WO2022063199A1 - 一种肺结节自动检测方法、装置及计算机系统 - Google Patents

一种肺结节自动检测方法、装置及计算机系统 Download PDF

Info

Publication number
WO2022063199A1
WO2022063199A1 PCT/CN2021/120100 CN2021120100W WO2022063199A1 WO 2022063199 A1 WO2022063199 A1 WO 2022063199A1 CN 2021120100 W CN2021120100 W CN 2021120100W WO 2022063199 A1 WO2022063199 A1 WO 2022063199A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
automatic detection
pulmonary nodules
detection
lung
Prior art date
Application number
PCT/CN2021/120100
Other languages
English (en)
French (fr)
Inventor
黄钢
聂生东
陈阳
Original Assignee
上海健康医学院
上海理工大学
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 上海健康医学院, 上海理工大学 filed Critical 上海健康医学院
Publication of WO2022063199A1 publication Critical patent/WO2022063199A1/zh

Links

Images

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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • 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
    • G06T7/10—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
    • G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
    • 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
    • G06T2207/10072—Tomographic images
    • G06T2207/10081—Computed x-ray tomography [CT]
    • 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
    • G06T2207/20021—Dividing image into blocks, subimages or windows
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20112—Image segmentation details
    • G06T2207/20132—Image cropping
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30061—Lung
    • G06T2207/30064—Lung nodule

Definitions

  • the invention relates to the field of computer-aided detection, in particular to a method, device and computer system for automatic detection of pulmonary nodules.
  • Lung cancer is one of the tumor diseases with the highest mortality rate. Every year, more than 1.3 million people in the world die from lung cancer. In the early stage of lung cancer, its performance is not obvious, and more than 70% of lung cancer patients are basically in the advanced stage of lung cancer. According to relevant medical statistics, if lung cancer patients can receive proper intervention treatment in the early stage of cancer, their 5-year survival rate can reach more than 90%, but the survival rate of lung cancer patients in stage 2 to 3 drops to 40% to 5%. Therefore, "early detection, early diagnosis and early treatment" is the key to improve the survival rate of lung cancer patients. Lung cancer is generally manifested as pulmonary nodules in the early stage, and the detection of pulmonary nodules is the first step in the early diagnosis of lung cancer.
  • CAD computer-aided detection system
  • the purpose of the present invention is to provide an automatic detection method, device and computer system for pulmonary nodules with high detection sensitivity and high precision in order to overcome the above-mentioned defects of the prior art.
  • a method for automatic detection of pulmonary nodules comprises the following steps:
  • the CT image sequence is segmented by a threshold method to obtain an image containing only the lung parenchyma area;
  • a 3D CNN model was used to automatically detect and identify the region of interest to obtain the detection results of pulmonary nodules.
  • the filtering enhancement processing includes image filtering and window width and window level adjustment.
  • segmentation process specifically includes:
  • the CT image was roughly segmented by the threshold method, and the holes were filled, and the initial lung parenchyma area was selected by the three-dimensional connected area labeling method;
  • the segmented initial lung parenchyma area is repaired, the edge of the lung is adjusted using morphological closing operation, and the edge is expanded outward by several pixels to obtain the final lung parenchyma area.
  • the patching is performed on the initial lung parenchyma region using a convex hull algorithm.
  • the U-Net network model of the multi-scale feature fusion introduces random inactivation, including a multi-scale feature fusion unit, a number of convolution max pooling units and a number of upsampling units, the convolution max pooling unit and the upper sampling unit.
  • the feature maps of the same scale and size obtained by the sampling unit are spliced together.
  • both the multi-scale feature fusion unit and the convolution maximum pooling unit include a combination of a convolution layer and a batch normalization layer.
  • the 3D CNN model includes several sub-models composed of a Dense Block unit and a pooling layer, and the output ends of each sub-model are connected, and after merging the prediction results of each sub-model, the final lung nodule detection result is output.
  • each sample includes CT images and corresponding pulmonary nodules labeling results, and the pulmonary nodules labeling results are: A fusion of at least three manually annotated results.
  • the present invention also provides an automatic detection device for pulmonary nodules, comprising:
  • the CT image acquisition module is used to acquire the CT image to be detected
  • a preprocessing module configured to perform filtering and enhancement processing on the CT image to be detected to obtain a CT image sequence enhanced by the lungs;
  • a lung parenchyma segmentation module used for segmenting the CT image sequence using a threshold method to obtain an image that only includes the lung parenchyma area;
  • a region of interest extraction module which is used for cropping the image obtained by the lung parenchyma segmentation module into several image blocks, and obtains the region of interest through a U-Net network model fused with multi-scale features;
  • the classification module is used to automatically detect and identify the region of interest by using a 3D CNN model to obtain the detection results of pulmonary nodules.
  • the present invention also provides a computer system for automatic detection of pulmonary nodules, comprising:
  • the processor is coupled to the memory for reading program instructions stored in the memory, and in response, executing the steps in the method as described above.
  • the present invention has the following beneficial effects:
  • the lung nodule detection method proposed by the present invention is realized by two models of the U-Net network model and 3D CNN model fused with multi-scale features. First, a segmentation model is constructed to locate the candidate region of the lung nodule, and then a classification model is constructed to identify the lung nodules. Nodules and non-nodules are classified and differentiated, so as to achieve the detection of pulmonary nodules. On the condition of maintaining 8FPs/Scan in the processing data, the sensitivity of the present invention can reach more than 90%, and the detection result is better.
  • the present invention extracts and fuses multi-scale features of pulmonary nodules, thereby improving the sensitivity of pulmonary nodule detection, with high detection sensitivity, and can accurately and efficiently detect pulmonary nodules.
  • the present invention has a preprocessing process for CT images, effectively reducing image noise, improving image quality, and further improving detection accuracy.
  • the U-Net network model of multi-scale feature fusion designed by the present invention can extract and fuse multi-scale characteristic information of pulmonary nodules, and can realize the repair of detailed information, with high feature accuracy, and can segment as many nodules as possible. .
  • Random deactivation is introduced into the U-Net model established by the present invention, and the convolutional layers in the constructed network include the combination of the convolutional layer and the batch normalization (BN) layer, which reduces the overfitting of the model and effectively improves the performance of the model. Accuracy of lung nodule segmentation.
  • the 3D CNN model established by the present invention outputs the prediction result after each pooling layer, and through the fusion of each prediction result, the final detection result is obtained, which further improves the detection accuracy.
  • Figure 1 is a flowchart of deep learning lung nodule detection based on multi-scale information
  • Figure 2 is a schematic diagram of a designed multi-scale feature fusion unit
  • Figure 3 is a schematic diagram of the designed U-Net network structure
  • Figure 4 is a schematic diagram of the designed 3D CNN network structure
  • Figure 5 is a schematic diagram of the structure of the Dense Block in the 3D CNN network.
  • this embodiment provides an automatic detection method for pulmonary nodules, which includes:
  • Step 1 Obtain CT image data.
  • Step 2 Sample preprocessing, including image filtering and window width and window level adjustment.
  • the lung CT images are filtered by means of median filtering, mean filtering, etc., to reduce image noise and improve image quality, and then adjust the window width and window level of the lung CT images for enhancement. Contrast of the lung parenchyma area to obtain a sequence of lung enhanced CT images.
  • the specific operation of adjusting the window width and the window level may be to set the pixels with a HU value greater than 400 to 400, and set the pixels with a HU value less than -1000 to -1000.
  • Step 3 Lung parenchyma segmentation.
  • the lung parenchyma segmentation is mainly performed by the "threshold method", and the segmentation results are repaired to accurately obtain the lung parenchyma area. Specifically, first, the gray value of the CT image is converted into the HU value, normalized, and then the lung parenchyma is roughly segmented using the set threshold T, and then the hole is filled, and then the "3D connected area marker" is used. method, select the lung parenchyma area, use the "convex hull” algorithm to repair the segmented lung parenchyma area, and finally use the morphological closing operation to adjust the edge of the lung, and expand the obtained lung parenchyma area outward by several pixel points to get the final lung parenchyma area. In this embodiment, the number of pixel expansions can be set to 10.
  • Step 4 Region of interest extraction.
  • a U-Net network model with multi-scale feature fusion is constructed.
  • the multi-scale feature fusion U-Net network model can extract multi-scale lung nodule information, and fuse the extracted multi-scale lung nodule information, so as to improve lung nodule information. Section segmentation accuracy.
  • the U-Net network model of multi-scale feature fusion in this embodiment introduces random deactivation, including a multi-scale feature fusion unit, several convolutional max pooling units and several upsampling units, and convolution max pooling units.
  • the feature maps of the same scale and size obtained by the unit and the upsampling unit are concatenated.
  • the multi-scale feature fusion unit includes four channels, and the convolution kernel sizes are 1*1, 3*3, 5*5, 7*7, and then extract the multi-scale
  • the feature information is spliced and input into the convolutional max pooling unit.
  • Both the multi-scale feature fusion unit and the convolution unit include a combination of a convolution layer and a batch normalization (BN) layer, thereby improving the accuracy of lung nodule segmentation.
  • BN batch normalization
  • the specific process of region of interest extraction includes:
  • Step 4.1 First, input the input m*m*n image blocks into the U-Net network model, and then go through a multi-scale feature fusion unit to extract and fuse the multi-scale feature information of lung nodules, and then go through four Convolution and max pooling units are used to finally extract feature maps of different scales of nodules.
  • Step 4.2 The extracted nodule feature map is subjected to an upsampling operation opposite to the pooling operation, and the feature map is gradually restored to the scale of the original image.
  • the same scale size in the feature extraction process is simultaneously
  • the feature map of , and the up-sampled feature map are spliced together for the repair of detailed information.
  • random inactivation is introduced into the established U-Net model, and the output of the model is m*m*
  • the segmentation result of n that is, the segmentation result of lung nodules.
  • Step 5 Classifier classification.
  • the obtained region of interest is saved as an image block of size x*x*w, and a 3D CNN model is used to automatically detect and identify the region of interest to obtain the detection results of lung nodules.
  • the 3D CNN model of this embodiment includes several sub-models composed of Dense Block units and pooling layers. After each pooling layer, a prediction result is output. At the end of the model, each prediction result is fused , to obtain the final pulmonary nodule detection results.
  • the 3D CNN model extracts features through multiple Dense Block units and pooling layers, and finally fuses and classifies the obtained feature information to identify nodules and non-nodules.
  • the structure of the Dense Block unit is shown in Figure 5.
  • each sample includes CT images and corresponding pulmonary nodule annotation results, and the pulmonary nodule annotation results are at least three manual Fusion of annotation results.
  • the construction of the sample data set used for training is as follows: remove cases with a slice thickness greater than 2.5 mm from the LIDC data set of the Lung Image Database Alliance, select 888 sets of CT images for training, and provide XML annotations through them. file, from which the coordinate information of pulmonary nodules is extracted, and the annotation results of four radiologists are fused. The doctor's annotation is not used as the annotation result, and the obtained CT image data and the annotation result are combined to form a sample data set.
  • model training is implemented after the samples in the sample data set are processed as in steps 2 and 3.
  • the above method studies the detection of pulmonary nodules by designing experiments, and constructs a pulmonary nodule detection model based on deep learning.
  • a segmentation model is constructed to locate the candidate regions of pulmonary nodules, and then a classification model is constructed to identify nodules and non-nodules. To classify and identify, so as to achieve the detection of pulmonary nodules.
  • the sensitivity of the above method can reach more than 90%, and the detection result is better.
  • This embodiment provides an automatic detection device for pulmonary nodules, including: a CT image acquisition module for acquiring a CT image to be detected; a preprocessing module for filtering and enhancing the CT image to be detected to obtain lung enhancement
  • the lung parenchyma segmentation module is used to segment the CT image sequence using the threshold method to obtain images that only contain the lung parenchyma area; the region of interest extraction module is used to obtain the lung parenchyma segmentation module.
  • the image is cropped into several image blocks, and the region of interest is obtained through a U-Net network model fused with multi-scale features; the classification module is used to automatically detect and identify the region of interest using a 3D CNN model to obtain lung nodule detection. result.
  • the rest are the same as in Example 1.
  • This embodiment provides a computer system for automatic detection of pulmonary nodules, including a processor and a memory storing executable instructions of the processor; wherein the processor is coupled to the memory and is used to read program instructions stored in the memory , and in response, perform the steps in the method described in Example 1.

Abstract

一种肺结节自动检测方法、装置及计算机系统，所述方法包括以下步骤：获取待检测CT影像；对所述待检测CT影像进行滤波增强处理，获得肺部增强的CT影像序列；采用阈值法对所述CT影像序列进行分割处理，获得仅包含肺实质区域的图像；将所述肺实质分割模块获得的图像裁剪为若干图像块，通过一多尺度特征融合的U-Net网络模型，获得感兴趣区域；采用一3D CNN模型对感兴趣区域进行自动检测识别，获得肺结节检测结果。该方法具有检测灵敏度和精度高等优点。

Description

一种肺结节自动检测方法、装置及计算机系统 技术领域
本发明涉及计算机辅助检测领域，尤其是涉及一种肺结节自动检测方法、装置及计算机系统。
背景技术
肺癌是死亡率最高的肿瘤疾病之一。每年世界上有超过130万人因为患肺癌而死亡。在肺癌的早期阶段，其表现不明显，70％以上的肺癌确诊患者基本上都处于肺癌晚期。根据有关医学统计数据，如果肺癌患者在癌症早期能获得正当的干预治疗，其5年存活率可达90％以上，然而处于2～3期的肺癌患者存活率下降到40％～5％。因此“早发现、早诊断、早治疗”是提高肺癌患者生存率的关键。肺癌早期一般表现为肺结节，对肺结节的检测是肺癌早期诊断的首要步骤。当前，肺癌的早期发现最常用的是通过CT断层扫描技术进行成像检查，然后由放射科医生进行筛查，但CT产生的大量影像加重了放射科医生的阅片负担，不但费时费力，而且随着阅片时间一长，会致使诊断的效率和准确率下降。在此背景下，提出了基于CT影像的肺癌计算机辅助检测系统(CAD)辅助放射科医生，作为“第二意见”辅助临床诊断。
从目前的国内外研究现状来看，有许多研究者致力于对肺结节的检测，但现有方法还存在敏感性不高，假阳过多的情况。
发明内容
本发明的目的就是为了克服上述现有技术存在的缺陷而提供一种检测灵敏度和精度高的肺结节自动检测方法、装置及计算机系统。
本发明的目的可以通过以下技术方案来实现：
一种肺结节自动检测方法，该方法包括以下步骤：
获取待检测CT影像；
对所述待检测CT影像进行滤波增强处理，获得肺部增强的CT影像序列；
采用阈值法对所述CT影像序列进行分割处理，获得仅包含肺实质区域的 图像；
将所述肺实质分割模块获得的图像裁剪为若干图像块，通过一多尺度特征融合的U-Net网络模型，获得感兴趣区域；
采用一3D CNN模型对感兴趣区域进行自动检测识别，获得肺结节检测结果。
进一步地，所述滤波增强处理包括图像滤波和窗宽窗位调节。
进一步地，所述分割处理具体包括：
使用阈值法对CT影像进行粗分割，并填充孔洞，采用三维联通区域标记法，选择初始肺实质区域；
对分割出的所述初始肺实质区域进行修补，使用形态学闭运算调整肺部边缘，并边缘处向外扩张若干像素点，获得最终的肺实质区域。
进一步地，使用凸包算法对初始肺实质区域进行所述修补。
进一步地，所述多尺度特征融合的U-Net网络模型引入随机失活，包括多尺度特征融合单元、若干卷积最大池化单元和若干上采样单元，所述卷积最大池化单元和上采样单元获得的相同尺度大小的特征图相拼接。
进一步地，所述多尺度特征融合单元和卷积最大池化单元均包括卷积层与batch normalization层的组合。
进一步地，所述3D CNN模型包括若干由Dense Block单元和池化层组成的子模型，各子模型的输出端相连接，融合各子模型的预测结果后，输出最终的肺结节检测结果。
进一步地，所述多尺度特征融合的U-Net网络模型和3D CNN模型进行训练时采用的样本数据集中，各样本包括CT影像及对应的肺结节标注结果，所述肺结节标注结果为至少三个手动标注结果的融合。
本发明还提供一种肺结节自动检测装置，包括：
CT影像获取模块，用于获取待检测CT影像；
预处理模块，用于对所述待检测CT影像进行滤波增强处理，获得肺部增强的CT影像序列；
肺实质分割模块，用于采用阈值法对所述CT影像序列进行分割处理，获得仅包含肺实质区域的图像；
感兴趣区域提取模块，用于将所述肺实质分割模块获得的图像裁剪为若干图像块，通过一多尺度特征融合的U-Net网络模型，获得感兴趣区域；
分类模块，用于采用一3D CNN模型对感兴趣区域进行自动检测识别，获得肺结节检测结果。
本发明还提供一种肺结节自动检测计算机系统，包括：
处理器；
存储处理器可执行指令的存储器；
其中，所述处理器耦合于所述存储器，用于读取所述存储器存储的程序指令，并作为响应，执行如上所述方法中的步骤。
与现有技术相比，本发明具有如下有益效果：
1、本发明提出的肺结节检测方法通过多尺度特征融合的U-Net网络模型和3D CNN模型两个模型实现，首先构建分割模型对肺结节的候选区域进行定位，然后构建分类模型对结节和非结节进行分类鉴别，从而达到对肺结节的检测。本发明在所述的处理数据上在保持8FPs/Scan的情况下，敏感性可达90％以上，具有较好的检测结果。
2、本发明对肺结节的多尺度特征进行提取、融合，从而提高肺结节检测的敏感性，检测灵敏度较高，可对肺结节进行准确、高效检测。
3、本发明具有对CT影像的预处理过程，有效降低图像的噪声，提升图像的质量，进而提高检测精度。
4、本发明设计的多尺度特征融合的U-Net网络模型可提取、融合肺结节的多尺度特性信息，可实现细节信息的修补，特征准确性高，可以尽可能多的分割出结节。
5、本发明建立的U-Net模型中引入了随机失活，且构建的网络中的卷积层均包括卷积层与batch normalization(BN)层的组合，降低模型过拟合，有效提高了肺结节分割的准确率。
6、本发明建立的3D CNN模型在每个池化层后均输出预测结果，通过各个预测结果的融合，得到最终的检测结果，进一步提高了检测精度。
附图说明
图1为基于多尺度信息的深度学习肺结节检测流程图；
图2为设计的多尺度特征融合单元示意图；
图3为设计的U-Net网络结构示意图；
图4为设计的3D CNN网络结构示意图；
图5为3D CNN网络中Dense Block结构示意图。
具体实施方式
下面结合附图和具体实施例对本发明进行详细说明。本实施例以本发明技术方案为前提进行实施，给出了详细的实施方式和具体的操作过程，但本发明的保护范围不限于下述的实施例。
实施例1
如图1所示，本实施例提供一种肺结节自动检测方法，该方法包括：
步骤1、CT影像数据获取。
步骤2、样本预处理，包括图像滤波和窗宽窗位调节。
本实施例中，利用中值滤波、均值滤波等方式对肺部CT影像进行滤波，降低图像的噪声，提升图像的质量，然后对肺部CT影像进行调节窗宽、窗位操作，用于增强肺实质区域的对比度，得到肺部增强的CT影像序列。其中，调节窗宽、窗位的具体操作可为将HU值大于400的像素设置为400，将HU值小于-1000的像素设置为-1000。
步骤3、肺实质分割。
肺实质分割主要通过“阈值法”进行肺实质的分割，并对分割结果进行修补，以准确获得肺实质区域。具体地，首先，将CT影像的灰度值转化为HU值，进行归一化，然后使用设定的阈值T对肺实质进行粗分割，再经过孔洞的填充，然后使用“三维联通区域标记”的方法，选择出肺实质区域，使用“凸包”算法对分割后的肺实质区域进行修补，最后再使用形态学闭运算进行调整肺部的边缘，将得到的肺实质区域向外扩张若干个像素点，得到最终的肺实质区域。本实施例中，像素点扩张个数可设置为10。
步骤4、感兴趣区域提取。
构建多尺度特征融合的U-Net网络模型，该多尺度特征融合的U-Net网络 模型可提取多尺度的肺结节信息，并将提取的多尺度肺结节信息进行融合，从而提高肺结节分割的准确率。
如图3所示，本实施例的多尺度特征融合的U-Net网络模型引入随机失活，包括多尺度特征融合单元、若干卷积最大池化单元和若干上采样单元，卷积最大池化单元和上采样单元获得的相同尺度大小的特征图相拼接。如图2所示，多尺度特征融合单元包括四个通道，其卷积核大小分别为1*1，3*3，5*5，7*7，然后再将四个通道提取到的多尺度特征信息进行拼接，输入到卷积最大池化单元中。多尺度特征融合单元与卷积单元均包括卷积层与batch normalization(BN)层的组合，从而提高肺结节分割的准确率。
感兴趣区域提取的具体过程包括：
步骤4.1、首先将输入的m*m*n的图像块输入到U-Net网络模型当中，然后经过一个多尺度特征融合单元用于提取、融合肺结节的多尺度特性信息，再经过四个卷积和最大池化单元，最终提取出结节的不同尺度的特征图。
步骤4.2、将提取到的结节特征图再经过与池化操作相反的上采样操作，将特征图逐步恢复到原图的尺度，在这一过程中，同时将特征提取过程中的相同尺度大小的特征图与上采样的到的特征图进行拼接，用于细节信息的修补，同时为了降低模型过拟合，建立的U-Net模型中引入了随机失活，模型的输出为m*m*n的分割结果，即肺结节分割结果。
步骤5、分类器分类。
将所得到的感兴趣区域保存成x*x*w大小的图像块，采用一3D CNN模型对感兴趣区域进行自动检测识别，获得肺结节检测结果。
如图4所示，本实施例的3D CNN模型包括若干由Dense Block单元和池化层组成的子模型，在每个池化层后均输出预测结果，模型的最后将各个预测的结果进行融合，得到最终的肺结节检测结果。3D CNN模型经过多个Dense Block单元和池化层提取特征，最后再将得到的特征信息进行融合、分类，从而识别出结节和非结节。Dense Block单元的结构如图5所示。
上述多尺度特征融合的U-Net网络模型和3D CNN模型进行训练时采用的样本数据集中，各样本包括CT影像及对应的肺结节标注结果，所述肺结节标注结果为至少三个手动标注结果的融合。
本实施例中，训练所用的样本数据集的构建具体为：从肺部图像数据库联盟LIDC数据集中去除层厚大于2.5mm的病例，筛选出用于训练的888套CT影像，通过其提供XML标注文件，从中提取出肺结节的坐标信息，并将四个放射科医生的标注结果进行融合，其中将至少被三个放射科医生标注过的结节纳为标注结果，少于三个放射科医生标注的则不作为标注结果，将获得的CT影像数据和标注结果一起组成样本数据集。进行训练时，对样本数据集中的样本进行如步骤2和3的处理后实现模型训练。
上述方法通过设计实验对肺结节的检测进行研究，构建基于深度学习的肺结节检测模型，首先构建分割模型对肺结节的候选区域进行定位，然后构建分类模型对结节和非结节进行分类鉴别，从而达到对肺结节的检测。上述方法在所述的处理数据上在保持8FPs/Scan的情况下，敏感性可达90％以上，具有较好的检测结果。
实施例2
本实施例提供一种肺结节自动检测装置，包括：CT影像获取模块，用于获取待检测CT影像；预处理模块，用于对所述待检测CT影像进行滤波增强处理，获得肺部增强的CT影像序列；肺实质分割模块，用于采用阈值法对所述CT影像序列进行分割处理，获得仅包含肺实质区域的图像；感兴趣区域提取模块，用于将所述肺实质分割模块获得的图像裁剪为若干图像块，通过一多尺度特征融合的U-Net网络模型，获得感兴趣区域；分类模块，用于采用一3D CNN模型对感兴趣区域进行自动检测识别，获得肺结节检测结果。其余同实施例1。
实施例3
本实施例提供一种肺结节自动检测计算机系统，包括处理器和存储处理器可执行指令的存储器；其中，所述处理器耦合于所述存储器，用于读取所述存储器存储的程序指令，并作为响应，执行如实施例1所述方法中的步骤。
以上详细描述了本发明的较佳具体实施例。应当理解，本领域的普通技术人员无需创造性劳动就可以根据本发明的构思作出诸多修改和变化。因此，凡本技术领域中技术人员依本发明的构思在现有技术的基础上通过逻辑分析、推理或者有限的实验可以得到的技术方案，皆应在由权利要求书所确定的保护范 围内。

Claims (10)

  1. 一种肺结节自动检测方法，其特征在于，该方法包括以下步骤：
    获取待检测CT影像；
    对所述待检测CT影像进行滤波增强处理，获得肺部增强的CT影像序列；
    采用阈值法对所述CT影像序列进行分割处理，获得仅包含肺实质区域的图像；
    将所述肺实质分割模块获得的图像裁剪为若干图像块，通过一多尺度特征融合的U-Net网络模型，获得感兴趣区域；
    采用一3D CNN模型对感兴趣区域进行自动检测识别，获得肺结节检测结果。
  2. 根据权利要求1所述的肺结节自动检测方法，其特征在于，所述滤波增强处理包括图像滤波和窗宽窗位调节。
  3. 根据权利要求1所述的肺结节自动检测方法，其特征在于，所述分割处理具体包括：
    使用阈值法对CT影像进行粗分割，并填充孔洞，采用三维联通区域标记法，选择初始肺实质区域；
    对分割出的所述初始肺实质区域进行修补，使用形态学闭运算调整肺部边缘，并边缘处向外扩张若干像素点，获得最终的肺实质区域。
  4. 根据权利要求3所述的肺结节自动检测方法，其特征在于，使用凸包算法对初始肺实质区域进行所述修补。
  5. 根据权利要求1所述的肺结节自动检测方法，其特征在于，所述多尺度特征融合的U-Net网络模型引入随机失活，包括多尺度特征融合单元、若干卷积最大池化单元和若干上采样单元，所述卷积最大池化单元和上采样单元获得的相同尺度大小的特征图相拼接。
  6. 根据权利要求5所述的肺结节自动检测方法，其特征在于，所述多尺度特征融合单元和卷积最大池化单元均包括卷积层与batch normalization层的组合。
  7. 根据权利要求1所述的肺结节自动检测方法，其特征在于，所述3D CNN 模型包括若干由Dense Block单元和池化层组成的子模型，各子模型的输出端相连接，融合各子模型的预测结果后，输出最终的肺结节检测结果。
  8. 根据权利要求1所述的肺结节自动检测方法，其特征在于，所述多尺度特征融合的U-Net网络模型和3D CNN模型进行训练时采用的样本数据集中，各样本包括CT影像及对应的肺结节标注结果，所述肺结节标注结果为至少三个手动标注结果的融合。
  9. 一种肺结节自动检测装置，其特征在于，包括：
    CT影像获取模块，用于获取待检测CT影像；
    预处理模块，用于对所述待检测CT影像进行滤波增强处理，获得肺部增强的CT影像序列；
    肺实质分割模块，用于采用阈值法对所述CT影像序列进行分割处理，获得仅包含肺实质区域的图像；
    感兴趣区域提取模块，用于将所述肺实质分割模块获得的图像裁剪为若干图像块，通过一多尺度特征融合的U-Net网络模型，获得感兴趣区域；
    分类模块，用于采用一3D CNN模型对感兴趣区域进行自动检测识别，获得肺结节检测结果。
  10. 一种肺结节自动检测计算机系统，其特征在于，包括：
    处理器；
    存储处理器可执行指令的存储器；
    其中，所述处理器耦合于所述存储器，用于读取所述存储器存储的程序指令，并作为响应，执行如权利要求1-8中任一项所述方法中的步骤。
PCT/CN2021/120100 2020-09-24 2021-09-24 一种肺结节自动检测方法、装置及计算机系统 WO2022063199A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011018485.9A CN112184657A (zh) 2020-09-24 2020-09-24 一种肺结节自动检测方法、装置及计算机系统
CN202011018485.9 2020-09-24

Publications (1)

Publication Number Publication Date
WO2022063199A1 true WO2022063199A1 (zh) 2022-03-31

Family

ID=73943409

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/120100 WO2022063199A1 (zh) 2020-09-24 2021-09-24 一种肺结节自动检测方法、装置及计算机系统

Country Status (2)

Country Link
CN (1) CN112184657A (zh)
WO (1) WO2022063199A1 (zh)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114757942A (zh) * 2022-05-27 2022-07-15 南通大学 一种基于深度学习的多层螺旋ct对活动性肺结核的检测方法
CN114974518A (zh) * 2022-04-15 2022-08-30 浙江大学 多模态数据融合的肺结节影像识别方法及装置
CN115358976A (zh) * 2022-08-10 2022-11-18 北京医准智能科技有限公司 一种图像识别方法、装置、设备及存储介质
CN115578309A (zh) * 2022-08-04 2023-01-06 云南师范大学 肺癌特征信息获取的方法、系统、电子设备和存储介质
CN115578564A (zh) * 2022-10-25 2023-01-06 北京医准智能科技有限公司 实例分割模型的训练方法、装置、电子设备及存储介质
CN116228685A (zh) * 2023-02-07 2023-06-06 重庆大学 一种基于深度学习的肺结节检测与剔除方法
CN116934757A (zh) * 2023-09-18 2023-10-24 电子科技大学 一种用于肺结节假阳性删减的方法、设备及存储介质
CN117542527A (zh) * 2024-01-09 2024-02-09 百洋智能科技集团股份有限公司 肺结节跟踪与变化趋势预测方法、装置、设备及存储介质
CN117809122A (zh) * 2024-02-29 2024-04-02 北京航空航天大学 一种颅内大血管图像的处理方法、系统、电子设备及介质

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112184657A (zh) * 2020-09-24 2021-01-05 上海健康医学院 一种肺结节自动检测方法、装置及计算机系统
CN112801964B (zh) * 2021-01-20 2022-02-22 中国人民解放军总医院 肺部ct图像的多标签智能检测方法、装置、设备和介质
CN112967254A (zh) * 2021-03-08 2021-06-15 中国计量大学 基于胸部ct影像肺部疾病识别和检测方法
CN113838026B (zh) * 2021-09-22 2024-02-02 中南大学 非小细胞肺癌检测方法、装置、计算机设备和存储介质
CN115619810B (zh) * 2022-12-19 2023-10-03 中国医学科学院北京协和医院 一种前列腺分区分割方法、系统及设备

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106780460A (zh) * 2016-12-13 2017-05-31 杭州健培科技有限公司 一种用于胸部ct影像的肺结节自动检测系统
CN108447062A (zh) * 2018-02-01 2018-08-24 浙江大学 一种基于多尺度混合分割模型的病理切片非常规细胞的分割方法
CN108986067A (zh) * 2018-05-25 2018-12-11 上海交通大学 基于跨模态的肺结节检测方法
EP3540692A1 (en) * 2018-03-16 2019-09-18 Infervision A computed tomography pulmonary nodule detection method based on deep learning
CN110930418A (zh) * 2019-11-27 2020-03-27 江西理工大学 融合W-net和条件生成对抗网络的视网膜血管分割方法
CN112184657A (zh) * 2020-09-24 2021-01-05 上海健康医学院 一种肺结节自动检测方法、装置及计算机系统

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5987094A (en) * 1996-10-30 1999-11-16 University Of South Florida Computer-assisted method and apparatus for the detection of lung nodules
CN107016665B (zh) * 2017-02-16 2021-05-04 浙江大学 一种基于深度卷积神经网络的ct肺结节检测方法
CN108022238B (zh) * 2017-08-09 2020-07-03 深圳科亚医疗科技有限公司 对3d图像中对象进行检测的方法、计算机存储介质和系统
US20190122073A1 (en) * 2017-10-23 2019-04-25 The Charles Stark Draper Laboratory, Inc. System and method for quantifying uncertainty in reasoning about 2d and 3d spatial features with a computer machine learning architecture
CN108765369B (zh) * 2018-04-20 2023-05-02 平安科技（深圳）有限公司 肺结节的检测方法、装置、计算机设备和存储介质
CN111402254B (zh) * 2020-04-03 2023-05-16 杭州华卓信息科技有限公司 一种ct图像肺结节高性能自动检测方法及装置

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106780460A (zh) * 2016-12-13 2017-05-31 杭州健培科技有限公司 一种用于胸部ct影像的肺结节自动检测系统
CN108447062A (zh) * 2018-02-01 2018-08-24 浙江大学 一种基于多尺度混合分割模型的病理切片非常规细胞的分割方法
EP3540692A1 (en) * 2018-03-16 2019-09-18 Infervision A computed tomography pulmonary nodule detection method based on deep learning
CN108986067A (zh) * 2018-05-25 2018-12-11 上海交通大学 基于跨模态的肺结节检测方法
CN110930418A (zh) * 2019-11-27 2020-03-27 江西理工大学 融合W-net和条件生成对抗网络的视网膜血管分割方法
CN112184657A (zh) * 2020-09-24 2021-01-05 上海健康医学院 一种肺结节自动检测方法、装置及计算机系统

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114974518A (zh) * 2022-04-15 2022-08-30 浙江大学 多模态数据融合的肺结节影像识别方法及装置
CN114757942A (zh) * 2022-05-27 2022-07-15 南通大学 一种基于深度学习的多层螺旋ct对活动性肺结核的检测方法
CN115578309A (zh) * 2022-08-04 2023-01-06 云南师范大学 肺癌特征信息获取的方法、系统、电子设备和存储介质
CN115358976A (zh) * 2022-08-10 2022-11-18 北京医准智能科技有限公司 一种图像识别方法、装置、设备及存储介质
CN115578564A (zh) * 2022-10-25 2023-01-06 北京医准智能科技有限公司 实例分割模型的训练方法、装置、电子设备及存储介质
CN115578564B (zh) * 2022-10-25 2023-05-23 北京医准智能科技有限公司 实例分割模型的训练方法、装置、电子设备及存储介质
CN116228685A (zh) * 2023-02-07 2023-06-06 重庆大学 一种基于深度学习的肺结节检测与剔除方法
CN116228685B (zh) * 2023-02-07 2023-08-22 重庆大学 一种基于深度学习的肺结节检测与剔除方法
CN116934757A (zh) * 2023-09-18 2023-10-24 电子科技大学 一种用于肺结节假阳性删减的方法、设备及存储介质
CN116934757B (zh) * 2023-09-18 2023-11-21 电子科技大学 一种用于肺结节假阳性删减的方法、设备及存储介质
CN117542527A (zh) * 2024-01-09 2024-02-09 百洋智能科技集团股份有限公司 肺结节跟踪与变化趋势预测方法、装置、设备及存储介质
CN117542527B (zh) * 2024-01-09 2024-04-26 百洋智能科技集团股份有限公司 肺结节跟踪与变化趋势预测方法、装置、设备及存储介质
CN117809122A (zh) * 2024-02-29 2024-04-02 北京航空航天大学 一种颅内大血管图像的处理方法、系统、电子设备及介质

Also Published As

Publication number Publication date
CN112184657A (zh) 2021-01-05

Similar Documents

Publication Publication Date Title
WO2022063199A1 (zh) 一种肺结节自动检测方法、装置及计算机系统
CN110310287B (zh) 基于神经网络的危及器官自动勾画方法、设备和存储介质
CN108464840B (zh) 一种乳腺肿块自动检测方法及系统
CN108133476B (zh) 一种肺结节自动检测方法及系统
CN109636808B (zh) 一种基于全卷积神经网络的肺叶分割方法
CN110245657B (zh) 病理图像相似性检测方法及检测装置
CN108648182B (zh) 一种基于分子亚型的乳腺癌核磁共振图像肿瘤区域分割方法
CN111553892B (zh) 基于深度学习的肺结节分割计算方法、装置及系统
CN111462049B (zh) 一种乳腺超声造影视频中病灶区形态自动标注方法
CN111340130A (zh) 一种基于深度学习和影像组学的泌尿结石检测分类方法
CN108062749B (zh) 肛提肌裂孔的识别方法、装置和电子设备
CN112150442A (zh) 基于深度卷积神经网络及多实例学习的新冠诊断系统
WO2020238043A1 (zh) 基于UNet网络的肺叶分割方法、装置及计算机可读存储介质
CN109670489B (zh) 基于多实例学习的弱监督式早期老年性黄斑病变分类方法
CN114565761A (zh) 一种基于深度学习的肾透明细胞癌病理图像肿瘤区域的分割方法
Yao et al. Pneumonia detection using an improved algorithm based on faster r-cnn
CN113744183A (zh) 肺结节检测方法及系统
CN114092450A (zh) 一种基于胃镜检查视频的实时图像分割方法、系统、装置
Li et al. Developing an image-based deep learning framework for automatic scoring of the pentagon drawing test
CN112116559A (zh) 一种基于深度学习的数字病理图像智能分析方法
CN116883341A (zh) 一种基于深度学习的肝脏肿瘤ct图像自动分割方法
CN114359279B (zh) 图像处理方法、装置、计算机设备及存储介质
Kiyuna et al. Automatic classification of hepatocellular carcinoma images based on nuclear and structural features
Liu et al. TrEnD: A transformer‐based encoder‐decoder model with adaptive patch embedding for mass segmentation in mammograms
CN114972272A (zh) 一种基于Grad-CAM的新冠肺炎病灶分割方法

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: 21871568

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: 21871568

Country of ref document: EP

Kind code of ref document: A1