvector, and x i is the weight vector of the linear combination)
  • S i represents the shape hypothesis labeled i.
  • FIG. 3 is a flow chart of the constrained multi-shape evolution of the present invention.
  • the constrained multi-shape evolution step of the present invention is based on clump evidence, while selecting a shape hypothesis for each cytoplasm from a set of established shape hypotheses to segment overlapping cytoplasms. It includes three steps: segmentation of the clump area to provide clump evidence (step 301); shape alignment to evaluate the quality of the current shape hypothesis (step 302); and shape evolution to find a better shape hypothesis for each cytoplasm (Step 303).
  • the shape alignment (step 302) and shape evolution (step 303) are performed iteratively.
  • the shape evolution takes the result of shape alignment as input, and the output of shape evolution (step 303) is used as the input of shape alignment to detect Do you need to start a new round of shape evolution. Once it is determined that no new shape evolution is needed, the current shape hypothesis aligned with the shape is regarded as the segmentation result of the cytoplasm (see Figure 1d).
  • the present invention uses a multi-size convolutional neural network (CNN) to segment the cytoplasm and nucleus region (see Figure 1b).
  • CNN convolutional neural network
  • the multi-size convolutional neural network divides each pixel in the image into three groups: nuclear part, cytoplasmic part and background part. It contains three parallel CNNs, and each CNN has a different size in terms of the resolution of the input patch. The outputs of these three CNNs are then merged to help capture more contextual information in different sizes. Finally, Markov random field is used to further optimize the segmentation results.
  • the shape alignment for the shape hypothesis s i , since it is only a vector for storing boundary point information, it is necessary to fill in the area inside the contour described by s i to obtain the corresponding binary image of s i ( That is, in a binary image), the pixels inside the contour are marked as 1, and the pixels outside the contour are marked as 0.
  • s i is assigned as the output of the shape evolution step, but the average shape of the instances collected from the shape hypothesis set is used as the initial s i of each cytoplasm.
  • the present invention can bypass the non-rigid transformation through the evolutionary shape assumption described below, the present invention limits the shape alignment to rigid alignment.
  • B c represents the image of the segmented blob area
  • B i represents the alignment result, which is obtained by rotating the area filled with s i by the angle ⁇ i and scaling it with the number r i
  • B i is the same as B c Binary images with the same size
  • the values of r i and ⁇ i are determined by grid search.
  • the shape evolution algorithm of the present invention can find a more suitable cytoplasmic shape hypothesis than s i .
  • the objective function as shown in the following formula (3) needs to be defined:
  • x is used to represent the collection
  • x k represent the k-th evolution x, as described in formula (1)
  • si is determined by x i ;
  • N represents the number of cytoplasm in the clump;
  • Align result The generated binary image; (x, y) represents the coordinates of the pixels in the image.
  • the objective function represented by the formula (3) is actually to detect the pixel difference between the segmented blob area and the blob area composed of the alignment result.
  • the objective function is designed in the method of the present invention. The main reason is to make full use of the boundary information of the clumps while minimizing the insufficient strength in the overlapping area. The impact of information.
  • the present invention uses the following formulas (4)-(6) to find the X k+1 having a lower value than x k .
  • formula (4) is obtained using Taylor's theorem here:
  • the minimum value of the overall area is the best x k+1 that can be used in the kth evolution.
  • the value of the scalar ⁇ is unknown, it cannot be directly analyzed Therefore, go back to x k to approximate Expressed by m k in the following formula (5) at x k
  • Equation (6) can be solved by the existing trust region method (see J. Nocedal and SJ Wright. “Numerical Optimization”. Springer, 2006), and finally, the output result of the kth evolution is shown in Equation (7) Out:
  • the shape assumption in can better restore any invisible cytoplasmic shape.
  • the present invention also adopts the step 203 of learning the importance of shape instances in the shape statistics calculation.
  • the method for learning the importance of shape instances of the present invention can solve a series of problems caused by manual collection of shape instances in the prior art.
  • the t first feature vector matrix M c constituting the matrix M (e 1 e 2 ... e t), the corresponding eigenvalues ⁇ 1 ⁇ 2 ⁇ ... ⁇ t .
  • the correlation between the step 203 of learning shape instance importance and the above-mentioned constrained multi-shape evolution step 202 and the establishment of shape hypothesis step 201 is: in each training image , Run the constrained multi-shape evolution step according to the initial ⁇ and M in the step of establishing the shape hypothesis set, and then detect the final in case Greater than a predetermined threshold Then the current ⁇ and M are not ideal, so they need to be recalculated according to formulas (9) and (10) in the step of learning the importance of shape instances to increase Thus, ⁇ and M in the shape hypothesis set are updated. During the update process, the importance of each shape instance It is determined by grid search. Continue the above increase Steps until It starts to get smaller. The steps of learning the importance of shape instances are based on formula (8) in Repeat until all final differences The sum is no longer reduced.
  • the present invention also relates to an overlapping cytoplasmic segmentation system, which includes: a shape hypothesis set module M1 for establishing a shape hypothesis set, an evolution module M2 for performing constrained multi-shape evolution, and an evolution module M2 for learning The learning module M3 of the importance of shape instances.
  • the shape hypothesis module M1 provides the initial ⁇ and M, then runs the constrained multi-shape evolution in the evolution module M2, and then detects the final in case Greater than a predetermined threshold It means that the current ⁇ and M are not ideal, you need to recalculate according to the above formulas (9) and (10) in the learning module M3 to increase Thus, ⁇ and M in the shape hypothesis set are updated. Continue the above increase Steps until It starts to get smaller.
  • This example is based on two typical cervical scraping data sets, which are the Pap staining data set and the H&E staining (H&E stain) data set.
  • H&E stain H&E stain
  • this data set includes 8 publicly available images, each Each image has 11 clumps with an average of 3.3 cytoplasm; the eosin staining data set is prepared by eosin staining, the data set includes 21 images, and each image has an average of 7 clumps with 6.1 cytoplasm.
  • Training set There are 72 clumps and 324 cytoplasmic instances, of which 28 isolated cytoplasms are used to initialize a small collection of shape instances There are 184 clumps containing 907 cytoplasm in the test set, and the number of cytoplasm in each clump is 2 to 13 (the average is 4.93, the standard deviation is 1.81).
  • a predetermined threshold for terminating the multi-shape evolution step And used to calculate the value of the eigenvector t of the matrix M.
  • a predetermined threshold It is set to be between 3% and 7% of the number of pixels in the blob, preferably about 5%.
  • a larger t will make the evaluated shape si show more details of the overall shape, but it will also consume more computing resources in the shape evolution process.
  • the results obtained by the method according to the present invention are compared with four prior art techniques (see Table 1 below).
  • These four existing technologies are the joint level set function method (see Z.Lu, G.Carneiro and APBradley.”An improved joint optimization of multiple level set functions for the segmentation of overlapping cells", IEEE Transactions on Image Processing ,24(4):1261–1272,2015.), multi-cell labeling (see Y. Song, ELTan, X. Jiang, etc., "Accurate Cervical Cell Segmentation from Overlapping Clumps in Pap Smear Images", IEEE Transactions on Medical Imaging,36(1):288-300,2017), multi-channel watershed method (see A. Tareef, Y. Song, H.
  • Table 1 lists the quantitative comparisons of segmentation results obtained by different methods under multiple overlap conditions.
  • the degree of overlap is to measure the degree of overlap, which is calculated from the ratio of the length of the occluded boundary portion to the entire boundary portion of the cytoplasm. It can be seen from the results in Table 1 that the method of the present invention obtains the best segmentation results. Compared with other methods, the accuracy of the method of the present invention is improved by about 5% on average. Specifically, when the degree of overlap is less than 0.5 (see the Column content), the accuracy of the method of the present invention is improved by about 3% on average; when the overlap is greater than 0.5 (see the table in the Column content), the accuracy improved by about 8% on average.
  • FIG. 5 shows a comparison image of the segmentation results obtained according to the method of the present invention and the method of the prior art.
  • Figure 5-(a) is the original input image
  • Figure 5-(b) to Figure 5-(f) respectively show the images of the segmentation results obtained by LSF, MCL, MPW, CF and the method of the present invention
  • Figure 5-(g) is the actual segmented image. It can be seen more intuitively from Figure 5 that the segmentation method of the present invention (as shown in Figure 5-(f)) obtains the best segmentation results. For some of the cytoplasmic examples, it can even be compared with the actual segmentation results (as shown in Figure 5). -(g) shows) exactly the same.
  • the method and system of the present invention overcome the problem that the cytoplasm cannot be accurately segmented due to lack of intensity information in the overlapping area. Compared with the existing prior shape-based technology, the method and system of the present invention provide an infinite set of shape hypotheses, calculate and evolve local prior shapes and overall prior shapes, and impose shape constraints on the final result. A more accurate method and system for cytoplasmic segmentation of overlapping regions.
  • the method and system of the present invention are not limited to the detection of cervical cancer, and those skilled in the art can make appropriate improvements so that the method and system of the present invention can be applied to other microscopy images that quantitatively measure cell-level features, such as pathological images. measuring.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Public Health (AREA)
  • General Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Primary Health Care (AREA)
  • Epidemiology (AREA)
  • Radiology & Medical Imaging (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Data Mining & Analysis (AREA)
  • Biomedical Technology (AREA)
  • Databases & Information Systems (AREA)
  • Pathology (AREA)
  • Multimedia (AREA)
  • Quality & Reliability (AREA)
  • Image Analysis (AREA)

Abstract

一种分割医学图像中的重叠细胞质的方法，其包括：建立细胞质形状假设集（201）；以及从建立的细胞质形状假设集中为每个细胞质选择形状假设，来进行受约束的多形状演进（202），从而分割医学图像中的重叠细胞质，其中，受约束的多形状演进（202）包括：对由多个重叠细胞质构成的团块区域进行分割以提供团块证据（301）；用于评估选择的形状假设的质量的形状对齐（302）；以及为每个细胞质确定更好的形状假设的形状演进（303）。还涉及一种分割医学图像中的重叠细胞质的系统。系统和方法能够更精确、更有效地基于先验形状进行重叠细胞质的分割。

Description

用于分割医学图像中的重叠细胞质的方法及系统 技术领域
本发明涉及一种用于分割医学图像中的重叠细胞质的方法及系统，更具体地，涉及一种用于宫颈癌筛查的医学图像中的重叠细胞质的分割方法及系统。
背景技术
宫颈癌在女性的恶性肿瘤中死亡率居第四位，也是发病率排名第四的癌症，而高质量的宫颈癌筛查能够极大地减少宫颈癌的发病率和死亡率。具体而言，宫颈癌筛查是指在显微镜下检查从宫颈处取样并放置在玻片上的每个宫颈细胞的异常程度，从而评估是否存在宫颈癌细胞。
对于发展宫颈癌自动筛查系统来说，对宫颈图像中的细胞的重叠的细胞质进行分割是其中很关键的任务，因为为了检查细胞的异常程度，提取细胞水平的特征(例如，细胞的形状、尺寸、和细胞质与细胞核的面积比)在临床上是很重要的，然而，由于在重叠区域中强度(或颜色)信息通常是混乱的，甚至会让人产生误解，因此，这种强度(或颜色)信息的不足导致这项任务是非常挑战的。
传统的分割医学图像中的重叠细胞质的方法是利用(leveraging)团块中的各细胞的细胞质之间的强度信息或者结合空间信息来实现的。普遍通过扩展典型的分割模型来实现这个目的，例如采用阈值分割法、分水岭分割法和图像切割法。这些方法在理论上假设了强度信息足以识别被遮挡的边界部分。然而，这种假设是有缺陷的，实际上，重叠区域的强度信息通常是混乱的，甚至会让人误解。
为了消除强度信息不足带来的问题，基于先验形状(shape prior-based)的方法显现出了很好的分割表现，其将额外的形状信息插入到分割方法 中。而这些方法要么通过简单的形状推测(例如细胞质具有椭圆形或星形的形状)、要么从预先收集的有限的形状集合中匹配形状实例，来建模先验形状。然后，利用强度信息，同时需要分割的结果与建模的先验形状尽可能地相似，来分割重叠的细胞质。通常采用主动轮廓模型或水平集模型来实现，其中，将先验形状设计成能量函数中的正则项，假设该能量函数的最小(或最大)值由产生最佳分割结果的分割函数获得。
尽管现有技术已改进了分割的精确度，但是这些现存的先验形状方法仍然具有三个主要的缺点。首先，这些方法都是通过有限的形状假设(例如，形状推测或收集的形状实例)来建模先验形状，由于这些具体的形状假设不能很好地还原细胞质被遮挡的边界部分，导致这些方法建模的先验形状通常不足以查明细胞质被遮挡的边界部分。其次，这些方法仅仅使用了局部先验形状(即，先验信息仅仅是单个的细胞质形状)来演进细胞质的形状，而没有考虑所有细胞质和团块之间的形状关系，这导致这些方法的分割结果通常与团块证据不一致，被分割的细胞质边界与团块的理想边界有偏差。再次，尽管这些方法要求最终形状与建模的先验形状尽可能地相似，但是这些方法没有对最终形状施加形状约束，实际上，这些方法试图通过平衡参数在强度证据和局部先验形状之间找到合适的妥协，因此，当强度证据与局部先验形状相矛盾时，这些方法会产生令人难以置信的分割结果。
此外，对于无限形状假设集的建立，现有技术中(例如，T.F.Cootes,C.J.Taylor,and et al.D.H.Cooper.“Active shape models-their training and application”，Computer Vision and Image Understanding,61(1):38–59,1995)公开了相关内容。虽然采用该方法能够通过选择无限的数值来得到无限大的形状假设集，但是其主要缺点在于：很难收集一组形状实例，使得能确保不可见的细胞质形状被建立的形状假设很好地还原。现有技术中，建立形状假设集依赖于如何收集到好的形状实例，通常，不得不采用手动选择具有代表性的形状实例，这是一种实验性并具有误差的方式，也是一种劳 动密集型方式，当需要收集无限多的实例用来近似复杂的形状时，这种方法是不可行的。
发明内容
因此，本发明要解决的一个技术问题是：提供一种能够更精确、更有效地基于先验形状的分割医学图像中的重叠细胞质的方法及系统。
在一个实施方式中，本发明涉及一种分割医学图像中的重叠细胞质的方法，其包括：建立细胞质形状假设集；以及从建立的细胞质形状假设集中为每个细胞质选择形状假设，来进行受约束的多形状演进，从而分割医学图像中的重叠细胞质，其中，所述受约束的多形状演进包括：对由多个重叠细胞质构成的团块区域进行分割以提供团块证据；用于评估选择的形状假设的质量的形状对齐；以及为每个细胞质确定更好的形状假设的形状演进。
所述分割医学图像中的重叠细胞质的方法还包括为所述建立的细胞质形状假设集更新细胞质形状实例的学习形状实例重要度的步骤。
优选地，所述形状对齐和所述形状演进是迭代进行的，形状对齐的输出作为形状演进的输入，而形状演进的输出作为形状对齐的输入，所述形状对齐评估是否需要开始新的形状演进，一旦确定不需要新的形状演进，则形状对齐后的当前的形状假设就是该重叠细胞质的分割结果。
优选地，由以下公式建立所述细胞质形状假设集：
Figure PCTCN2019091761-appb-000001
其中，μ表示收集的形状实例的平均形状，Mx i表示收集的形状的协方差矩阵的特征向量的线性组合，其中，矩阵M的每一列是一个特征向量，x i是该线性组合的权重向量，s i表示标号为i的形状假设。
优选地，所述形状对齐包括填充从细胞质形状假设集中选择的形状假设以获得二值图像，并利用下述公式来获得所述二值图像与对应的细胞质对齐所需要旋转的角度和缩放的大小：
Figure PCTCN2019091761-appb-000002
其中，B c表示分割后的团块区域的图像；B i表示对齐的结果图像，B i是与B c具有相同尺寸的二值图像，且需要B i在B c的内部。
优选地，所述形状演进包括：设定目标函数，所述目标函数为：
Figure PCTCN2019091761-appb-000003
其中，x表示集合
Figure PCTCN2019091761-appb-000004
x k表示第k次演进的x；N表示所述团块区域中细胞质的数量；
Figure PCTCN2019091761-appb-000005
是由对齐结果
Figure PCTCN2019091761-appb-000006
生成的二值图像；(x，y)表示图像中像素的坐标；以及确定能够使
Figure PCTCN2019091761-appb-000007
具有比x k更低的值的x k+1。
优选地，所述确定能够使
Figure PCTCN2019091761-appb-000008
具有比x k更低的值的x k+1包括：对于从所述目标函数得到的矩阵p，使用泰勒定理得到公式：
Figure PCTCN2019091761-appb-000009
其中，
Figure PCTCN2019091761-appb-000010
和
Figure PCTCN2019091761-appb-000011
分别表示梯度和海森矩阵计算；γ是间隔(0,1)之间的一个标量，从而获得以x k为圆心、以||p|| 2为半径的区域内的
Figure PCTCN2019091761-appb-000012
的整体区域最小值；回到x k处近似
Figure PCTCN2019091761-appb-000013
其中，m k表示x k处的
Figure PCTCN2019091761-appb-000014
由m k的最小值来近似
Figure PCTCN2019091761-appb-000015
的最小值，并通过信赖域方法来解公式：
Figure PCTCN2019091761-appb-000016
以及得到第k次演进的输出结果：x k+1＝x k+p *。
优选地，所述学习形状实例重要度的步骤包括：随机选择一组形状实例，根据以下公式计算所述一组形状实例的平均形状：
Figure PCTCN2019091761-appb-000017
其中，K s是选择的形状实例的数量；w i是每个形状实例s i的重要度；W是所有w i的和；并且，根据得到的μ计算协方差矩阵：
Figure PCTCN2019091761-appb-000018
其中，矩阵M c的第一t个特征向量构成矩阵M＝(e 1 e 2…e t)，其对应的特征值为λ 1≥λ 2≥…≥λ t。
优选地，如果由所述受约束的多形状演进得到的分割结果大于预定阈值，则由所述学习形状实例重要度的步骤重新计算，以更新所述形状假设集；直到所述分割结果不能再减小或达到了所述预定阈值，才停止所述更新。
在另一方面，本发明还涉及一种用于分割医学图像中的重叠细胞质的系统，其包括：用于建立细胞质形状假设集的形状假设集模块；以及多形状演进模块，其用于从建立的细胞质形状假设集中为每个细胞质选择形状假设，来进行受约束的多形状演进，从而分割医学图像中的重叠细胞质，其中，所述多形状演进模块被配置为：对由多个重叠细胞质构成的团块区域进行分割以提供团块证据；用于评估选择的形状假设的质量的形状对齐；以及为每个细胞质确定更好的形状假设的形状演进。
优选地，所述用于分割医学图像中的重叠细胞质的系统还包括为所述建立的细胞质形状假设集更新细胞质形状实例的学习形状实例重要度模块。
优选地，所述学习形状实例重要度模块被配置为：随机选择一组形状实例，根据以下公式计算所述一组形状实例的平均形状：
Figure PCTCN2019091761-appb-000019
其中，K s是选择的形状实例的数量；w i是每个形状实例s i的重要度；W是所有w i的和；并且，根据得到的μ计算协方差矩阵：
Figure PCTCN2019091761-appb-000020
其中，矩阵M c的第一t个特征向量构成矩阵M＝(e 1 e 2…e t)，其对应的特征值为λ 1≥λ 2≥…≥λ t。
优选地，如果由所述受约束的多形状演进模块得到的分割结果大于预定阈值，则由所述学习形状实例重要度模块重新计算，以更新所述形状假设集；直到所述分割结果不能再减小或达到了所述预定阈值，才停止所述 更新。本发明的方法和系统通过无限大的形状假设集来建模先验形状，同时将局部先验形状和整体先验形状与强度信息合并用于演进，并且将每次演进中的结果形状约束在形状假设集中。相比于现有的分割医学图像中的重叠细胞质的方法和系统，本发明的方法和系统能够更好的查明被遮挡的边界部分，由此更好的分割重叠的细胞质，从而为医学诊断提供了更加精确的形状特征。本发明建立的无限大的形状假设集能够更好的描述细胞质所有可能的形状，由此，更有效率地分割具有不同形状的重叠细胞质；本发明受约束的多形状演进算法通过考虑所有细胞质和整个团块的形状关系，来将局部先验形状和整体先验形状与强度信息合并用于演进，从而获得更多的信息用于分割；本发明在形状统计计算中利用了每个形状实例的重要度，从而是的不可见的形状都能由形状假设集中的形状假设来很好地近似；本发明的多形状演进步骤中植入学习步骤，能够更加有效地获得有用的信息。因此，相比于现有技术，本发明的分割医学图像中的重叠细胞质的方法和系统能更有效地获得更精确的结果。
附图说明
通过附图以及下面的描述，可以更好地理解本发明的技术方案，其中：
图1a-1d示出了根据本发明的方法分割具有重叠的细胞质的团块的示意图；
图2示出了本发明的方法的流程图；
图3示出了本发明的受约束的多形状演进步骤的流程图；
图4示出了本发明的重叠细胞质分割系统的示意图；以及
图5示出了根据本发明的方法和现有技术的方法得到的分割结果的比较图像
具体实施方式
本发明通过演进由建模的局部先验形状和整体先验形状引导的细胞 质形状、同时演进细胞质的相互形状约束，来分割团块中的所有重叠细胞质，使得本发明的基于先验形状的分割医学图像中的重叠细胞质的方法能够精确而高效地获得细胞质的分割结果，从而提高了宫颈癌筛查的精确性和效率。具体而言，通过利用统计的形状信息建模局部先验形状，从而建立具有无限大的细胞质的形状假设的形状假设集；在多形状演进步骤中，除了考虑局部先验形状，本发明通过演进细胞质的形状、继而使用算法使分割结果与团块证据一致，从而获得整体先验形状；此外，在多形状演进步骤中，要求演进过程得到的最终形状在形状假设集中，从而减少了现有技术中难以置信的分割结果；而为了使建立的形状假设更好地还原任何不可见的细胞质形状，本发明还在形状统计计算中增加了学习形状实例重要度的步骤。
本发明采用以下新的算法和步骤来实现新的分割医学图像中的重叠细胞质的团块的方法。图1a-1d示出了采用本发明的方法分割具有重叠细胞质的团块的示意图，。其中，图1a示出了输入的团块图像；图1b示出了最初被分割的团块区域；图1c示出了在随机演进中的偏差区域11和由所有对齐后的形状覆盖的区域12；图1d示出了采用本发明的方法得到的分割结果。图2示出了本发明的方法的流程图，该方法包括三个步骤：建立形状假设集步骤201、受约束的多形状演进步骤202，以及学习形状实例重要度步骤203。
以下内容对于上述建立形状假设集步骤201、受约束的多形状演进步骤202以及学习形状实例重要度步骤203三个步骤进行详细的描述。
建立形状假设集
首先，对细胞的细胞质的形状进行参数化。采用向量s形式的边界点来描述每个细胞的细胞质的形状。第k项s存储极坐标系中角度值等于k的边界点的距离值，该极坐标系的原点位于细胞核的质心。需要注意的是，每个细胞都由细胞质和细胞核组成，本发明选择细胞核的质心而不是细胞质的质心作为极坐标系的原点是出于可实施性的考虑，因为当各细胞的细胞质重叠时(见图1a)，相比于检测细胞质的质心，检测细胞核的质心会容易地多。
其次，根据现有技术已有的建立无限形状假设集的方法，利用细胞质统计形状信息建立形状假设集。在本发明中，将形状假设集表示为以下的公式(1)：
Figure PCTCN2019091761-appb-000021
其中μ表示收集的形状实例的平均形状，Mx i表示收集的形状的协方差矩阵的特征向量的线性组合(其中，矩阵M的每一列是一个特征向量，x i是该线性组合的权重向量)，s i表示标号为i的形状假设。
通过选取不同值的x i代入公式(1)中，可以获得不同的形状假设s i，由于可以选择无穷多的x i，因此，可以建立无穷大的形状假设集，但是依靠公式(1)建立的形状假设集很难收集到能够很好地还原不可见的细胞质形状的形状实例，而本发明通过在形状统计计算中运用以下将描述的学习形状实例重要度的步骤，克服了这一缺陷。
受约束的多形状演进
图3是本发明的受约束的多形状演进的流程图。本发明的受约束的多形状演进步骤是基于团块证据、同时从建立的形状假设集中为每个细胞质选择形状假设，来分割重叠的细胞质。其包括三个步骤：用于提供团块证据的团块区域分割(步骤301)；评估当前形状假设的质量的形状对齐(步骤302)；以及为每个细胞质寻找更好的形状假设的形状演进(步骤303)。其中形状对齐(步骤302)和形状演进(步骤303)是迭代进行的，形状演进(步骤303)将形状对齐的结果作为输入，而形状演进(步骤303)的输出作为形状对齐的输入，从而检测是否需要开始新一轮的形状演进。一旦确定不需要新的形状演进，则经形状对齐的当前的形状假设就被视为该细胞质的分割结果(见图1d)。
在团块区域分割(步骤301)中，本发明采用了多尺寸卷积神经网络(CNN)来分割细胞质和细胞核区域(见图1b)。该多尺寸卷积神经网络将图像中的每个像素分成三组：细胞核部分、细胞质部分和背景部分。其包含三个平行CNN，每个CNN在输入块(input patch)的分辨率方面具有不同尺寸。然后将这三个CNN的输出进行融合，以助于以不同的尺寸捕 获更情境化(contextual)的信息。最后，采用马可夫随机场来进一步优化分割结果。
在形状对齐(步骤302)中，对于形状假设s i，由于其仅仅是存储边界点信息的矢量，因此，需要通过填充s i描述的轮廓内部的区域，来获得s i相应的二值图像(即，二进制图像)，该轮廓内部的像素被标为1，轮廓外部的像素被标为0。如上所述，s i被赋值为形状演进步骤的输出，但从形状假设集中收集的实例的平均形状作为每个细胞质的初始s i。此外，由于本发明可以通过以下将描述的演进形状假设来绕开非刚性变换，所以本发明将形状对齐限定为刚性对齐。
具体而言，对于每个s i，首先将s i被填充的区域的质心与图像中的细胞核的质心对齐。然后，通过以下公式(2)得到用于对齐的缩放因子(r i)和旋转系数θ i。
Figure PCTCN2019091761-appb-000022
其中，B c表示分割后的团块区域的图像；B i表示对齐的结果，其通过将s i被填充的区域旋转角度θ i并使用数字r i进行缩放后得到，B i是与B c具有相同尺寸的二值图像；其中r i与θ i的值由格点搜索来确定。
需要对齐的结果B i在B c的内部，如果没有这种约束，则实际得到的形状假设是与整个团块区域对齐的，而不是与细胞质本身对齐的。
在形状演进(步骤303)中，对于对齐的结果B i，本发明的形状演进算法能够找到比s i更合适的细胞质的形状假设。首先，需要限定如以下公式(3)所示的目标函数：
Figure PCTCN2019091761-appb-000023
其中，x用来表示集合
Figure PCTCN2019091761-appb-000024
用x k来表示第k次演进的x，如公式(1)所述，由x i来确定s i；N表示团块中细胞质的数量；
Figure PCTCN2019091761-appb-000025
其是由对齐结果
Figure PCTCN2019091761-appb-000026
生成的二值图像；(x，y)表示图像中像素的坐标。
由此可见，公式(3)表示的目标函数实际上是检测已分割的团块区域与由对齐结果组成的团块区域之间关于像素的差。在理想状态下，如果 非常精确地分割所有的细胞质，
Figure PCTCN2019091761-appb-000027
等于0。如上所述，这种理想状态是很难实现的，由此，在本发明的方法中设计了目标函数，其主要原因是充分利用团块的边界信息，同时最小化重叠区域中不充分的强度信息的影响。
因此，本发明通过以下公式(4)-(6)来找到能够使
Figure PCTCN2019091761-appb-000028
具有比x k更低的值的x k+1。其中，对于任何从公式(3)得到的矩阵p，在此使用泰勒定理得到以下公式(4)：
Figure PCTCN2019091761-appb-000029
其中，
Figure PCTCN2019091761-appb-000030
和
Figure PCTCN2019091761-appb-000031
分别表示梯度和海森矩阵(Hessian)计算；γ是间隔(0,1)之间的某个标量。上述公式(4)表示可使用仅仅关于函数值、x k处的一阶导数和二阶导数的信息来近似x k附近的
Figure PCTCN2019091761-appb-000032
从而获得以x k为圆心、以||p|| 2为半径的区域内的
Figure PCTCN2019091761-appb-000033
的整体区域最小值。
从理论上而言，该整体区域最小值是能够在第k次演进时使用的最佳x k+1。但是，由于标量γ的值是未知的，不能直接分析得到
Figure PCTCN2019091761-appb-000034
因此，回到x k处来近似
Figure PCTCN2019091761-appb-000035
由以下公式(5)中的m k来表示x k处的
Figure PCTCN2019091761-appb-000036
Figure PCTCN2019091761-appb-000037
当||p|| 2很小时是非常准确的，近似误差为
Figure PCTCN2019091761-appb-000038
然后由m k的最小值来近似
Figure PCTCN2019091761-appb-000039
的最小值，如公式(6)所示：
Figure PCTCN2019091761-appb-000040
可通过现有的信赖域方法(可参见J.Nocedal和S.J.Wright.“Numerical Optimization”.Springer,2006)来解公式(6)，最终，第k次演进的输出结果在公式(7)中示出：
x k+1＝x k+p *   (7)
一旦得到x k+1，就将新的形状假设与图像对齐，然后开始新一轮的演进计算，直到
Figure PCTCN2019091761-appb-000041
不能再减小或达到了预定阈值
Figure PCTCN2019091761-appb-000042
对于细胞质i而言，对齐后的最终形状假设s i为最终的分割结果。
学习形状实例重要度
为了使根据上述公式(1)计算的形状假设集
Figure PCTCN2019091761-appb-000043
中的形状假设能够更好地还原任何不可见的细胞质形状，本发明还在形状统计计算中采用了学习形状实例重要度的步骤203。本发明的学习形状实例重要度的方法能够解决现有技术中需要手动收集形状实例所带来的一系列问题。
具体而言，先随机地选择一组K个输入-输出对，见以下公式(8)：
Figure PCTCN2019091761-appb-000044
其中，
Figure PCTCN2019091761-appb-000045
是训练图像j中团块区域被分割的图像；
Figure PCTCN2019091761-appb-000046
表示图像j中细胞质i的形状向量；N j是图像j中细胞质的数量。从上述公式(8)得到
Figure PCTCN2019091761-appb-000047
选择一小组形状实例
Figure PCTCN2019091761-appb-000048
每个实例s i最初具有一个重要度，由w i表示。然后，利用以下公式(9)计算平均形状：
Figure PCTCN2019091761-appb-000049
其中，W是所有w i的和。然后，利用以下公式(10)计算协方差矩阵：
Figure PCTCN2019091761-appb-000050
其中，矩阵M c的第一t个特征向量构成矩阵M＝(e 1 e 2…e t)，其对应的特征值为λ 1≥λ 2≥…≥λ t。
学习形状实例重要度的步骤203与上述的受约束的多形状演进步骤202和建立形状假设集步骤201的关联(见图2)在于：在每个训练图像
Figure PCTCN2019091761-appb-000051
中，根据建立形状假设集步骤中初始的μ和M，运行所述受约束的多形状演进步骤，随后检测最终的
Figure PCTCN2019091761-appb-000052
如果
Figure PCTCN2019091761-appb-000053
大于预定阈值
Figure PCTCN2019091761-appb-000054
则当前的μ和M不够理想，因此需要根据学习形状实例重要度的步骤中的公式(9)和(10)重新计算以增加
Figure PCTCN2019091761-appb-000055
从而更新形状假设集中的μ和M。在更新过程中，每个形状实例重要度
Figure PCTCN2019091761-appb-000056
是由格点搜索来确定的。继续上述增加
Figure PCTCN2019091761-appb-000057
的步骤，直到
Figure PCTCN2019091761-appb-000058
开始变小。整个学习形状实例重要度的步骤都根据公式(8)在
Figure PCTCN2019091761-appb-000059
中重复进行，直到所有最终的差值
Figure PCTCN2019091761-appb-000060
的和不再减小。
如图4所示，本发明还涉及一种重叠细胞质分割系统，其包括：用于建立形状假设集的形状假设集模块M1、用于进行受约束的多形状演进的 演进模块M2和用于学习形状实例重要度的学习模块M3。首先，形状假设集模块M1提供初始的μ和M，然后在演进模块M2中运行受约束的多形状演进，随后检测最终的
Figure PCTCN2019091761-appb-000061
如果
Figure PCTCN2019091761-appb-000062
大于预定阈值
Figure PCTCN2019091761-appb-000063
说明当前的μ和M不够理想，则需要在学习模块M3中根据上述公式(9)和(10)重新计算以增加
Figure PCTCN2019091761-appb-000064
从而更新形状假设集中的μ和M。继续上述增加
Figure PCTCN2019091761-appb-000065
的步骤，直到
Figure PCTCN2019091761-appb-000066
开始变小。
实施例
本实例基于两种典型的宫颈刮片数据集，分别是巴氏染色数据集和伊红染色(H&E stain)数据集，其中巴氏染色数据集的获得，参见(Z.Lu,G.Carneiro,A.Bradley等，“Evaluationof three algorithms for the segmentation of overlappingcervical cells”,IEEE Journal of Biomedical and Health Informatics,21(2):441–450,2017)，这个数据集包括8个公众可获得的图像，每个图像具有平均有3.3个细胞质的11个团块；伊红染色数据集由伊红染色制备，该数据集包括21个图像，每个图像具有平均有6.1个细胞质的7个团块。
首先，从巴氏染色数据集随机选取3个图像，并从伊红染色数据集随机选取5个图像，来构造训练集
Figure PCTCN2019091761-appb-000067
剩下的图像构成测试集。训练集
Figure PCTCN2019091761-appb-000068
具有72个团块和324个细胞质实例，其中的28个孤立的细胞质用于初始化小的形状实例集合
Figure PCTCN2019091761-appb-000069
在测试集中具有包含907个细胞质的184个团块，每个团块中的细胞质的数量是2至13(平均值是4.93，标准偏差为1.81)。
在本发明描述的方法中，需要设定两个参数：用于终止多形状演进步骤的预定阈值
Figure PCTCN2019091761-appb-000070
以及用于计算矩阵M的特征向量t的数值。尽管通常小的
Figure PCTCN2019091761-appb-000071
有助于改进分割结果的精度，但越小的
Figure PCTCN2019091761-appb-000072
会使计算时间越长。本发明为 了平衡分割结果的精度和计算时长，将预定阈值
Figure PCTCN2019091761-appb-000073
设定约为团块中像素数量的3％-7％之间，优选约为5％。对于特征向量t的数值，越大的t会使评估的形状s i呈现出更多整体形状的细节，但是同样会使形状演进过程消耗更多的计算资源。在本发明中，根据现有技术中(参见T.F.Cootes,C.J.Taylor，D.H.Cooper等，“Active shape models-their training and application”，Computer Vision and Image Understanding,61(1):38–59,1995)已有的公式
Figure PCTCN2019091761-appb-000074
来确定t，在本实例中t值设定为20。
在本实施例中，将根据本发明的方法得到的结果与四个现有技术进行了比较(见以下的表1)。这四个现有技术分别是联合水平集函数法(参见Z.Lu,G.Carneiro和A.P.Bradley.“An improved joint optimization of multiple level set functions for the segmentation of overlapping cervical cells”,IEEE Transactions on Image Processing,24(4):1261–1272,2015.)、多细胞标记法(参见Y.Song,E.L.Tan,X.Jiang等，“Accurate cervical cell segmentation from overlapping clumps in pap smear images”,IEEE Transactions on Medical Imaging,36(1):288–300,2017)、多通道分水岭法(参见A.Tareef,Y.Song，H.Huang等，“Multi-pass fast watershed for accurate segmentation of overlapping cervical cells”，IEEE Transactions on Medical Imaging,2018.)以及轮廓分段法(参见Y.Song,J.Qin,B.Lei等，“Automated segmentation of overlapping cytoplasm in cervical smear images via contour fragments”,In Proceedings of the 32th AAAI Conference on Artificial Intelligence,168–175页.AAAI,2018.)，在表1中分别使用LSF、MCL、MPW和CF来表示上述四种现有技术的分割结果。其中，LSF、MCL和CF属于现有的基于先验形状的 方法，MPW属于基于分水岭方法的变型。
Figure PCTCN2019091761-appb-000075
表1
表1列出了多个重叠度条件下的、由不同方法得到的分割结果的定量比较。重叠度是为了衡量重叠的程度，其是由被遮挡的边界部分与细胞质的整个边界部分的长度比率来计算的。从表1的结果可以看出：本发明的方法得到了最好的分割结果，相比于其他方法，本发明的方法的精确度平均改进了大约5％。具体而言，当重叠度小于0.5(见表中的
Figure PCTCN2019091761-appb-000076
栏的内容)时，本发明的方法的精确度平均改进了大约3％；当重叠度大于0.5(见表中的
Figure PCTCN2019091761-appb-000077
栏的内容)时，精确度平均改进了大约8％。
此外，图5示出了根据本发明的方法和现有技术的方法得到的分割结果的比较图像。其中，图5-(a)是原始输入的图像，图5-(b)至图5-(f)分别依次表示由LSF、MCL、MPW、CF和本发明的方法获得的分割结果的图像，图5-(g)是实际的分割图像。从图5可以更直观地看出，本发明的分割方法(如图5-(f)所示)获得了最好的分割结果，对于其中部分细胞质实例，甚至与实际的分割结果(如图5-(g)所示)完全一致。
本发明的方法和系统克服了重叠区域中由于缺少强度信息而导致不能精确地分割细胞质的问题。相比于现有的基于先验形状的技术，本发明的方法和系统通过建立无限大的形状假设集、计算演进局部先验形状和整体先验形状、并对最终结果施加形状约束而提供了一种更加精确的重叠区域细胞质分割方法与系统。
本发明的方法和系统不仅限于用于宫颈癌检测中，而且本领域技术人员能够进行适当的改进，使本发明的方法和系统应用于定量测量细胞水平特征的其他显微镜检查图像，例如病理学图像测量。
尽管上面已经参考基于先验形状的检测重叠细胞质的分割方法及系统的具体实施例描述了本发明，但当然能想到，本领域普通技术人员可以推导出许多变型，因此，本领域普通技术人员容易想到的变型被认作本发明的一部分。本发明的范围在所附的权利要求书中限定。

Claims (18)

  1. 一种分割医学图像中的重叠细胞质的方法，其包括：
    建立细胞质形状假设集；以及
    从建立的细胞质形状假设集中为每个细胞质选择形状假设，来执行受约束的多形状演进，从而分割医学图像中的重叠细胞质，
    其中，所述受约束的多形状演进包括以下步骤：
    团块区域分割，其对由多个重叠细胞质构成的所述医学图像中的团块区域进行分割以提供团块证据；
    形状对齐，其用于评估选择的形状假设的质量；以及
    形状演进，其为每个细胞质确定更好的形状假设。
  2. 根据权利要求1所述的方法，还包括为所建立的细胞质形状假设集更新细胞质形状实例的学习形状实例重要度的步骤。
  3. 根据权利要求1或2所述的方法，其中，所述形状对齐和所述形状演进是迭代进行的，形状对齐的输出作为形状演进的输入，而形状演进的输出作为形状对齐的输入，所述形状对齐评估是否需要开始新的形状演进，一旦确定不需要新的形状演进，则形状对齐后的当前的形状假设就是该重叠细胞质的分割结果。
  4. 根据权利要求3所述的方法，其中，由以下公式建立所述细胞质形状假设集：
    Figure PCTCN2019091761-appb-100001
    其中，μ表示收集的形状实例的平均形状，Mx i表示收集的形状的协方差矩阵的特征向量的线性组合，其中，矩阵M的每一列是一个特征向量，x i是该线性组合的权重向量，s i表示标号为i的形状假设。
  5. 根据权利要求4所述的方法，其中，所述形状对齐包括填充从细胞质形状假设集中选择的形状假设以获得二值图像，并利用下述公式来获 得所述二值图像与对应的细胞质对齐所需要旋转的角度和缩放的大小：
    Figure PCTCN2019091761-appb-100002
    其中，B c表示分割后的团块区域的图像；B i表示对齐的结果图像，B i是与B c具有相同尺寸的二值图像，且需要B i在B c的内部。
  6. 根据权利要求5所述的方法，其中，所述形状演进包括：
    -设定目标函数，所述目标函数为：
    Figure PCTCN2019091761-appb-100003
    其中，x表示集合
    Figure PCTCN2019091761-appb-100004
    x k表示第k次演进的x；N表示所述团块区域中细胞质的数量；
    Figure PCTCN2019091761-appb-100005
    是由对齐结果
    Figure PCTCN2019091761-appb-100006
    生成的二值图像；(x，y)表示图像中像素的坐标；以及
    -确定能够使
    Figure PCTCN2019091761-appb-100007
    具有比x k更低的值的x k+1。
  7. 根据权利要求6所述的方法，其中，所述确定能够使
    Figure PCTCN2019091761-appb-100008
    具有比x k更低的值的x k+1包括：
    -对于从所述目标函数得到的矩阵p，使用泰勒定理得到公式：
    Figure PCTCN2019091761-appb-100009
    其中，
    Figure PCTCN2019091761-appb-100010
    和
    Figure PCTCN2019091761-appb-100011
    分别表示梯度和海森矩阵计算；γ是间隔(0,1)之间的一个标量，从而获得以x k为圆心、以||p|| 2为半径的区域内的
    Figure PCTCN2019091761-appb-100012
    的整体区域最小值；
    -回到x k处近似
    Figure PCTCN2019091761-appb-100013
    Figure PCTCN2019091761-appb-100014
    其中，m k表示x k处的
    Figure PCTCN2019091761-appb-100015
    -由m k的最小值来近似
    Figure PCTCN2019091761-appb-100016
    的最小值，并通过信赖域方法来解公式：
    Figure PCTCN2019091761-appb-100017
    以及
    -得到第k次演进的输出结果：x k+1＝x k+p *。
  8. 根据权利要求2所述的方法，其中，所述学习形状实例重要度的 步骤包括：
    -随机选择一组形状实例，根据以下公式计算所述一组形状实例的平均形状：
    Figure PCTCN2019091761-appb-100018
    其中，K a是选择的形状实例的数量；ω i是每个形状实例s i的重要度；W是所有ω i的和；并且，
    -根据得到的μ计算协方差矩阵：
    Figure PCTCN2019091761-appb-100019
    其中，矩阵M c的第一t个特征向量构成矩阵
    Figure PCTCN2019091761-appb-100020
    其对应的特征值为λ 1≥λ 2≥…≥λ t。
  9. 根据权利要求8所述的方法，其中，如果由所述受约束的多形状演进得到的分割结果大于预定阈值，则由所述学习形状实例重要度的步骤重新计算，以更新所述形状假设集；直到所述分割结果不能再减小或达到了所述预定阈值，才停止所述更新。
  10. 一种用于分割医学图像中的重叠细胞质的系统，其包括：
    用于建立细胞质形状假设集的形状假设集模块；以及
    多形状演进模块，其用于从建立的细胞质形状假设集中为每个细胞质选择形状假设，来进行受约束的多形状演进，从而分割医学图像中的重叠细胞质，
    其中，所述多形状演进模块被配置为执行以下步骤：
    团块区域分割，其对由多个重叠细胞质构成的所述医学图像中的团块区域进行分割以提供团块证据；
    形状对齐，其用于评估选择的形状假设的质量；以及
    形状演进，其为每个细胞质确定更好的形状假设。
  11. 根据权利要求10所述的系统，还包括为所述建立的细胞质形状假设集更新细胞质形状实例的学习形状实例重要度模块。
  12. 根据权利要求10或11所述的系统，其中，所述形状对齐和所述形状演进是迭代进行的，形状对齐的输出作为形状演进的输入，而形状演进的输出作为形状对齐的输入，所述形状对齐评估是否需要开始新的形状演进，一旦确定不需要新的形状演进，则形状对齐后的当前的形状假设就是该重叠细胞质的分割结果。
  13. 根据权利要求12所述的系统，其中，由以下公式建立所述细胞质形状假设集：
    Figure PCTCN2019091761-appb-100021
    其中，μ表示收集的形状实例的平均形状，Mx i表示收集的形状的协方差矩阵的特征向量的线性组合，其中，矩阵M的每一列是一个特征向量，x i是该线性组合的权重向量，s i表示标号为i的形状假设。
  14. 根据权利要求13所述的系统，其中，所述形状对齐包括填充从细胞质形状假设集中选择的形状假设以获得二值图像，并利用下述公式来获得所述二值图像与对应的细胞质对齐所需要旋转的角度和缩放的大小：
    Figure PCTCN2019091761-appb-100022
    其中，B c表示分割后的团块区域的图像；B i表示对齐的结果图像，B i是与B c具有相同尺寸的二值图像，且需要B i在B c的内部。
  15. 根据权利要求14所述的系统，其中，所述形状演进包括：
    -设定目标函数，所述目标函数为：
    Figure PCTCN2019091761-appb-100023
    其中，x表示集合
    Figure PCTCN2019091761-appb-100024
    x k表示第k次演进的x；N表示所述团块区域中细胞质的数量；
    Figure PCTCN2019091761-appb-100025
    是由对齐结果
    Figure PCTCN2019091761-appb-100026
    生成的二值图像；(x，y)表示图像中像素的坐标；以及
    -确定能够使
    Figure PCTCN2019091761-appb-100027
    具有比x k更低的值的x k+1。
  16. 根据权利要求15所述的系统，其中，所述确定能够使
    Figure PCTCN2019091761-appb-100028
    具有比x k更低的值的x k+1包括：
    -对于从所述目标函数得到的矩阵p，使用泰勒定理得到公式：
    Figure PCTCN2019091761-appb-100029
    其中，
    Figure PCTCN2019091761-appb-100030
    和
    Figure PCTCN2019091761-appb-100031
    分别表示梯度和海森矩阵计算；γ是间隔(0,1)之间的一个标量，从而获得以x k为圆心、以||p|| 2为半径的区域内的
    Figure PCTCN2019091761-appb-100032
    的整体区域最小值；
    -回到x k处近似
    Figure PCTCN2019091761-appb-100033
    Figure PCTCN2019091761-appb-100034
    其中，m k表示x k处的
    Figure PCTCN2019091761-appb-100035
    -由m k的最小值来近似
    Figure PCTCN2019091761-appb-100036
    的最小值，并通过信赖域方法来解公式：
    Figure PCTCN2019091761-appb-100037
    以及
    -得到第k次演进的输出结果：x k+1＝x k+P *。
  17. 根据权利要求11所述的系统，其中，所述学习形状实例重要度模块被配置为：
    -随机选择一组形状实例，根据以下公式计算所述一组形状实例的平均形状：
    Figure PCTCN2019091761-appb-100038
    其中，K a是选择的形状实例的数量；ω i是每个形状实例s i的重要度；W是所有ω i的和；并且，
    -根据得到的μ计算协方差矩阵：
    Figure PCTCN2019091761-appb-100039
    其中，矩阵M c的第一t个特征向量构成矩阵
    Figure PCTCN2019091761-appb-100040
    其对应的特征值为λ 1≥λ 2≥…≥λ t。
  18. 根据权利要求17所述的系统，其中，如果由所述受约束的多形状演进模块得到的分割结果大于预定阈值，则由所述学习形状实例重要度模块重新计算，以更新所述形状假设集；直到所述分割结果不能再减小或 达到了所述预定阈值，才停止所述更新。
PCT/CN2019/091761 2019-06-18 2019-06-18 用于分割医学图像中的重叠细胞质的方法及系统 WO2020252665A1 (zh)

Priority Applications (3)

Application Number Priority Date Filing Date Title
US17/596,382 US20220237782A1 (en) 2019-06-18 2019-06-18 Method and system for segmenting overlapping cytoplasm in medical image
CN201980000985.9A CN112470166A (zh) 2019-06-18 2019-06-18 用于分割医学图像中的重叠细胞质的方法及系统
PCT/CN2019/091761 WO2020252665A1 (zh) 2019-06-18 2019-06-18 用于分割医学图像中的重叠细胞质的方法及系统

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2019/091761 WO2020252665A1 (zh) 2019-06-18 2019-06-18 用于分割医学图像中的重叠细胞质的方法及系统

Publications (1)

Publication Number Publication Date
WO2020252665A1 true WO2020252665A1 (zh) 2020-12-24

Family

ID=74037189

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/091761 WO2020252665A1 (zh) 2019-06-18 2019-06-18 用于分割医学图像中的重叠细胞质的方法及系统

Country Status (3)

Country Link
US (1) US20220237782A1 (zh)
CN (1) CN112470166A (zh)
WO (1) WO2020252665A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112614142A (zh) * 2020-12-25 2021-04-06 华侨大学 一种基于多通道图像融合的细胞弱标签制作方法及系统

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116564418B (zh) * 2023-04-20 2024-06-11 深圳湾实验室 细胞类群相关性网络构建方法和装置、设备及存储介质
CN117253228B (zh) * 2023-11-14 2024-02-09 山东大学 一种基于核像距离内编码的细胞团簇空间约束方法及系统
CN117576098B (zh) * 2024-01-16 2024-04-19 武汉互创联合科技有限公司 基于分割的细胞分裂均衡度评估方法及设备

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9292933B2 (en) * 2011-01-10 2016-03-22 Anant Madabhushi Method and apparatus for shape based deformable segmentation of multiple overlapping objects

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP2779089B1 (en) * 2010-07-30 2018-11-14 Fundação D. Anna Sommer Champalimaud E Dr. Carlos Montez Champalimaud Systems and methods for segmentation and processing of tissue images and feature extraction from same for treating, diagnosing, or predicting medical conditions
CN103310449B (zh) * 2013-06-13 2016-03-02 沈阳航空航天大学 基于改进形状模型的肺分割方法
CN106611416B (zh) * 2015-10-19 2020-06-26 上海联影医疗科技有限公司 一种医学图像中肺分割的方法及装置
CN105549079A (zh) * 2016-01-12 2016-05-04 中国矿业大学(北京) 一种地球物理参数的全波形反演模型的建立方法和装置
WO2018089783A1 (en) * 2016-11-11 2018-05-17 University Of South Florida Automated stereology for determining tissue characteristics
CN106548521A (zh) * 2016-11-24 2017-03-29 北京三体高创科技有限公司 一种联合2d+3d主动外观模型的人脸对齐方法及系统
CN108334860A (zh) * 2018-03-01 2018-07-27 北京航空航天大学 细胞图像的处理方法和装置
CN108596932A (zh) * 2018-04-18 2018-09-28 哈尔滨理工大学 一种重叠宫颈细胞图像分割方法

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9292933B2 (en) * 2011-01-10 2016-03-22 Anant Madabhushi Method and apparatus for shape based deformable segmentation of multiple overlapping objects

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
AFAF TAREEF, YANG SONG, WEIDONG CAI, HENG HUANG, HANG CHANG, YUE WANG, MICHAEL FULHAM, DAGAN FENG, MEI CHEN: "Automatic segmentation of overlapping cervical smear cells based on local distinctive features and guided shape deformation", NEUROCOMPUTING, ELSEVIER, AMSTERDAM, NL, vol. 221, 1 January 2017 (2017-01-01), AMSTERDAM, NL, pages 94 - 107, XP055767280, ISSN: 0925-2312, DOI: 10.1016/j.neucom.2016.09.070 *
SONG YOUYI; ZHU LEI; QIN JING; LEI BAIYING; SHENG BIN; CHOI KUP-SZE: "Segmentation of Overlapping Cytoplasm in Cervical Smear Images via Adaptive Shape Priors Extracted From Contour Fragments", IEEE TRANSACTIONS ON MEDICAL IMAGING, IEEE SERVICE CENTER, PISCATAWAY, NJ., US, vol. 38, no. 12, 1 December 2019 (2019-12-01), US, pages 2849 - 2862, XP011758299, ISSN: 0278-0062, DOI: 10.1109/TMI.2019.2915633 *
SONG, YOUYI ET AL.: "Accurate Cervical Cell Segmentation from Overlapping Clumps in Pap Smear Images", IEEE TRANSACTIONS ON MEDICAL IMAGING, vol. 36, no. 1,, 31 January 2017 (2017-01-31), XP011638276, ISSN: 0278-0062, DOI: 20200217083858A *
TAREEF AFAF; SONG YANG; HUANG HENG; WANG YUE; FENG DAGAN; CHEN MEI; CAI WEIDONG: "Optimizing the cervix cytological examination based on deep learning and dynamic shape modeling", NEUROCOMPUTING, ELSEVIER, AMSTERDAM, NL, vol. 248, 8 March 2017 (2017-03-08), AMSTERDAM, NL, pages 28 - 40, XP029997982, ISSN: 0925-2312, DOI: 10.1016/j.neucom.2017.01.093 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112614142A (zh) * 2020-12-25 2021-04-06 华侨大学 一种基于多通道图像融合的细胞弱标签制作方法及系统
CN112614142B (zh) * 2020-12-25 2023-05-30 华侨大学 一种基于多通道图像融合的细胞弱标签制作方法及系统

Also Published As

Publication number Publication date
CN112470166A (zh) 2021-03-09
US20220237782A1 (en) 2022-07-28

Similar Documents

Publication Publication Date Title
WO2020252665A1 (zh) 用于分割医学图像中的重叠细胞质的方法及系统
WO2021009258A1 (en) Object detection and instance segmentation of 3d point clouds based on deep learning
Wan et al. Accurate segmentation of overlapping cells in cervical cytology with deep convolutional neural networks
WO2021203795A1 (zh) 一种基于显著性密集连接扩张卷积网络的胰腺ct自动分割方法
WO2018108129A1 (zh) 用于识别物体类别的方法及装置、电子设备
CN112069874B (zh) 胚胎光镜图像中细胞的识别方法及系统、设备及存储介质
CN112580416A (zh) 基于深暹罗网络和贝叶斯优化的视频跟踪
CN112435243A (zh) 一种全切片数字病理图像的自动分析系统及方法
CN109492706B (zh) 一种基于循环神经网络的染色体分类预测装置
WO2020238256A1 (zh) 基于弱分割的损伤检测方法及装置
WO2010111916A1 (zh) 多类目标的检测装置及检测方法
TWI701608B (zh) 用於圖片匹配定位的神經網路系統、方法及裝置
CN111080596A (zh) 一种融合局部阴影与全局特征的尘肺病辅助筛查方法及系统
CN112037180B (zh) 染色体分割方法及装置
CN111524132A (zh) 识别待检测样本中异常细胞的方法、装置和存储介质
CN113256634A (zh) 基于深度学习的宫颈癌tct切片排阴方法和系统
US11521323B2 (en) Systems and methods for generating bullseye plots
Larrazabal et al. Maximum Entropy on Erroneous Predictions (MEEP): Improving model calibration for medical image segmentation
CN113327666A (zh) 一种胸片疾病多分类网络的多标签局部至全局学习方法
CN110717916B (zh) 一种基于卷积神经网络的肺栓塞检测系统
CN108765384B (zh) 一种联合流形排序和改进凸包的显著性检测方法
Sarmadian et al. Optimizing the snake model using honey-bee mating algorithm for road extraction from very high-resolution satellite images
Zhao et al. Overlapping region reconstruction in nuclei image segmentation
CN112861878B (zh) 基于结构偏移特征的异常匹配识别方法
US20230343438A1 (en) Systems and methods for automatic image annotation

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19933928

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19933928

Country of ref document: EP

Kind code of ref document: A1