WO2020019671A1 - 一种乳腺肿块检测与分类***、计算机可读存储介质 - Google Patents

一种乳腺肿块检测与分类***、计算机可读存储介质 Download PDF

Info

Publication number
WO2020019671A1
WO2020019671A1 PCT/CN2018/124655 CN2018124655W WO2020019671A1 WO 2020019671 A1 WO2020019671 A1 WO 2020019671A1 CN 2018124655 W CN2018124655 W CN 2018124655W WO 2020019671 A1 WO2020019671 A1 WO 2020019671A1
Authority
WO
WIPO (PCT)
Prior art keywords
pixels
breast
pixel
mass
breast image
Prior art date
Application number
PCT/CN2018/124655
Other languages
English (en)
French (fr)
Inventor
徐勇
刘宏
Original Assignee
哈尔滨工业大学（深圳）
北京大学深圳研究生院
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 哈尔滨工业大学（深圳）, 北京大学深圳研究生院 filed Critical 哈尔滨工业大学（深圳）
Publication of WO2020019671A1 publication Critical patent/WO2020019671A1/zh

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
    • 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
    • G06T2207/30068—Mammography; Breast

Definitions

  • the invention relates to the medical field, in particular to a breast mass detection and classification system and a computer-readable storage medium.
  • Mammography mammography images are widely used mammography images, which have the advantages of low cost, high quality, and high cost performance. This type of image mainly reflects the contours of breast tissue. Using mammography images of mammary glands, doctors can better identify breast masses and judge the nature of the masses. However, doctors' manual discrimination has the problems of relying on subjective experience and different discrimination results between different doctors. Due to the development of computer vision understanding technology, automatic identification of breast masses and their qualitative computer becomes possible.
  • breast mass detection and classification Most of the existing methods for breast mass detection and identification are divided into two steps: breast mass detection and classification.
  • the current method is implemented in two steps mainly as follows: It is generally believed that breast tumors have a relatively special texture structure, and tumors lack clear edges. Therefore, breast tumor detection is a task that is quite different from ordinary target detection tasks. The texture features in ordinary target edges generally have significant features. Therefore, people tend to use special methods to detect tumors before classifying them. Such a processing method has the disadvantage of high computational complexity. Moreover, these two independent steps cut off the intrinsic link between the two closely linked tasks of breast mass detection and benign and malignant mass classification; the separate use of these two steps also limits the overall performance of the method. If this The accuracy of both steps is 80%, and the overall accuracy is only 64%. Therefore, it is urgent to explore new methods and technical routes to overcome the shortcomings of current methods. Collins
  • an object of the present invention is to provide a breast mass detection and classification system and a computer-readable storage medium for implementing breast mass detection and classification.
  • a breast mass detection and classification system including:
  • An image acquisition unit configured to acquire a breast image
  • a pixel classification unit configured to classify each pixel of the breast image, and the types of the pixels include ordinary pixels, benign lumps pixels, and malignant lumps pixels;
  • a detection and classification unit is configured to detect and classify a breast mass from a pixel-classified breast image.
  • the breast mass detection and classification system further includes a neural network unit, and the neural network unit includes:
  • a network construction module configured to construct a deep neural network, which is used to classify each pixel of the breast image, and the types of the pixels include ordinary pixels, benign tumor pixels, and malignant tumor pixels; the depth
  • the neural network is divided into six layers, which are a first convolution layer, a second convolution layer, a third convolution layer, a first fully connected layer, a second fully connected layer, and a network output layer in this order.
  • the deep neural network The number of neurons in each layer is the same as the number of pixels in the breast image;
  • a network training module is configured to train the deep neural network by using a breast image sample set.
  • the first convolutional layer, the second convolutional layer, the third convolutional layer, and the first fully connected layer adopt an improved LU activation function
  • the improved LU activation function is a calculation value of the LU activation function smaller than When it is equal to a preset value, the calculated value of the LU activation function is set to 0.
  • the second fully connected layer uses a sigmoid activation function.
  • the breast image is calculated through the first convolution layer, the second convolution layer, the third convolution layer, the first fully connected layer, and the second fully connected layer to obtain a scalar corresponding to each pixel;
  • the network output layer is used to calculate the absolute value of the difference between the scalar and the pixel digital class.
  • the pixel digital class includes a normal pixel class, a benign mass pixel class, and a malignant mass pixel class.
  • the pixel class of the pixel corresponding to the smallest difference is output as the class of the pixel.
  • the normal pixel class is labeled -1
  • the benign mass pixel class is labeled 0
  • the malignant mass pixel class is labeled 1.
  • the network training module includes
  • An image processing sub-module configured to process the breast image sample set, where the breast image sample set includes multiple breast image samples, and a method for processing the breast image sample set includes:
  • Transforming the pixel gray value of the breast image sample includes increasing the gray value of the pixels of the breast image sample by the same proportion or reducing the gray value of the pixels of the breast image sample by the same proportion;
  • a training sub-module is configured to train the deep neural network by using a breast image sample set processed by the image processing sub-module.
  • the detection and classification unit includes
  • a region division module configured to divide a pixel-classified breast image into a plurality of partially overlapping subregions, the subregions having different sizes
  • a mass detection module is configured to determine whether a proportion of benign mass pixels of the sub-region to the total pixels of the sub-region or a proportion of malignant mass pixels to the total pixels of the sub-region is greater than a preset ratio. When the determination result is yes When determining that the sub-region is a breast lump;
  • a tumor classification module is configured to determine whether the proportion of the benign tumor pixels in the total pixels of the sub-region is greater than the proportion of the malignant tumor pixels in the total pixels of the sub-region.
  • a benign breast mass and conversely, the subregion is a malignant breast mass.
  • the preset ratio is 0.3.
  • Another technical solution adopted by the present invention is: a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:
  • Classify each pixel of the breast image the types of the pixels include ordinary pixels, benign lumps pixels, and malignant lumps pixels;
  • the invention relates to a breast mass detection and classification system and a computer-readable storage medium.
  • the pixels are directly classified into three categories: ordinary pixels, benign mass pixels, and malignant mass pixels, to achieve fast and accurate Performing mass detection and mass classification on breast images, overcoming the two steps of splitting mass detection and classification in the prior art, leads to technical problems of low accuracy and low efficiency.
  • FIG. 1 is a schematic structural diagram of a specific embodiment of a breast mass detection and classification system in the present invention
  • FIG. 2 is a function curve diagram of a specific embodiment of an improved LU activation function in the present invention.
  • FIG. 3 is a schematic diagram of a specific embodiment for increasing the gray value of pixels in the same proportion according to the present invention.
  • FIG. 4 is a schematic diagram of a specific embodiment of reducing gray values of pixels in the same proportion according to the present invention.
  • FIG. 1 is a schematic structural diagram of a specific embodiment of a breast mass detection and classification system according to the present invention. including:
  • An image acquisition unit is used to acquire a breast image.
  • the breast image is a mammography target X-ray radiographic image obtained by using an X-ray apparatus.
  • the original mammography target X-ray radiographic image may be different in size. Therefore, the breast of the present invention
  • the image is scaled to a size of 200 pixels by 200 pixels, and then entered into a pixel classification unit.
  • a pixel classification unit configured to classify each pixel of a breast image, and the types of pixels include ordinary pixels, benign tumor pixels, and malignant tumor pixels;
  • a detection and classification unit is configured to detect and classify a breast mass from a pixel-classified breast image.
  • the image acquisition unit, the pixel classification unit, and the detection and classification unit may be computers.
  • the computer uses a computer program to implement image acquisition, pixel classification, and mass detection and classification.
  • the image acquisition unit may directly acquire a breast image stored in a computer in advance or sent to the computer in real time, or may acquire a breast image taken by the breast image by communicating with the X-ray apparatus.
  • the pixel classification unit in the present invention is directly based on the breast image, and performs pixel classification discrimination on a single pixel of the breast image.
  • the detection and classification unit can further simultaneously locate a breast tumor based on the pixel discrimination result and identify the tumor type (benign breast tumor or Malignant breast mass) to achieve fast and effective breast mass detection and classification.
  • the pixel classification unit uses a deep neural network to classify pixels of a breast image. Therefore, before classifying a pixel of a breast image, a deep neural network needs to be designed first.
  • the breast mass detection and classification system also includes a neural network unit.
  • the neural network unit includes:
  • a network building module is used to build a deep neural network.
  • the deep neural network is used to classify each pixel of the breast image.
  • the types of pixels include ordinary pixels, benign lumps, and malignant lumps.
  • the deep neural network is divided into six layers. The order is the first convolution layer, the second convolution layer, the third convolution layer, the first fully connected layer, the second fully connected layer, and the network output layer. The number of neurons in each layer of the deep neural network and The number of pixels in the breast image is the same;
  • a network training module for training a deep neural network using a breast image sample set.
  • the first convolution layer, the second convolution layer, the third convolution layer, and the first fully connected layer adopt an improved LU activation function.
  • the improved LU activation function is that the calculated value of the LU activation function is less than or equal to a preset value.
  • the calculated value of the LU activation function is set to 0.
  • the second fully connected layer uses a sigmoid activation function.
  • the sigmoid activation function enhances the non-linear transformation capabilities of the "classifier.”
  • the deep neural network in the present invention does not have a pooling layer, mainly to reduce the loss of information.
  • the breast image is calculated through the first convolution layer, the second convolution layer, the third convolution layer, the first fully connected layer, and the second fully connected layer to obtain a scalar corresponding to each pixel; the network output layer is used for calculation The absolute value of the difference between a scalar and a pixel number class.
  • the pixel number class includes ordinary pixel class, benign mass pixel class, and malignant mass pixel class, and the pixel number class corresponding to the smallest absolute value.
  • the output is used as the pixel's class standard, and it is quantized to one of the three based on the similarity of the scalar with the three class targets.
  • the normal pixel class is labeled -1
  • the benign mass pixel class is labeled 0
  • the malignant mass pixel class is labeled 1.
  • the settings of the three categories of -1, 0, and 1 are also in line with the visual description of the differences between the three pixel categories.
  • the difference between -1 and 0 is less than the difference between -1 and 1. It is consistent with ordinary pixels and benign mass pixels. The fact that the difference between them is smaller than the difference between ordinary pixels and malignant mass pixels.
  • the network output layer calculates the absolute value of the difference between 3 and -1, 0, and 1 and compares 3 The absolute value of the absolute value, select the category corresponding to the difference with the smallest absolute value as the category of this pixel.
  • the smallest absolute value is the absolute value of the difference between 3 and 1, then this pixel
  • the class is labeled 1.
  • the corresponding pixel type of the breast image is output in the network output layer; that is, the final output of the neurons in the network output layer is -1, 0, or 1.
  • the corresponding pixels of the input breast image are respectively Discriminated as normal pixels, benign mass pixels or malignant mass pixels.
  • the deep neural network is constructed, it is trained using a network training module in order to better implement pixel classification processing of breast images.
  • the network training module includes
  • An image processing sub-module is used to process a breast image sample set.
  • the breast image sample set includes multiple breast image samples.
  • the breast image samples have been labeled with tumors and their types in the breast image.
  • the method for processing the breast image sample set includes:
  • Transforming the pixel gray value of the breast image sample including increasing the gray value of the pixels of the breast image sample by the same proportion or reducing the gray value of the pixels of the breast image sample by the same proportion;
  • a training sub-module is used to train a deep neural network using the breast image sample set processed by the image processing sub-module, and input the processed breast image sample set to the deep neural network for pixel classification to implement network training.
  • the present invention uses limited image samples during the training phase to generate as many breast image samples as possible, enhancing the diversity of breast image sample sets, as they can represent different breast molybdenum targets
  • the image results of the instrument can enrich the "experience" of the neural network, enhance the adaptability of the training results to the samples, and the neural network after training will have better robustness.
  • FIG. 3 is a schematic diagram of a specific embodiment of increasing the gray value of the pixels in the same proportion in the present invention; When it reaches 255, the pixel value is increased by the same proportion (as shown by the y coordinate).
  • the formula for reducing the gray value in the same proportion is c + dx, c and d are coefficients and d is greater than 0 but less than 1, and x is still the original pixel value.
  • FIG. 4 is a schematic diagram of a specific embodiment of reducing the gray value of the pixels in the same proportion in the present invention; When it reaches 255, the pixel value becomes smaller in proportion (as shown in the y coordinate).
  • the above four schemes describe the variability of the samples from different angles.
  • the first scheme can simulate the systematic differences of different models of X-ray instruments.
  • the two methods of increasing and decreasing the pixel value respectively make the new sample better. Representativeness and biased in two possible directions of change.
  • the second, third, and fourth schemes can simulate the random errors of the instruments, and have better simulation capabilities for the differences between the same model of instruments. Using these four schemes at the same time makes the pixel change in the new sample generated more comprehensive.
  • the image processing sub-module modifies the pixels of the original breast image to increase the diversity of the breast image sample set.
  • the pixel classification unit uses the trained neural network to classify the pixels of the breast image to obtain a pixel class label corresponding to each pixel.
  • the detection and classification unit performs breast mass detection and classification on the breast image classified by the pixels.
  • the detection and classification unit includes:
  • a region division module is used to divide a pixel-classified breast image into a plurality of partially overlapping sub-regions.
  • the sub-regions have different sizes.
  • the sub-regions are rectangular regions, and the dimensions of the rectangular regions are different (such as 7 pixels * 7 pixels, 9 pixels * 9 pixels, 11 pixels * 11 pixels, 13 pixels * 13 pixels); because the actual breast masses are of different sizes, the detection and classification system of the present invention judges sub-regions of different sizes to make the detected breast masses more realistic. If only the input breast image is divided into non-overlapping sub-regions, and the sub-regions have the same size, when the breast mass crosses two adjacent sub-regions, it is likely to miss detection.
  • a mass detection module is used to determine whether the proportion of benign mass pixels in the sub-region to the total pixels of the sub-region or the proportion of malignant mass pixels in the total pixels of the sub-region is greater than a preset ratio.
  • the determination result is yes, the sub-region is determined to be Breast lump; in this embodiment, the preset ratio is 0.3.
  • a tumor classification module is used to determine whether the proportion of benign tumor pixels in the total area of the sub-region is greater than that of malignant tumor pixels in the total area of the sub-region. If the result of the determination is yes, the sub-region is a benign breast tumor. For malignant breast mass.
  • the mass detection module and mass classification module perform mass detection and classification on each sub-region of the breast image, and directly perform mass detection and classification based on breast image pixels, which is efficient, convenient, and highly accurate.
  • the present invention also provides a computer-readable storage medium on which a computer program is stored.
  • a computer program is stored on a computer-readable storage medium.
  • Classify each pixel of the breast image the types of pixels include ordinary pixels, benign lumps pixels and malignant lumps pixels;
  • the present invention directly identifies pixels of breast images and determines them as ordinary pixels, benign mass pixels, There are three categories of pixels for malignant masses; the categories of pixels for the breast image output by the neural network; the location of breast masses and the discrimination results of benign or malignant masses are obtained according to the pixel class.

Landscapes

  • Engineering & Computer Science (AREA)
  • Quality & Reliability (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Apparatus For Radiation Diagnosis (AREA)
  • Image Processing (AREA)

Abstract

一种乳腺肿块检测与分类***、计算机可读存储介质，通过对乳腺图像的每一个像素进行分类，直接将像素分为三类：普通像素、良性肿块像素和恶性肿块像素，实现快速准确地对乳腺图像进行肿块检测和肿块分类，克服现有技术割裂肿块检测和分类两个步骤，导致准确率低下且效率低的技术问题。

Description

一种乳腺肿块检测与分类***、计算机可读存储介质
技术领域
本发明涉及医学领域，尤其是一种乳腺肿块检测与分类***、计算机可读存储介质。
背景技术
乳腺钼靶X线摄影图像是广泛使用的乳腺检测图像，具有成本低，质量高和高性价比的优点。该类图像主要反映乳腺组织的等密度线。利用乳腺钼靶X线摄影图像，医生可以较好地判别出乳腺肿块以及对肿块性质进行判断。但是，医生人工判别存在着依赖主观经验，以及不同医生之间的判别结果不同的问题。由于计算机视觉理解技术的发展，乳腺肿块及其定性的计算机自动判识成为可能。
绝大部分已有的乳腺肿块检测与判识方法分为乳腺肿块检测与分类两个步骤。目前方法分两个步骤实施的主要因为如下：人们普遍认为乳腺肿块有相对特殊的纹理结构，肿块缺乏比较清晰的边缘，因此乳腺肿块检测是一个与普通目标检测任务区别较大的任务。普通的目标边缘内纹理特征一般具有显著的特点。因此，人们倾向于使用特别的方法检测出肿块后再分类。这样的处理方法存在计算复杂度高的缺点。而且，这两个独立的步骤割裂了乳腺肿块检测与良性与恶性肿块分类这两个紧密联系的任务之间的内在联系；这两个步骤的分别使用也使得方法的整体性能得到限制，假如这两个步骤的准确率均为80%，则整体上最终的准确率仅有64%。因此，急需探索新的方法与技术路线，克服目前方法的缺陷。
发明内容
本发明旨在至少在一定程度上解决相关技术中的技术问题之一。为此，本发明的一个目的是提供一种乳腺肿块检测与分类***、计算机可读存储介质，用于实现乳腺肿块检测和分类。
本发明所采用的技术方案是：一种乳腺肿块检测与分类***，包括
图像获取单元，用于获取乳腺图像；
像素分类单元，用于对所述乳腺图像的每一个像素进行分类，所述像素的类别包括普通像素、良性肿块像素和恶性肿块像素；
检测与分类单元，用于对像素分类后的乳腺图像进行乳腺肿块检测和分类。
进一步地，所述乳腺肿块检测与分类***还包括神经网络单元，所述神经网络单元包括：
网络构建模块，用于构建深度神经网络，所述深度神经网络用于对所述乳腺图像的每一个像素进行分类，所述像素的类别包括普通像素、良性肿块像素和恶性肿块像素；所述深度神经网络分为六层，按照顺序依次为第一卷积层、第二卷积层、第三卷积层、第一全连接层、第二全连接层和网络输出层，所述深度神经网络的每一层的神经元个数与乳腺图像的像素个数相同；
网络训练模块，用于利用乳腺图像样本集训练所述深度神经网络。
进一步地，所述第一卷积层、第二卷积层、第三卷积层和第一全连接层采用改进的LU激活函数，所述改进的LU激活函数为LU激活函数的计算值小于或等于预设值时，将所述LU激活函数的计算值设置为0。
进一步地，所述第二全连接层采用sigmoid激活函数。
进一步地，所述乳腺图像经过所述第一卷积层、第二卷积层、第三卷积层、第一全连接层和第二全连接层计算得到对应每个像素的标量；
所述网络输出层用于计算所述标量与像素数字类标的差值的绝对值，所述像素数字类标分别包括普通像素类标、良性肿块像素类标和恶性肿块像素类标，并将绝对值最小的差值所对应的像素数字类标作为像素的类标进行输出。
进一步地，所述普通像素类标为-1，所述良性肿块像素类标为0，所述恶性肿块像素类标为1。
进一步地，所述网络训练模块包括
图像处理子模块，用于对所述乳腺图像样本集进行处理，所述乳腺图像样本集包括多个乳腺图像样本，处理所述乳腺图像样本集的方法包括：
对所述乳腺图像样本进行像素灰度值进行变换，包括对所述乳腺图像样本的像素进行同比例增大灰度值或对所述乳腺图像样本的像素进行同比例减小灰度值；
和/或，
对全部或者部分的所述乳腺图像样本的像素添加高斯噪声；
和/或，
对全部或者部分的所述乳腺图像样本的像素添加椒盐噪声；
训练子模块，用于利用所述图像处理子模块处理后的乳腺图像样本集训练所述深度神经网络。
进一步地，所述检测与分类单元包括
区域划分模块，用于将像素分类后的乳腺图像划分成多个部分重叠的子区域，所述子区域的大小不同；
肿块检测模块，用于判断所述子区域的良性肿块像素占所述子区域的总像素的比例或恶性肿块像素占所述子区域的总像素的比例是否大于预设比例，当判断结果为是时，判断所述子区域为乳腺肿块；
肿块分类模块，用于判断所述良性肿块像素占所述子区域的总像素的比例是否大于恶性肿块像素占所述子区域的总像素的比例，若判断结果为是，则所述子区域为良性乳腺肿块，反之，所述子区域为恶性乳腺肿块。
进一步地，所述预设比例为0.3。
本发明所采用的另一技术方案是：一种计算机可读存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时实现以下步骤：
获取乳腺图像；
对所述乳腺图像的每一个像素进行分类，所述像素的类别包括普通像素、良性肿块像素和恶性肿块像素；
对像素分类后的乳腺图像进行乳腺肿块检测和分类。
本发明的有益效果是：
本发明一种乳腺肿块检测与分类***、计算机可读存储介质，通过对乳腺图像的每一个像素进行分类，直接将像素分为三类：普通像素、良性肿块像素和恶性肿块像素，实现快速准确地对乳腺图像进行肿块检测和肿块分类，克服现有技术割裂肿块检测和分类两个步骤，导致准确率低下且效率低的技术问题。
附图说明
下面结合附图对本发明的具体实施方式作进一步说明：
图1是本发明中一种乳腺肿块检测与分类***的一具体实施例结构示意图；
图2是本发明中改进的LU激活函数的一具体实施例函数曲线图；
图3是本发明中对像素进行同比例增大灰度值的一具体实施例示意图；
图4是本发明中对像素进行同比例减小灰度值的一具体实施例示意图。
具体实施方式
需要说明的是，在不冲突的情况下，本申请中的实施例及实施例中的特征可以相互组合。
一种乳腺肿块检测与分类***，参考图1，图1是本发明中一种乳腺肿块检测与分类***的一具体实施例结构示意图；包括：
图像获取单元，用于获取乳腺图像；本实施例中，乳腺图像为利用X射线仪获取的乳腺钼靶X线摄影图像，原始的乳腺钼靶X线摄影图像可能大小不同，因此，本发明乳腺图像缩放到200像素*200像素的大小，然后输入像素分类单元。
像素分类单元，用于对乳腺图像的每一个像素进行分类，像素的类别包括普通像素、良性肿块像素和恶性肿块像素；
检测与分类单元，用于对像素分类后的乳腺图像进行乳腺肿块检测和分类。本实施例中，图像获取单元、像素分类单元和检测与分类单元可以为计算机，计算机利用计算机程序实现图像获取、像素分类、肿块检测和分类。图像获取单元可以直接获取预先存储在计算机中或者实时送入计算机中的乳腺图像，也可以通过与X射线仪通信以获取其拍摄的乳腺图像。
本发明中的像素分类单元直接基于乳腺图像，对乳腺图像的单个像素进行像素类别判别，检测与分类单元进而可以基于像素判别结果同时定位出乳腺肿块并给肿块类型判识结果（良性乳腺肿块或恶性乳腺肿块），实现快速有效的乳腺肿块检测和分类。
作为技术方案的进一步改进，本实施例中，像素分类单元是利用深度神经网络对乳腺图像进行像素分类，因此，在对乳腺图像进行像素分类前，需要先设计深度神经网络。乳腺肿块检测与分类***还包括神经网络单元，神经网络单元包括：
网络构建模块，用于构建深度神经网络，深度神经网络用于对乳腺图像的每一个像素进行分类，像素的类别包括普通像素、良性肿块像素和恶性肿块像素；深度神经网络分为六层，按照顺序依次为第一卷积层、第二卷积层、第三卷积层、第一全连接层、第二全连接层和网络输出层，深度神经网络的每一层的神经元个数与乳腺图像的像素个数相同；
网络训练模块，用于利用乳腺图像样本集训练深度神经网络。
其中，第一卷积层、第二卷积层、第三卷积层和第一全连接层采用改进的LU激活函数，改进的LU激活函数为LU激活函数的计算值小于或等于预设值时，将LU激活函数的计算值设置为0。具体设置如下：如果z>e(Z表示当前卷积计算的结果，e为大于零的较小的常数), 则f(z)=z, 否则，f(z)=0。假设e=1，则z>e时f(z)=z的函数曲线如图2所示，图2是本发明中改进的LU激活函数的一具体实施例函数曲线图。另外，网络中总是大的输出值起较大作用，而小的输出值作用较小。而深度网络常用的增强网络鲁棒性的方法为阻断其中一些神经元之间的连接，人为达到连接的稀疏性。阻断连接与将神经元的输出值设为0实际上是相同的效果，但后者更简单，不用增加任何计算量。本发明为了增强网络稀疏性，提出了上述改进的LU激活函数，以简便的方式使得网络具有结构稀疏的特点，从而增强肿块检测与分类结果的鲁棒性。而第二全连接层采用sigmoid激活函数。sigmoid激活函数可增强“分类程序”的非线性变换能力。与传统的卷积神经网络不同，本发明中的深度神经网络不设pooling层，主要是为了减小信息的损失。
具体地，乳腺图像经过第一卷积层、第二卷积层、第三卷积层、第一全连接层和第二全连接层计算得到对应每个像素的标量；网络输出层用于计算标量与像素数字类标的差值的绝对值，像素数字类标分别包括普通像素类标、良性肿块像素类标和恶性肿块像素类标，并将绝对值最小的差值所对应的像素数字类标作为像素的类标进行输出，根据标量与三个类标的相似性将其量化为三者之一。本实施例中，普通像素类标为-1，良性肿块像素类标为0，恶性肿块像素类标为1。这使得乳腺肿块检测与肿块分类的步骤合二为一，且能同时提升训练与应用阶段的计算效率。此外，-1、0、1三个类标的设置也符合三个像素类别之间差异的直观描述，-1与0之间差值小于-1与1之间差值符合普通像素与良性肿块像素之间差异小于普通像素与恶性肿块像素之间差异的事实。举个例子，假设一个像素输入神经网络后，在第二全连接层最后输出的标量为3，则网络输出层计算3与-1、0、1之间的差值的绝对值，并比较3个绝对值的大小，选择绝对值最小的那个差值对应的类标作为这个像素的类标，本实施例中，绝对值最小的是3与1之间的差值的绝对值，则这个像素的类标为1。综上，乳腺图像输入神经网络后，在网络输出层输出乳腺图像的相应像素的类别；即网络输出层的神经元的最终输出为-1，0或1，输入的乳腺图像的相应像素分别被判别为普通像素、良性肿块像素或者恶性肿块像素。
作为技术方案的进一步改进，在深度神经网络构建完成后，利用网络训练模块对其进行训练，以便更好地实现对乳腺图像的像素分类处理。具体地，网络训练模块包括
图像处理子模块，用于对乳腺图像样本集进行处理，乳腺图像样本集包括多个乳腺图像样本，乳腺图像样本已标注了乳腺图像中的肿块及其类别，处理乳腺图像样本集的方法包括：
对乳腺图像样本进行像素灰度值进行变换，包括对乳腺图像样本的像素进行同比例增大灰度值或对乳腺图像样本的像素进行同比例减小灰度值；
和/或，
对全部或者部分的乳腺图像样本的像素添加高斯噪声；
和/或，
对全部或者部分的乳腺图像样本的像素添加椒盐噪声；
训练子模块，用于利用图像处理子模块处理后的乳腺图像样本集训练深度神经网络，将处理后的乳腺图像样本集输入深度神经网络进行像素分类以实现对网络的训练。
在实际应用中，不同厂家的X射线仪得到的同一个乳腺的影像具有较大差异，甚至同一型号的X射线仪得到的影像也有差异的现象，此外，实际应用中不可能获取所有厂家的乳腺钼靶仪的大量摄影图像，也不可能获取同一型号的乳腺钼靶仪的大量图像样例。为了增强***对乳腺图像差异的适应性，本发明在训练阶段利用有限的图像样例，生成尽可能多的乳腺图像样本，增强乳腺图像样本集的多样性，由于它们可以表示不同的乳腺钼靶仪的影像结果，所以可以丰富神经网络的“阅历”，增强训练结果对样本的适应性，训练后的神经网络将有较好的鲁棒性。通过对乳腺图像样本的像素进行处理修改，以获得更多的乳腺图像样本；通过同时使用上述像素灰度值修改和噪声添加的修改方法，产生的样例更多，效果更好。具体方式如下：
1.对乳腺图像样本的像素灰度值进行变换，分别对像素进行同比例增大灰度值与同比例减小灰度值，以获得新的样例。同比例增大灰度值的方式中，如果像素修改后的值大于255，则强制设置为255。同比例增大灰度值的公式为a+bx，a、b为系数且b大于1，x为原像素值。对像素进行同比例增大灰度值的结果如图3所示，图3是本发明中对像素进行同比例增大灰度值的一具体实施例示意图；图3中给出x从0变化至255时，像素值同比例增大的结果（如y坐标所示）。同比例减小灰度值的公式为c+dx，c、d为系数且d大于0但小于1，x仍为原像素值。对像素进行同比例减小灰度值的结果如图4所示，图4是本发明中对像素进行同比例减小灰度值的一具体实施例示意图；图4中给出x从0变化至255时，像素值同比例变小的结果（如y坐标所示）。
2.对所有乳腺图像样本的像素均添加一定程度的高斯噪声。
3.对所有乳腺图像样本的像素均添加一定程度的椒盐噪声。
4.对一个乳腺图像样本的图像中的像素随机添加高斯或椒盐噪声。
上述4个方案从不同角度刻画了样本的可变性，其中第一个方案可模拟不同型号X射线仪的***性差异，分别增大和减小像素值两种方式使得得出的新样本有更好的代表性，并在两个可能的变化方向上都不失偏颇。第二、三、四个方案可模拟仪器的随机误差，对同一型号仪器间的差异有比较好的模拟能力。同时使用这四个方案，使得产生的新样本中的像素变化的全面性较好。综上，图像处理子模块对原始乳腺图像像素进行修改，以增大乳腺图像样本集的多样性。
作为技术方案的进一步改进，实际使用中，图像获取单元获取乳腺图像后，像素分类单元利用训练好的神经网络对乳腺图像进行像素分类，获得对应每一个像素的像素类标。检测与分类单元对像素分类后的乳腺图像进行乳腺肿块检测和分类，具体地，检测与分类单元包括：
区域划分模块，用于将像素分类后的乳腺图像划分成多个部分重叠的子区域，子区域的大小不同，本实施例中，子区域为矩形区域，矩形区域的尺寸不同（如7像素*7像素, 9像素*9像素, 11像素*11像素, 13像素*13像素）；因为实际的乳腺肿块是大小不一的，因此，本发明的检测与分类***对大小不同的子区域进行判断，以使检测出的乳腺肿块尺寸更符合实际。假如只将输入的乳腺图像划分为非重叠的子区域，子区域的尺寸相同，则当乳腺肿块横跨两个相邻的子区域时，很可能出现漏检。
肿块检测模块，用于判断子区域的良性肿块像素占子区域的总像素的比例或恶性肿块像素占子区域的总像素的比例是否大于预设比例，当判断结果为是时，判断子区域为乳腺肿块；本实施例中，预设比例为0.3。
肿块分类模块，用于判断良性肿块像素占子区域的总像素的比例是否大于恶性肿块像素占子区域的总像素的比例，若判断结果为是，则子区域为良性乳腺肿块，反之，子区域为恶性乳腺肿块。
肿块检测模块和肿块分类模块对乳腺图像的每一个子区域进行肿块检测和分类，基于乳腺图像像素直接进行肿块检测和分类，高效便捷，且准确性高。
本发明还提供一种计算机可读存储介质，其上存储有计算机程序，计算机程序被处理器执行时实现以下步骤：
获取乳腺图像；
对乳腺图像的每一个像素进行分类，像素的类别包括普通像素、良性肿块像素和恶性肿块像素；
对像素分类后的乳腺图像进行乳腺肿块检测和分类。
一种计算机可读存储介质存储的计算机程序的工作过程请参照上述乳腺肿块检测与分类***的描述，不再赘述。
为了高效与便捷地实现基于乳腺钼靶X线摄影图像进行乳腺肿块检测与判识，不同于传统方法，本发明直接对乳腺图像的像素进行判识，将其判定为普通像素、良性肿块像素、恶性肿块像素三个类别；神经网络输出乳腺图像的像素的类别；根据像素类标获得乳腺肿块的定位及其为良性或恶性肿块的判别结果。
以上是对本发明的较佳实施进行了具体说明，但本发明创造并不限于所述实施例，熟悉本领域的技术人员在不违背本发明精神的前提下还可做出种种的等同变形或替换，这些等同的变形或替换均包含在本申请权利要求所限定的范围内。

Claims (10)

  1. 一种乳腺肿块检测与分类***，其特征在于，包括
    图像获取单元，用于获取乳腺图像；
    像素分类单元，用于对所述乳腺图像的每一个像素进行分类，所述像素的类别包括普通像素、良性肿块像素和恶性肿块像素；
    检测与分类单元，用于对像素分类后的乳腺图像进行乳腺肿块检测和分类。
  2. 根据权利要求1所述的乳腺肿块检测与分类***，其特征在于，所述乳腺肿块检测与分类***还包括神经网络单元，所述神经网络单元包括：
    网络构建模块，用于构建深度神经网络，所述深度神经网络用于对所述乳腺图像的每一个像素进行分类，所述像素的类别包括普通像素、良性肿块像素和恶性肿块像素；所述深度神经网络分为六层，按照顺序依次为第一卷积层、第二卷积层、第三卷积层、第一全连接层、第二全连接层和网络输出层，所述深度神经网络的每一层的神经元个数与乳腺图像的像素个数相同；
    网络训练模块，用于利用乳腺图像样本集训练所述深度神经网络。
  3. 根据权利要求2所述的乳腺肿块检测与分类***，其特征在于，所述第一卷积层、第二卷积层、第三卷积层和第一全连接层采用改进的LU激活函数，所述改进的LU激活函数为LU激活函数的计算值小于或等于预设值时，将所述LU激活函数的计算值设置为0。
  4. 根据权利要求2所述的乳腺肿块检测与分类***，其特征在于，所述第二全连接层采用sigmoid激活函数。
  5. 根据权利要求2至4任一项所述的乳腺肿块检测与分类***，其特征在于，所述乳腺图像经过所述第一卷积层、第二卷积层、第三卷积层、第一全连接层和第二全连接层计算得到对应每个像素的标量；
    所述网络输出层用于计算所述标量与像素数字类标的差值的绝对值，所述像素数字类标分别包括普通像素类标、良性肿块像素类标和恶性肿块像素类标，并将绝对值最小的差值所对应的像素数字类标作为像素的类标进行输出。
  6. 根据权利要求5所述的乳腺肿块检测与分类***，其特征在于，所述普通像素类标为-1，所述良性肿块像素类标为0，所述恶性肿块像素类标为1。
  7. 根据权利要求2至4任一项所述的乳腺肿块检测与分类***，其特征在于，所述网络训练模块包括
    图像处理子模块，用于对所述乳腺图像样本集进行处理，所述乳腺图像样本集包括多个乳腺图像样本，处理所述乳腺图像样本集的方法包括：
    对所述乳腺图像样本进行像素灰度值进行变换，包括对所述乳腺图像样本的像素进行同比例增大灰度值或对所述乳腺图像样本的像素进行同比例减小灰度值；
    和/或，
    对全部或者部分的所述乳腺图像样本的像素添加高斯噪声；
    和/或，
    对全部或者部分的所述乳腺图像样本的像素添加椒盐噪声；
    训练子模块，用于利用所述图像处理子模块处理后的乳腺图像样本集训练所述深度神经网络。
  8. 根据权利要求1至4任一项所述的乳腺肿块检测与分类***，其特征在于，所述检测与分类单元包括
    区域划分模块，用于将像素分类后的乳腺图像划分成多个部分重叠的子区域，所述子区域的大小不同；
    肿块检测模块，用于判断所述子区域的良性肿块像素占所述子区域的总像素的比例或恶性肿块像素占所述子区域的总像素的比例是否大于预设比例，当判断结果为是时，判断所述子区域为乳腺肿块；
    肿块分类模块，用于判断所述良性肿块像素占所述子区域的总像素的比例是否大于恶性肿块像素占所述子区域的总像素的比例，若判断结果为是，则所述子区域为良性乳腺肿块，反之，所述子区域为恶性乳腺肿块。
  9. 根据权利要求8所述的乳腺肿块检测与分类***，其特征在于，所述预设比例为0.3。
  10. 一种计算机可读存储介质，其特征在于，其上存储有计算机程序，所述计算机程序被处理器执行时实现以下步骤：
    获取乳腺图像；
    对所述乳腺图像的每一个像素进行分类，所述像素的类别包括普通像素、良性肿块像素和恶性肿块像素；
    对像素分类后的乳腺图像进行乳腺肿块检测和分类。
PCT/CN2018/124655 2018-07-23 2018-12-28 一种乳腺肿块检测与分类***、计算机可读存储介质 WO2020019671A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810810971.0 2018-07-23
CN201810810971.0A CN109191424B (zh) 2018-07-23 2018-07-23 一种乳腺肿块检测与分类***、计算机可读存储介质

Publications (1)

Publication Number Publication Date
WO2020019671A1 true WO2020019671A1 (zh) 2020-01-30

Family

ID=64937114

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/124655 WO2020019671A1 (zh) 2018-07-23 2018-12-28 一种乳腺肿块检测与分类***、计算机可读存储介质

Country Status (2)

Country Link
CN (1) CN109191424B (zh)
WO (1) WO2020019671A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111554383A (zh) * 2020-04-24 2020-08-18 浙江杜比医疗科技有限公司 一种用于乳腺肿瘤检测的神经网络及其检测***
CN112241954A (zh) * 2020-10-22 2021-01-19 上海海事大学 基于肿块差异化分类的全视野自适应分割网络配置方法
CN112699948A (zh) * 2020-12-31 2021-04-23 无锡祥生医疗科技股份有限公司 超声乳腺病灶的分类方法、装置及存储介质
CN115619641A (zh) * 2022-10-24 2023-01-17 中山大学附属第五医院 基于ffdm的乳腺影像处理方法、***、终端和介质

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109785320A (zh) * 2019-01-28 2019-05-21 太原理工大学 一种采用改进型AlexNet模型对乳腺钼靶X线图像进行分类和识别的方法
CN109872307B (zh) * 2019-01-30 2022-01-07 腾讯科技（深圳）有限公司 一种检测生物组织图像中肿块的方法、相应设备及介质
CN111767760A (zh) * 2019-04-01 2020-10-13 北京市商汤科技开发有限公司 活体检测方法和装置、电子设备及存储介质
CN110827296B (zh) * 2019-11-01 2023-06-23 南京信息工程大学 多目标集成式深度神经网络的乳腺x光影像分析方法
CN113627416B (zh) * 2021-10-12 2022-01-25 上海蜜度信息技术有限公司 图片分类和对象检测的同步处理方法、***、存储介质及终端

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060116579A1 (en) * 2004-11-29 2006-06-01 Pai-Chi Li Ultrasound imaging apparatus and method thereof
CN104851101A (zh) * 2015-05-25 2015-08-19 电子科技大学 一种基于深度学习的脑肿瘤自动分割方法
CN107464250A (zh) * 2017-07-03 2017-12-12 深圳市第二人民医院 基于三维mri图像的乳腺肿瘤自动分割方法
CN107749061A (zh) * 2017-09-11 2018-03-02 天津大学 基于改进的全卷积神经网络脑肿瘤图像分割方法及装置

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP4800127B2 (ja) * 2006-06-29 2011-10-26 富士フイルム株式会社 医用画像分割装置、及び、医用画像分割プログラム
EP2727514B1 (en) * 2012-03-21 2017-05-10 Olympus Corporation Image processing device
CN104408482B (zh) * 2014-12-08 2019-02-12 电子科技大学 一种高分辨率sar图像目标检测方法
CN104966100A (zh) * 2015-06-17 2015-10-07 北京交通大学 基于纹理基元的图像肿块良恶性的分类方法
CN107507195B (zh) * 2017-08-14 2019-11-15 四川大学 基于超图模型的pet-ct多模态鼻咽癌图像分割方法
CN107590806B (zh) * 2017-09-19 2021-06-01 陈烨 一种基于大脑医学成像的检测方法和***
CN107845098A (zh) * 2017-11-14 2018-03-27 南京理工大学 基于随机森林和模糊聚类的肝癌图像全自动分割方法
CN107886514B (zh) * 2017-11-22 2021-04-23 浙江中医药大学 基于深度残差网络的乳腺钼靶图像肿块语义分割方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060116579A1 (en) * 2004-11-29 2006-06-01 Pai-Chi Li Ultrasound imaging apparatus and method thereof
CN104851101A (zh) * 2015-05-25 2015-08-19 电子科技大学 一种基于深度学习的脑肿瘤自动分割方法
CN107464250A (zh) * 2017-07-03 2017-12-12 深圳市第二人民医院 基于三维mri图像的乳腺肿瘤自动分割方法
CN107749061A (zh) * 2017-09-11 2018-03-02 天津大学 基于改进的全卷积神经网络脑肿瘤图像分割方法及装置

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111554383A (zh) * 2020-04-24 2020-08-18 浙江杜比医疗科技有限公司 一种用于乳腺肿瘤检测的神经网络及其检测***
CN111554383B (zh) * 2020-04-24 2023-09-05 浙江杜比医疗科技有限公司 一种用于乳腺肿瘤检测的神经网络及其检测***
CN112241954A (zh) * 2020-10-22 2021-01-19 上海海事大学 基于肿块差异化分类的全视野自适应分割网络配置方法
CN112241954B (zh) * 2020-10-22 2024-03-15 上海海事大学 基于肿块差异化分类的全视野自适应分割网络配置方法
CN112699948A (zh) * 2020-12-31 2021-04-23 无锡祥生医疗科技股份有限公司 超声乳腺病灶的分类方法、装置及存储介质
CN115619641A (zh) * 2022-10-24 2023-01-17 中山大学附属第五医院 基于ffdm的乳腺影像处理方法、***、终端和介质

Also Published As

Publication number Publication date
CN109191424A (zh) 2019-01-11
CN109191424B (zh) 2022-04-22

Similar Documents

Publication Publication Date Title
WO2020019671A1 (zh) 一种乳腺肿块检测与分类***、计算机可读存储介质
JP2882468B2 (ja) 画像の動的ヒストグラムワーピング装置及び方法
WO2020000643A1 (zh) Ct图像肺结节检测方法、设备及可读存储介质
WO2013174231A1 (zh) 增强现实交互的实现方法和***
JP4597391B2 (ja) 顔領域検出装置およびその方法並びにコンピュータ読み取り可能な記録媒体
TWI419082B (zh) 偵測視訊序列中的移動物體影像的方法以及影像處理系統
CN111612736A (zh) 一种电力设备故障检测方法、计算机和计算机程序
WO2021012508A1 (zh) Ai影像识别方法、装置、设备及存储介质
CN111291605A (zh) 人流分析***及人流分析方法
WO2022111682A1 (zh) 移动行人检测方法、电子设备及机器人
Zhou et al. A multi-image stitching method and quality evaluation
JP2007293559A (ja) 非定常映像検出装置，非定常映像検出方法及びその方法を実装したプログラム
CN113724143A (zh) 用于图像修复的方法和装置
WO2020101300A1 (ko) 영상 처리 장치 및 그 동작방법
Chen et al. Image segmentation in thermal images
Lee et al. Vehicle counting based on a stereo vision depth maps for parking management
JPS6376578A (ja) 自動２値化方式
JP2001229371A (ja) 画像アライメント方法
WO2020111382A1 (ko) 단일 영상을 기반으로 한 인버스 톤 매핑의 최적화 장치 및 방법, 상기 방법을 수행하기 위한 기록 매체
WO2023055013A1 (ko) 뉴럴 네트워크(Neural Network)에 기초한 영상 처리 방법 및 영상 처리 장치
Han et al. An efficient estimation method for intensity factor of illumination changes
WO2021150016A1 (en) Methods and systems for performing tasks on media using attribute specific joint learning
WO2024076169A1 (ko) 공간정보를 이용한 객체 인식 모델의 학습 방법 및 이를 수행하기 위한 컴퓨팅 장치
WO2024076020A1 (ko) 공간 맵을 생성하는 방법 및 서버
WO2014003507A1 (ko) 영상 부호화기의 전처리 장치와 3차원 얼굴 생성 장치 및 방법

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: 18927743

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: 18927743

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 18927743

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 13/08/2021)

122 Ep: pct application non-entry in european phase

Ref document number: 18927743

Country of ref document: EP

Kind code of ref document: A1