WO2022063200A1 - 用于非小细胞肺癌预后生存预测的方法、介质及电子设备 - Google Patents

用于非小细胞肺癌预后生存预测的方法、介质及电子设备 Download PDF

Info

Publication number
WO2022063200A1
WO2022063200A1 PCT/CN2021/120101 CN2021120101W WO2022063200A1 WO 2022063200 A1 WO2022063200 A1 WO 2022063200A1 CN 2021120101 W CN2021120101 W CN 2021120101W WO 2022063200 A1 WO2022063200 A1 WO 2022063200A1
Authority
WO
WIPO (PCT)
Prior art keywords
survival
prognosis
layer
interest
lung cancer
Prior art date
Application number
PCT/CN2021/120101
Other languages
English (en)
French (fr)
Inventor
黄钢
聂生东
陈雯
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 WO2022063200A1 publication Critical patent/WO2022063200A1/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
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • 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
    • G06T2207/20081—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
    • G06T2207/30061—Lung
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30096—Tumor; Lesion
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/03—Recognition of patterns in medical or anatomical images
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Definitions

  • the invention relates to the field of computer-aided medicine, and relates to a computer electronic device, in particular to a method, a medium and an electronic device for predicting the prognosis and survival of non-small cell lung cancer.
  • NSCLC non-small cell lung cancer
  • the purpose of the present invention is to provide a method, medium and electronic device for predicting the prognosis and survival of non-small cell lung cancer with high prediction accuracy and convenient realization in order to overcome the above-mentioned defects of the prior art.
  • a method for prognostic survival prediction of non-small cell lung cancer comprising the steps of:
  • the deep learning-based prognosis survival model is a deep learning convolutional neural network model, including 5 convolution blocks, 1 fully connected layer and 1 classification layer, extracts tumor abstract features layer by layer, and obtains prognosis survival classification results.
  • the Bottleneck architecture is introduced into the middle three convolution blocks, and a fusion layer is added to the last convolution block based on the Bottleneck architecture.
  • extracting the region of interest specifically includes: reading the three-dimensional tumor markers and clinical data corresponding to the CT image to be predicted, extracting the structure information of the region of interest, comparing it with the CT image to be predicted, and intercepting the region of interest.
  • the data set used in the training of the prognostic survival model includes a test set, a verification set and a training set that do not intersect with each other, the network parameters are optimized based on the training set, and the generalization error during or after the training is estimated based on the verification set.
  • Hyperparameters to estimate model performance on a test set each sample in the dataset includes CT images, 3D tumor markers, clinical data, and survival.
  • the survival period includes long survival period and short survival period.
  • y i is the output of the neural network
  • t i is the positive solution label
  • ⁇ is the weight factor
  • the present invention also provides a computer-readable medium having stored thereon a computer program that, when executed by a processor, implements the steps of the method for prognosis survival prediction of non-small cell lung cancer as described above.
  • the present invention also provides an electronic device for predicting the prognosis and survival of non-small cell lung cancer, including:
  • the CT image acquisition module is used to acquire the CT image to be predicted, perform grayscale normalization processing on the CT image to be predicted, and extract the region of interest;
  • the module maintains a deep learning-based prognosis survival model, and based on the region of interest, uses the trained deep learning-based prognosis survival model to predict and obtain a corresponding prognosis survival period classification result;
  • the deep learning-based prognosis survival model is a deep learning convolutional neural network model, including 5 convolution blocks, 1 fully connected layer and 1 classification layer, extracts tumor abstract features layer by layer, and obtains prognosis survival classification results.
  • the Bottleneck architecture is introduced into the middle three convolution blocks, and a fusion layer is added to the last convolution block based on the Bottleneck architecture.
  • extracting the region of interest is specifically: reading the three-dimensional tumor markers and clinical data corresponding to the CT image to be predicted, extracting the structure information of the region of interest, comparing it with the CT image to be predicted, and intercepting. Obtain the region of interest.
  • the data set used in the training of the prognostic survival model includes a test set, a verification set and a training set that do not intersect with each other, the network parameters are optimized based on the training set, and the generalized parameters during or after training are estimated based on the verification set. Errors were reduced and hyperparameters were updated to estimate model performance on a test set, each sample in the dataset including CT images, 3D tumor markers, clinical data, and survival.
  • y i is the output of the neural network
  • t i is the positive solution label
  • ⁇ is the weight factor
  • the present invention has the following beneficial effects:
  • the present invention preprocesses CT images, then performs ROI extraction, and uses a trained deep learning-based prognosis survival model for survival prediction, which is convenient and simple to implement, improves prediction efficiency, and effectively assists clinicians to make correct plan decisions.
  • the present invention designs a new convolutional neural network model for prognosis and survival prediction of non-small cell lung cancer.
  • extracting tumor abstract features layer by layer and introducing Bottleneck architecture and fusion layer, the combination of extracted information from each layer is realized and the extracted features are improved. reliability, thereby improving the prediction accuracy.
  • the present invention designs an effective loss function for the deep learning convolutional neural network model, which improves the prediction accuracy of the model.
  • the present invention can accurately predict the specific situation of the prognosis of a single patient, assist clinicians to formulate suitable and effective treatment plans, realize precise medical treatment for different patients and different lesions, reduce subjective judgment, formulate objective evaluation standards, and improve the quality of patient prognosis, It has great potential and clinical application value to reduce the cost of seeing a doctor, avoid excessive treatment and waste of medical resources, and improve the medical level.
  • Fig. 1 is the training process schematic diagram of the prognosis survival model of the present invention
  • FIG. 2 is a schematic structural diagram of a deep learning convolutional neural network model adopted in the present invention.
  • Fig. 3 is the schematic diagram of convolution block 1 in the convolutional neural network model of the present invention.
  • FIG. 4 is a schematic diagram of convolution blocks 2-4 in the convolutional neural network model of the present invention.
  • FIG. 5 is a schematic diagram of the convolution block 5 in the convolutional neural network model of the present invention.
  • Deep learning is based on the deep superposition of artificial neural network (ANN) in machine learning, which is an extension of traditional neural network. It forms more abstract high-level features by combining low-level features to achieve classification and prediction. Due to its unique advantages, deep learning is developing rapidly in various fields of medicine, such as cancer prognosis analysis.
  • Radiomics is the use of automated algorithms to mine a large amount of feature information from the area of interest (ROI) of radiological images as research objects, and to extract effective key information through statistical methods for auxiliary diagnosis, classification or grading of diseases. .
  • Computed tomography (CT) is one of the common methods of lung examination and one of the important modalities in radiomics. There are good outcomes in terms of response or patient survival.
  • This embodiment provides a method for predicting the prognosis and survival of non-small cell lung cancer based on the above-mentioned foundation, including the following steps: acquiring a CT image to be predicted, performing grayscale normalization processing on the CT image to be predicted, and extracting a region of interest ; Based on the region of interest, use the trained deep learning-based prognosis survival model to predict and obtain the corresponding prognosis survival period classification result.
  • the deep learning-based prognosis survival model used in this method is a deep learning convolutional neural network (CNN) model, as shown in Figure 2, including 5 convolution blocks, 1 fully connected layer and 1
  • the classification layer extracts tumor abstract features layer by layer and obtains the classification results of prognosis and survival time.
  • the Bottleneck architecture is introduced into the middle three convolution blocks, and the last convolution block is added on the basis of the Bottleneck architecture. There are fusion layers.
  • the entire convolutional neural network model mainly includes a 3*3 convolutional layer, a 1*1 convolutional layer, a 2*2 maximum pooling layer, and a differentiable learning adaptive normalization layer ( Switchable Normalization, SN) and advanced activation function Exponential Linear Units (Exponential Linear Units, ELU).
  • Switchable Normalization, SN Switchable Normalization
  • ELU Exponential Linear Units
  • FIGs 3-5 the convolution block 1 consists of a two-dimensional convolution layer with a convolution kernel size of 3*3, an SN layer, and an ELU layer.
  • the structure of convolution block 2, convolution block 3, and convolution block 4 is the same, and the Bottleneck architecture is introduced.
  • This architecture has two channels and can combine the feature information between the two channels, including one convolution 2D convolutional layers with kernel size of 3*3, 3 2D convolutional layers with kernel size of 1*1, 3 SN layers, 3 ELU layers, 1 convolutional kernel size of 2*2
  • the max pooling layer of convolution block 5 adds a fusion layer on the basis of Bottleneck architecture, splicing the feature channels of convolution block 4 and convolution block 5, including a two-dimensional convolution with a convolution kernel of 3*3 layer, 1 2D convolution layer with 2*2 convolution kernel, 3 2D convolution layers with 1*1 convolution kernel, 3 SN layers, 3 ELU layers, 1 convolution kernel size It is a 3*3 max pooling layer.
  • the training process of the prognostic survival model includes:
  • Step 1 Data preprocessing: Obtain CT images, corresponding three-dimensional tumor markers and related clinical data, classify the CT images for survival, and perform grayscale normalization processing on the images.
  • 3D tumor markers are 3D tumors manually delineated by radiation oncologists with overall volume information.
  • the NSCLC patient dataset of Maastricht University was first obtained from The Cancer Imaging Archive (TCIA) database, including CT images of 422 NSCLC patients confirmed by histology or cytology data, 3D tumor total volume manually delineated by radiation oncologists, and clinical data; then, based on the length of patient survival in the data set and the status of survival at the time of survival, 165 cases (105 men, 60 women) were screened for 2 years. Survival period was used as the boundary to divide into long- and short-term survival groups (82 cases in the long-term survival group and 83 cases in the short-term survival group).
  • TCIA Cancer Imaging Archive
  • Step 2 Extract the region of interest.
  • matlab software was used to read the DICOM-RT files manually drawn by radiation oncologists, and the ROI structure information was extracted; then, the corresponding CT slices and the location of the tumor were found in the original image; finally, a 64*64 pixel ROI was intercepted.
  • Step 3 Data set division and data enhancement.
  • Step 4 Build a deep learning convolutional neural network model.
  • the training set in step 3 is fed into a convolutional neural network (CNN) as input to extract tumor abstract features through layer-by-layer training. ; Then, use the backpropagation algorithm and stochastic gradient descent algorithm to minimize the loss function to optimize the network parameters, and use the validation set to estimate the generalization error during or after training, update the hyperparameters; finally obtain the optimal prognosis survival model and save.
  • CNN convolutional neural network
  • the feature data of the shallow layer is first extracted through the first convolution block, and then the SN layer enters the ELU advanced activation layer to calculate the corresponding feature map output, which is used as the input of the next layer, the mathematical expression of the ELU activation function. for:
  • the feature map is sent to the second convolution block, and after convolution calculation layer by layer, it enters the pooling layer, and the feature map distinguishes The rate becomes the original 1/s, so as to gradually extract the high-level features of the image; then the obtained feature map is used as the input of the next layer of convolution block, and abstract features are extracted in turn; finally, the final classification result is output through the fully connected layer and the classification layer.
  • the hyperparameters of the network are adjusted through the validation set, and the backpropagation algorithm and the Adamax gradient descent optimization algorithm are used to supervised to minimize the loss function to obtain a convolutional neural network that optimizes the network connection weights.
  • the mathematical expression of the loss function used is:
  • y i is the output of the neural network
  • t i is the positive solution label
  • ⁇ is the weight factor
  • Step 5 Use the optimal prognostic survival model to predict the survival outcome.
  • the accuracy represents the probability of all prediction pairs of all samples
  • the sensitivity represents the probability that the actual positive samples are judged as positive samples
  • the specificity represents the probability that the actual negative samples are judged as negative samples
  • the AUC value is a value used to evaluate the binary classification A common indicator of model pros and cons. A higher AUC value usually indicates a better model.
  • This embodiment provides a computer-readable medium on which a computer program is stored, and when the computer program is executed by a processor, implements the steps of the method for prognosis and survival prediction of non-small cell lung cancer as described in Embodiment 1.
  • This embodiment provides an electronic device for predicting prognosis and survival of non-small cell lung cancer, including a CT image acquisition module and a prediction module, wherein the CT image acquisition module is used to acquire a CT image to be predicted, and grayscale the CT image to be predicted. degree normalization, and extract the region of interest; the prediction module maintains a deep learning-based prognostic survival model, and based on the region of interest, uses the trained deep learning-based prognostic survival model to predict and obtain the corresponding prognostic survival time classification result.
  • the deep learning-based prognosis survival model is a deep learning convolutional neural network model, including 5 convolution blocks, 1 fully connected layer and 1 classification layer, extracts tumor abstract features layer by layer, and obtains prognosis survival classification results. , Among the five convolution blocks, the Bottleneck architecture is introduced into the middle three convolution blocks, and a fusion layer is added to the last convolution block based on the Bottleneck architecture.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biophysics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Medical Informatics (AREA)
  • Multimedia (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Quality & Reliability (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)
  • Apparatus For Radiation Diagnosis (AREA)

Abstract

一种用于非小细胞肺癌预后生存预测的方法、介质及电子设备，所述方法包括：获取待预测CT影像，对该待预测CT影像进行灰度归一化处理，并提取感兴趣区域；基于所述感兴趣区域，采用训练好的基于深度学习的预后生存模型预测获得对应的预后生存期分类结果；所述基于深度学习的预后生存模型为深度学习卷积神经网络模型，包括5个卷积块、1个全连接层和1个分类层，逐层提取肿瘤抽象特征，并获得预后生存期分类结果，所述5个卷积块中，中间3个卷积块引入有Bottleneck架构，最后一个卷积块在Bottleneck架构的基础上添加有融合层。

Description

用于非小细胞肺癌预后生存预测的方法、介质及电子设备 技术领域
本发明涉及计算机辅助医学领域，涉及一种计算机电子设备，尤其是涉及一种用于非小细胞肺癌预后生存预测的方法、介质及电子设备。
背景技术
2018年国际癌症研究机构最新报告显示，肺癌是全球发病率和死亡率最高的癌症，其中，非小细胞肺癌(non-small cell lung cancer,NSCLC)患者占肺癌患病总人数的80％～85％，约3/4的患者发现时已处于中晚期，5年生存率较低。此外，由于肿瘤的异质性，不同的个体身上表现出不一样的治疗效果及预后，甚至在同一个体身上的肿瘤细胞也存在不同的特性和差异。因此，迫切需要一种精准客观且泛化性强的NSCLC预后生存预测系统，以期辅助临床医生高效治疗NSCLC患者，制定个性化的治疗和随访方案，进而提高其治愈率和存活率。
从目前国内外的研究现状来看，临床上研究人员普遍从临床试验、医疗记录中收集患者的年龄、性别、临床分期、吸烟史、组织病理类型、肿瘤标记物等量化指标；通过统计学中多元线性回归方法对临床特征与预后关系进行单因素和多因素生存分析，获得与NSCLC患者预后相关的预后因素。但通过这类方法获取的预后因素有限，可能不与当前治疗方法相符，且不能准确地预测及全面的评估单个患者的具体预后情况。为寻找更多相关的预后因素，研究人员采用机器学习方法(包括人工神经网络、决策树、随机森林等)从大量复杂的医学数据中挖掘关键特征，预测NSCLC患者的预后情况。这类方法在一定程度上提高了预测精度，但从医学图像中提取感兴趣特征需要经验丰富的医生来识别或描述，较为费时费力，且手工制作特征的有效性也存在不确定性。
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预测精度高且方便实现的用于非小细胞肺癌预后生存预测的方法、介质及电子设备。
本发明的目的可以通过以下技术方案来实现：
一种用于非小细胞肺癌预后生存预测的方法，该方法包括以下步骤：
获取待预测CT影像，对该待预测CT影像进行灰度归一化处理，并提取感兴趣区域；
基于所述感兴趣区域，采用训练好的基于深度学习的预后生存模型预测获得对应的预后生存期分类结果；
所述基于深度学习的预后生存模型为深度学习卷积神经网络模型，包括5个卷积块、1个全连接层和1个分类层，逐层提取肿瘤抽象特征，并获得预后生存期分类结果，所述5个卷积块中，中间3个卷积块引入有Bottleneck架构，最后一个卷积块在Bottleneck架构的基础上添加有融合层。
进一步地，提取所述感兴趣区域具体为：读取待预测CT影像对应的三维肿瘤标记及临床数据提取感兴趣区域结构信息，与待预测CT影像比对，截取获得所述感兴趣区域。
进一步地，所述预后生存模型训练时采用的数据集包括互不相交的测试集、验证集和训练集，基于训练集优化网络参数，基于验证集估计训练中或训练后的泛化误差，更新超参数，以测试集估计模型性能，所述数据集中的各样本包括CT影像、三维肿瘤标记、临床数据和生存期。
进一步地，所述生存期包括长生存期和短生存期。
进一步地，所述预后生存模型训练时使用的损失函数数学表达式为：
Figure PCTCN2021120101-appb-000001
其中y i是神经网络的输出，t i是正解标签，
Figure PCTCN2021120101-appb-000002
是交叉熵损失函数，
Figure PCTCN2021120101-appb-000003
是新添加的损失项，β是权重因子。
本发明还提供一种计算机可读介质，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如上所述用于非小细胞肺癌预后生存预测的方法的步骤。
本发明还提供一种用于非小细胞肺癌预后生存预测的电子设备，包括：
CT影像获取模块，用于获取待预测CT影像，对该待预测CT影像进行灰度归一化处理，并提取感兴趣区域；
预测模块，该模块维护一基于深度学习的预后生存模型，基于所述感兴趣区域，采用训练好的基于深度学习的预后生存模型预测获得对应的预后生存期分类结果；
所述基于深度学习的预后生存模型为深度学习卷积神经网络模型，包括5个卷积块、1个全连接层和1个分类层，逐层提取肿瘤抽象特征，并获得预后生存期分类结果，所述5个卷积块中，中间3个卷积块引入有Bottleneck架构，最后一个卷积块在Bottleneck架构的基础上添加有融合层。
进一步地，所述CT影像获取模块中，提取所述感兴趣区域具体为：读取待预测CT影像对应的三维肿瘤标记及临床数据提取感兴趣区域结构信息，与待预测CT影像比对，截取获得所述感兴趣区域。
进一步地，所述预测模块中，预后生存模型训练时采用的数据集包括互不相交的测试集、验证集和训练集，基于训练集优化网络参数，基于验证集估计训练中或训练后的泛化误差，更新超参数，以测试集估计模型性能，所述数据集中的各样本包括CT影像、三维肿瘤标记、临床数据和生存期。
进一步地，所述预后生存模型训练时使用的损失函数数学表达式为：
Figure PCTCN2021120101-appb-000004
其中y i是神经网络的输出，t i是正解标签，
Figure PCTCN2021120101-appb-000005
是交叉熵损失函数，
Figure PCTCN2021120101-appb-000006
是新添加的损失项，β是权重因子。
与现有技术相比，本发明具有如下有益效果：
1、本发明对CT影像进行预处理，而后进行ROI提取，利用训练好的基于深度学习的预后生存模型进行生存期预测，实现方便简单，提高预测效率，有效辅助临床医生做出正确方案决策。
2、本发明针对非小细胞肺癌预后生存预测设计了新型的卷积神经网络模型，通过逐层提取肿瘤抽象特征，并引入Bottleneck架构和融合层，实现各层提取信息的结合，提高所提取特征的可靠性，进而提高预测准确度。
3、本发明对深度学习卷积神经网络模型设计了有效的损失函数，提高了模型预测精度。
4、本发明可准确地预测单个患者的预后具体情况，协助临床医生制定合适 且有效的治疗方案，针对不同患者、不同病灶实现精准医疗，减少主观判断，制定客观评价标准，提高患者预后质量，降低看病成本，避免过度治疗及医疗资源的浪费，提高医疗水平，具有极大潜力与临床应用价值。
附图说明
图1为本发明预后生存模型的训练过程示意图；
图2为本发明采用的深度学习卷积神经网络模型的结构示意图；
图3为本发明卷积神经网络模型中卷积块1的示意图；
图4为本发明卷积神经网络模型中卷积块2-4的示意图；
图5为本发明卷积神经网络模型中卷积块5的示意图。
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
以深度学习根据机器学习中人工神经网络(Artificial neural network,ANN)的深度叠加进行学习，是传统神经网络的拓展，通过低层特征组合形成更加抽象的高层特征从而实现分类和预测。因自身的独特优势，深度学习在医学各个领域的发展越来越快，例如癌症的预后分析方面。影像组学是采用自动化算法从放射影像的感兴趣区域(Area of interest,ROI)内挖掘大量特征信息作为研究对象，通过统计学方法提取有效的关键信息，用于疾病的辅助诊断、分类或分级。计算机断层扫描(Computed tomography，CT)是肺部检查的常用手段之一，也是影像组学中的重要模态之一，易于采集与比对，在区分不同组织病理学特征的肿瘤病变和预测治疗反应或患者生存方面有很好的结果。
本实施例基于上述基础提供一种用于非小细胞肺癌预后生存预测的方法，包括以下步骤：获取待预测CT影像，对该待预测CT影像进行灰度归一化处理，并提取感兴趣区域；基于所述感兴趣区域，采用训练好的基于深度学习的预后生存模型预测获得对应的预后生存期分类结果。
该方法采用的所述基于深度学习的预后生存模型为深度学习卷积神经网络模型(Convolutional neural network，CNN)，如图2所示，包括5个卷积块、1个全连接层和1个分类层，逐层提取肿瘤抽象特征，并获得预后生存期分类结果，所述5个卷积块中，中间3个卷积块引入有Bottleneck架构，最后一个卷积块在Bottleneck架构的基础上添加有融合层。
本实施例中，整个卷积神经网络模型主要包括3*3的卷积层、1*1的卷积层、2*2的最大池化层、可微分学习的自适配归一化层(Switchable Normalization,SN)和高级激活函数指数线性单元(Exponential Linear Units,ELU)。5个卷积块的具体实施方式如图3-图5。如图3所示，卷积块1由1个卷积核大小为3*3的二维卷积层、1个SN层、1个ELU层构成。如图4所示，卷积块2、卷积块3、卷积块4的结构相同，均引入Bottleneck架构，该架构具有两个通道，能够合并两通道间的特征信息，包括1个卷积核大小为3*3的二维卷积层、3个卷积核大小为1*1的二维卷积层、3个SN层、3个ELU层、1个卷积核大小为2*2的最大池化层。如图5所示，卷积块5在Bottleneck架构的基础上添加了融合层，拼接卷积块4与卷积块5的特征通道，包括1个卷积核为3*3的二维卷积层、1个卷积核为2*2的二维卷积层、3个卷积核为1*1的二维卷积层、3个SN层、3个ELU层、1个卷积核大小为3*3的最大池化层。
如图1所示，预后生存模型的训练过程包括：
步骤1、数据预处理：获取CT影像、对应的三维肿瘤标记及相关临床数据，对CT影像进行生存情况分类，并对影像进行灰度归一化处理。三维肿瘤标记为放射肿瘤学家手动描绘的三维肿瘤，具有总体体积信息。
本实施例中，首先从癌症影像档案(The Cancer Imaging Archive,TCIA)数据库中获取马斯特里赫特大学的NSCLC患者数据集，包含422例经组织学或细胞学确认的NSCLC病人的CT影像数据、放射肿瘤学家手动描绘的三维肿瘤总体体积以及临床数据；然后，根据数据集中患者生存时间的长短、截至存活情况，筛选出165例数据(105名男性，60名女性)，以2年生存期为界限划分长短期生存组(长生存期组82例，短生存期组83例)；最后，对筛选出CT影像进行灰度归一化处理。
步骤2、提取感兴趣区域。
首先，采用matlab软件读取放射肿瘤专家手动描绘的DICOM-RT文件，提取其中ROI结构信息；然后，在原图像中找到其对应CT切片和肿瘤所在位置；最后，截取64*64像素大小的ROI。
步骤3、数据集划分与数据增强。
首先，将步骤2中获得ROI按照病人随机抽取20％作为测试集；然后，在剩下的数据中随机抽取10％作为验证集，其余作为训练集，测试集、验证集、训练集互不相交；最后对训练集、验证集进行平移、随机旋转、错切、放缩等操作以扩充数据。
步骤4、构建深度学习卷积神经网络模型。
将NSCLC患者生存预测转化为基于CT影像组学的二分类问题，首先，将步骤3中的训练集作为输入送进卷积神经网络(Convolutional neural network，CNN)，通过逐层训练提取肿瘤抽象特征；然后，利用反向传播算法和随机梯度下降算法最小化损失函数，从而优化网络参数，同时使用验证集估计训练中或训练后的泛化误差，更新超参数；最后获得最优预后生存模型并保存。
训练过程中，首先通过第一个卷积块提取浅层的特征数据，经SN层后进入ELU高级激活层计算得到相应的特征图输出，作为下一层的输入，ELU激活函数的数学表达式为：
Figure PCTCN2021120101-appb-000007
其中x代表输入，α是一个可调整的介于0～1之间的参数；然后，将特征图送入第二个卷积块，逐层经过卷积计算后进入池化层，特征图分辨率变为原来的1/s，从而逐步提取图像的高级特征；接着将获得的特征图作为下一层卷积块的输入，依次提取抽象特征；最后通过全连接层和分类层输出最终分类结果。
整个预后生存模型的训练过程中，通过验证集调整网络的超参数，利用反向传播算法和Adamax梯度下降优化算法有监督地最小化损失函数，获得最优化网络连接权重的卷积神经网络。所使用的损失函数数学表达式为：
Figure PCTCN2021120101-appb-000008
其中y i是神经网络的输出，t i是正解标签，
Figure PCTCN2021120101-appb-000009
是交叉熵损失函 数，
Figure PCTCN2021120101-appb-000010
是新添加的损失项，β是权重因子。
步骤5、利用最优预后生存模型预测生存结果。
首先，加载训练好的预后生存模型，将测试集输入模型，根据输出层的节点值，获得患者生存期分类结果；然后，计算准确性、敏感度、特异度、受试者工作特征曲线下面积(Area Under the Curve,AUC)值评估模型分类性能。其中，准确性表示所有样本全部预测对的概率，敏感度代表实际为正样本判断为正样本的概率，特异度代表实际为负样本判断为负样本的概率，AUC值就是一个用来评价二分类模型优劣的常用指标，AUC值越高通常表明模型的效果越好。
实施例2
本实施例提供一种计算机可读介质，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如实施例1所述用于非小细胞肺癌预后生存预测的方法的步骤。
实施例3
本实施例提供一种用于非小细胞肺癌预后生存预测的电子设备，包括CT影像获取模块和预测模块，其中，CT影像获取模块用于获取待预测CT影像，对该待预测CT影像进行灰度归一化处理，并提取感兴趣区域；预测模块维护一基于深度学习的预后生存模型，基于所述感兴趣区域，采用训练好的基于深度学习的预后生存模型预测获得对应的预后生存期分类结果。所述基于深度学习的预后生存模型为深度学习卷积神经网络模型，包括5个卷积块、1个全连接层和1个分类层，逐层提取肿瘤抽象特征，并获得预后生存期分类结果，所述5个卷积块中，中间3个卷积块引入有Bottleneck架构，最后一个卷积块在Bottleneck架构的基础上添加有融合层。
其余同实施例1。
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. 一种用于非小细胞肺癌预后生存预测的方法，其特征在于，该方法包括以下步骤：
    获取待预测CT影像，对该待预测CT影像进行灰度归一化处理，并提取感兴趣区域；
    基于所述感兴趣区域，采用训练好的基于深度学习的预后生存模型预测获得对应的预后生存期分类结果；
    所述基于深度学习的预后生存模型为深度学习卷积神经网络模型，包括5个卷积块、1个全连接层和1个分类层，逐层提取肿瘤抽象特征，并获得预后生存期分类结果，所述5个卷积块中，中间3个卷积块引入有Bottleneck架构，最后一个卷积块在Bottleneck架构的基础上添加有融合层。
  2. 根据权利要求1所述的用于非小细胞肺癌预后生存预测的方法，其特征在于，提取所述感兴趣区域具体为：读取待预测CT影像对应的三维肿瘤标记及临床数据提取感兴趣区域结构信息，与待预测CT影像比对，截取获得所述感兴趣区域。
  3. 根据权利要求1所述的用于非小细胞肺癌预后生存预测的方法，其特征在于，所述预后生存模型训练时采用的数据集包括互不相交的测试集、验证集和训练集，基于训练集优化网络参数，基于验证集估计训练中或训练后的泛化误差，更新超参数，以测试集估计模型性能，所述数据集中的各样本包括CT影像、三维肿瘤标记、临床数据和生存期。
  4. 根据权利要求3所述的用于非小细胞肺癌预后生存预测的方法，其特征在于，所述生存期包括长生存期和短生存期。
  5. 根据权利要求1所述的用于非小细胞肺癌预后生存预测的方法，其特征在于，所述预后生存模型训练时使用的损失函数数学表达式为：
    Figure PCTCN2021120101-appb-100001
    其中y i是神经网络的输出，t i是正解标签，
    Figure PCTCN2021120101-appb-100002
    是交叉熵损失函数，
    Figure PCTCN2021120101-appb-100003
    是新添加的损失项，β是权重因子。
  6. 一种计算机可读介质，其上存储有计算机程序，其特征在于，所述计算机程序被处理器执行时实现如权利要求1-5任一所述用于非小细胞肺癌预后生存预测的方法的步骤。
  7. 一种用于非小细胞肺癌预后生存预测的电子设备，其特征在于，包括：
    CT影像获取模块，用于获取待预测CT影像，对该待预测CT影像进行灰度归一化处理，并提取感兴趣区域；
    预测模块，该模块维护一基于深度学习的预后生存模型，基于所述感兴趣区域，采用训练好的基于深度学习的预后生存模型预测获得对应的预后生存期分类结果；
    所述基于深度学习的预后生存模型为深度学习卷积神经网络模型，包括5个卷积块、1个全连接层和1个分类层，逐层提取肿瘤抽象特征，并获得预后生存期分类结果，所述5个卷积块中，中间3个卷积块引入有Bottleneck架构，最后一个卷积块在Bottleneck架构的基础上添加有融合层。
  8. 根据权利要求7所述的用于非小细胞肺癌预后生存预测的电子设备，其特征在于，所述CT影像获取模块中，提取所述感兴趣区域具体为：读取待预测CT影像对应的三维肿瘤标记及临床数据提取感兴趣区域结构信息，与待预测CT影像比对，截取获得所述感兴趣区域。
  9. 根据权利要求7所述的用于非小细胞肺癌预后生存预测的电子设备，其特征在于，所述预测模块中，预后生存模型训练时采用的数据集包括互不相交的测试集、验证集和训练集，基于训练集优化网络参数，基于验证集估计训练中或训练后的泛化误差，更新超参数，以测试集估计模型性能，所述数据集中的各样本包括CT影像、三维肿瘤标记、临床数据和生存期。
  10. 根据权利要求7所述的用于非小细胞肺癌预后生存预测的电子设备，其特征在于，所述预后生存模型训练时使用的损失函数数学表达式为：
    Figure PCTCN2021120101-appb-100004
    其中y i是神经网络的输出，t i是正解标签，
    Figure PCTCN2021120101-appb-100005
    是交叉熵损失函数，
    Figure PCTCN2021120101-appb-100006
    是新添加的损失项，β是权重因子。
PCT/CN2021/120101 2020-09-24 2021-09-24 用于非小细胞肺癌预后生存预测的方法、介质及电子设备 WO2022063200A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011018497.1 2020-09-24
CN202011018497.1A CN112184658B (zh) 2020-09-24 2020-09-24 用于非小细胞肺癌预后生存预测的方法、介质及电子设备

Publications (1)

Publication Number Publication Date
WO2022063200A1 true WO2022063200A1 (zh) 2022-03-31

Family

ID=73943403

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/120101 WO2022063200A1 (zh) 2020-09-24 2021-09-24 用于非小细胞肺癌预后生存预测的方法、介质及电子设备

Country Status (2)

Country Link
CN (1) CN112184658B (zh)
WO (1) WO2022063200A1 (zh)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115083574A (zh) * 2022-08-22 2022-09-20 四川大学华西医院 癌症预后生存预测方法、系统、计算机设备及存储介质
CN115132275A (zh) * 2022-05-25 2022-09-30 西北工业大学 基于端到端三维卷积神经网络预测egfr基因突变状态方法
CN115984622A (zh) * 2023-01-10 2023-04-18 深圳大学 基于多模态和多示例学习分类方法、预测方法及相关装置
CN116228753A (zh) * 2023-05-06 2023-06-06 中山大学孙逸仙纪念医院 肿瘤预后评估方法、装置、计算机设备和存储介质
CN116664953A (zh) * 2023-06-28 2023-08-29 北京大学第三医院（北京大学第三临床医学院） 2.5d肺炎医学ct影像分类装置及设备
CN116665017A (zh) * 2023-07-28 2023-08-29 神州医疗科技股份有限公司 一种基于影像组学的前列腺癌预测系统及构建方法
CN117173167A (zh) * 2023-11-02 2023-12-05 中日友好医院(中日友好临床医学研究所) 影像组学机器学习生存模型预测肿瘤预后的方法及装置

Families Citing this family (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112184658B (zh) * 2020-09-24 2023-11-24 上海健康医学院 用于非小细胞肺癌预后生存预测的方法、介质及电子设备
CN112819768B (zh) * 2021-01-26 2022-06-17 复旦大学 基于dcnn的癌症全视野数字病理切片生存分析方法
CN113688942A (zh) * 2021-02-09 2021-11-23 四川大学 基于深度学习的头侧片腺样体图像自动评估方法和装置
CN112927203A (zh) * 2021-02-25 2021-06-08 西北工业大学深圳研究院 基于多序列mri全局信息的胶质瘤患者术后生存期预测方法
CN113222887A (zh) * 2021-03-03 2021-08-06 复旦大学附属华山医院 基于深度学习的纳米铁标记神经干细胞示踪的方法
CN112991295B (zh) * 2021-03-12 2023-04-07 中国科学院自动化研究所 基于深度学习的淋巴结转移影像分析系统、方法及设备
CN112861994B (zh) * 2021-03-12 2023-04-28 中国科学院自动化研究所 基于Unet迁移学习的胃印戒细胞癌图像智能分类系统
CN113257413B (zh) * 2021-06-22 2021-10-19 安翰科技(武汉)股份有限公司 基于深度学习的癌症预后生存预测方法、设备及存储介质
CN113689382B (zh) * 2021-07-26 2023-12-01 北京知见生命科技有限公司 基于医学影像和病理图像的肿瘤术后生存期预测方法及系统
CN114283938A (zh) * 2021-11-25 2022-04-05 湖南大学 一种基于CNN-XGBoost的胶质母细胞瘤预后预测方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109658411A (zh) * 2019-01-21 2019-04-19 杭州英库医疗科技有限公司 一种基于ct影像学特征与非小细胞肺癌患者预后情况的相关性分析方法
US20190254611A1 (en) * 2018-02-21 2019-08-22 Case Western Reserve University Predicting disease recurrence following trimodality therapy in non-small cell lung cancer using computed tomography derived radiomic features and clinico-patholigic features
CN110956626A (zh) * 2019-12-09 2020-04-03 北京推想科技有限公司 一种基于图像的预后评估方法及装置
CN111354442A (zh) * 2018-12-20 2020-06-30 中国医药大学附设医院 肿瘤影像深度学习辅助子宫颈癌患者预后预测系统、方法
CN111462042A (zh) * 2020-03-03 2020-07-28 西北工业大学 癌症预后分析方法及系统
CN112184658A (zh) * 2020-09-24 2021-01-05 上海健康医学院 用于非小细胞肺癌预后生存预测的方法、介质及电子设备

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105005714A (zh) * 2015-06-18 2015-10-28 中国科学院自动化研究所 一种基于肿瘤表型特征的非小细胞肺癌预后方法
WO2020176762A1 (en) * 2019-02-27 2020-09-03 University Of Iowa Research Foundation Methods and systems for image segmentation and analysis
CN110472629B (zh) * 2019-08-14 2022-03-25 青岛大学附属医院 一种基于深度学习的病理图像自动识别系统及其训练方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190254611A1 (en) * 2018-02-21 2019-08-22 Case Western Reserve University Predicting disease recurrence following trimodality therapy in non-small cell lung cancer using computed tomography derived radiomic features and clinico-patholigic features
CN111354442A (zh) * 2018-12-20 2020-06-30 中国医药大学附设医院 肿瘤影像深度学习辅助子宫颈癌患者预后预测系统、方法
CN109658411A (zh) * 2019-01-21 2019-04-19 杭州英库医疗科技有限公司 一种基于ct影像学特征与非小细胞肺癌患者预后情况的相关性分析方法
CN110956626A (zh) * 2019-12-09 2020-04-03 北京推想科技有限公司 一种基于图像的预后评估方法及装置
CN111462042A (zh) * 2020-03-03 2020-07-28 西北工业大学 癌症预后分析方法及系统
CN112184658A (zh) * 2020-09-24 2021-01-05 上海健康医学院 用于非小细胞肺癌预后生存预测的方法、介质及电子设备

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115132275A (zh) * 2022-05-25 2022-09-30 西北工业大学 基于端到端三维卷积神经网络预测egfr基因突变状态方法
CN115132275B (zh) * 2022-05-25 2024-02-27 西北工业大学 基于端到端三维卷积神经网络预测egfr基因突变状态方法
CN115083574A (zh) * 2022-08-22 2022-09-20 四川大学华西医院 癌症预后生存预测方法、系统、计算机设备及存储介质
CN115083574B (zh) * 2022-08-22 2022-12-06 四川大学华西医院 癌症预后生存预测方法、系统、计算机设备及存储介质
CN115984622A (zh) * 2023-01-10 2023-04-18 深圳大学 基于多模态和多示例学习分类方法、预测方法及相关装置
CN115984622B (zh) * 2023-01-10 2023-12-29 深圳大学 基于多模态和多示例学习分类方法、预测方法及相关装置
CN116228753A (zh) * 2023-05-06 2023-06-06 中山大学孙逸仙纪念医院 肿瘤预后评估方法、装置、计算机设备和存储介质
CN116664953A (zh) * 2023-06-28 2023-08-29 北京大学第三医院（北京大学第三临床医学院） 2.5d肺炎医学ct影像分类装置及设备
CN116665017A (zh) * 2023-07-28 2023-08-29 神州医疗科技股份有限公司 一种基于影像组学的前列腺癌预测系统及构建方法
CN117173167A (zh) * 2023-11-02 2023-12-05 中日友好医院(中日友好临床医学研究所) 影像组学机器学习生存模型预测肿瘤预后的方法及装置

Also Published As

Publication number Publication date
CN112184658A (zh) 2021-01-05
CN112184658B (zh) 2023-11-24

Similar Documents

Publication Publication Date Title
WO2022063200A1 (zh) 用于非小细胞肺癌预后生存预测的方法、介质及电子设备
Santos et al. Artificial intelligence, machine learning, computer-aided diagnosis, and radiomics: advances in imaging towards to precision medicine
US11416716B2 (en) System and method for automatic assessment of cancer
Neal Joshua et al. 3D CNN with visual insights for early detection of lung cancer using gradient-weighted class activation
CN108464840B (zh) 一种乳腺肿块自动检测方法及系统
Kasinathan et al. Cloud-based lung tumor detection and stage classification using deep learning techniques
Zhong et al. Cancer image classification based on DenseNet model
CN112270666A (zh) 基于深度卷积神经网络的非小细胞肺癌病理切片识别方法
Zewdie et al. Classification of breast cancer types, sub-types and grade from histopathological images using deep learning technique
Elayaraja et al. An efficient approach for detection and classification of cancer regions in cervical images using optimization based CNN classification approach
EP4138091A1 (en) Cancer prognosis
Wankhade et al. Lung cell cancer identification mechanism using deep learning approach
CN115690056A (zh) 基于her2基因检测的胃癌病理图像分类方法及系统
CN114693671A (zh) 基于深度学习的肺结节半自动分割方法、装置、设备及介质
KR20230097127A (ko) 주의-기반 다중 인스턴스 학습
Nalla et al. Influence of Convolutional Neural Network Depth on the Efficacy of Automated Breast Cancer Screening Systems
Alshayeji et al. Lung cancer classification and identification framework with automatic nodule segmentation screening using machine learning
Babu et al. ICMFKC with optimize XGBoost classification for breast cancer image screening and detection
Baughan Artificial Intelligence for Breast Cancer Risk Assessment in Mammography and Methods for Dataset Balancing and Distribution Sampling
Suganya et al. An Automated Lion-Butterfly Optimization (LBO) based Stacking Ensemble Learning Classification (SELC) Model for Lung Cancer Detection
Karthikeyan et al. Android Applications for Lung Nodules Classification Using Convolutional Neural Network
Seth Automated localization of breast ductal carcinoma in situ in whole slide images
Farrag Towards a trustworthy data-driven clinical decision support system: breast cancer use-case
Rajesh et al. Prostate Cancer Detection using Radiomics-based Feature Analysis with ML Algorithms and MR Images
Sajiv et al. Analysis of Breast Cancer Recognition in Histopathological Images using Convolutional Neural Network

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: 21871569

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: 21871569

Country of ref document: EP

Kind code of ref document: A1