CN-122000068-A - ARDS prediction method and system based on multi-modal data and deep learning

Abstract

The invention provides a method and a system for predicting acute respiratory distress syndrome (Acute Respiratory Distress Syndrome, ARDS) based on multi-modal data and deep learning, relates to the technical field of computer aided diagnosis, and provides a computer algorithm for intelligently predicting the severity of an ARDS patient based on computed tomography (Computed Tomography, CT) imaging and clinical indexes so as to assist doctors in early intervention and treatment. The method has the technical key points that the method can effectively divide inflammation by using a 2.5D U-Net model and a sub-visual recognition algorithm, predicts ARDS by introducing a Extreme Gradient Boosting (XGBoost) algorithm after extracting the segmentation inflammation characteristics, realizes ARDS prediction based on imaging characteristics, can shorten diagnosis time, reduce death rate and has important significance for early diagnosis of ARDS. Experimental results show that the method provided by the invention has higher accuracy in predicting the ARDS severity, the Dice coefficient of lung inflammation segmentation is 90.2%, and the accuracy of ARDS severity prediction is 86.5%.

Inventors

• KANG KAI
• DU XUE
• GAO YANG
• WANG KUO
• ZHANG JIANNAN
• ZHANG WEITING
• QIU ZHAOWEN
• WANG ZHIMING
• YIN ZEHUA

Assignees

• 哈尔滨医科大学

Dates

Publication Date

20260508

Application Date

20260130

Claims (9)

1. 1. The ARDS prediction method based on the multi-mode data and the deep learning is characterized by comprising the following steps of: S100, collecting CT image data and clinical diagnosis and treatment data of ARDS patients; S200, carrying out normalization processing on the acquired CT image data, then adopting a 2.5D U-Net segmentation model to carry out data segmentation, namely simplifying a 3D segmentation task into 2D segmentation tasks with different visual angles, and then fusing the 2D segmentation tasks to obtain a final segmentation result; S300, after segmentation of CT image data is completed, deleting blood vessels and air ducts in a CT image by using a sub-visual recognition algorithm Laplacian feature mapping model for subsequent feature enhancement and feature extraction; s400, based on XGBoost model, fusing the CT image data with enhanced characteristics in step S300 and the clinical data collected in step S100, constructing ARDS prediction model and executing prediction analysis.
2. 2. The method of claim 1, wherein in step S100, the data acquisition object is defined as a patient who is diagnosed as ARDS by Berlin' S diagnostic criteria and who is in an intensive care unit for a period of time longer than 24 hours.
3. 3. The ARDS prediction method based on multi-modal data and deep learning as claimed in claim 2, wherein the clinical data in step S100 covers patient demographics including sex, age, height and weight, laboratory examination indexes including D-dimer count, white blood cell count, neutrophil absolute count, lymphocyte absolute count, C-reactive protein, IL-6 and procalcitonin level, respiratory support related parameters including current day respiratory support mode, mechanical ventilation duration, ventilator setting parameters, concentration and flow parameters of high flow oxygen therapy, arterial blood oxygen partial pressure and arterial oxygenation index, and patient prognosis related information including whether or not to succeed in offline and clinical outcome.
4. 4. The ARDS prediction method based on multi-mode data and deep learning according to claim 3, wherein the data normalization preprocessing in the step S200 comprises spatial normalization and signal normalization, the spatial normalization is used for unifying resolution and pixel size of CT image data and eliminating spatial dimension deviation caused by scanning equipment difference, the signal normalization is based on CT scanner lung window setting, and the signal intensity of each voxel is subjected to standardization processing to ensure consistency of signal amplitude.
5. 5. The ARDS prediction method based on multi-modal data and deep learning as claimed in claim 4, wherein in step S200, the 2.5D U-Net segmentation model performs 2D segmentation using a three-view collaborative segmentation strategy, the three views being cross-section, coronal plane and sagittal plane.
6. 6. The ARDS prediction method based on multi-modal data and depth learning as claimed in claim 5, wherein in step S300, 20000 voxels are randomly extracted from the remaining lung parenchyma area obtained by each CT scan, the CT signal median value of the sampled voxels is extracted by taking a baseline CT image as a reference, and the CT signal standard deviation of the healthy lung parenchyma area is calculated at the same time; For each CT image, the CT signal is truncated, so that the linear scaling of the original CT signal is enhanced, and the scanning level deviation is effectively eliminated; wherein ENHANCED CT represents enhanced CT, original CT represents original CT, and baseline represents baseline.
7. 7. The ARDS prediction method based on multi-mode data and deep learning according to claim 6 is characterized in that in step S400, K-means clustering algorithm collaborative prediction is further introduced, prediction is completed based on label information of a preset number of training samples closest to a new data point by locating the samples, and if a regression prediction scene is adopted, a prediction value average value of K nearest neighbor samples is adopted as a final prediction result, so that prediction stability and precision are improved.
8. 8. The ARDS prediction system based on the multi-mode data and the deep learning is characterized by comprising a series of functional program modules, wherein the program modules and the steps of the prediction method according to any one of claims 1-7 form a one-to-one corresponding adaptation relation, and when the system is operated, all the steps of the ARDS prediction method based on the multi-mode data and the deep learning can be completely executed through the cooperative scheduling of the functional modules.
9. 9. A computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program configured to implement the steps of the multi-modal data and deep learning based ARDS prediction method of any one of claims 1-7 when called by a processor.

Description

ARDS prediction method and system based on multi-modal data and deep learning Technical Field The invention relates to the technical field of computer aided diagnosis, in particular to an ARDS prediction method and system based on multi-mode data and deep learning. Background ARDS is a serious clinical manifestation of acute hypoxia respiratory failure, is a critical respiratory disease co-induced by intrapulmonary and/or extrapulmonary factors, and has the core pathophysiological characteristics of refractory hypoxia, chang Ji is caused by various critical diseases such as serious infection, wound, burn, shock and the like. The syndrome has the remarkable characteristics of hidden onset, rapid progress and high clinical mortality, and the diagnosis standard is well-defined in 2012 that dyspnea symptoms are progressively aggravated within 1 week after the occurrence of causes, chest X-Ray (CXR) or CT shows diffuse double-lung infiltration shadow, and dyspnea cannot be explained by other causes such as atelectasis, lung tumor or lung nodule. In recent years, with the global pandemic of new coronavirus infections (Coronavirus Disease 2019, COVID-19), the total number of ARDS patients caused by COVID-19 has increased greatly, with the death rate remaining high, leading to a broad focus in the world-wide medical field. The relevant study data showed that COVID-19 related ARDS had a mortality rate of about 26% -50% of which about 25% were mild cases and about 75% progressed to moderate to severe ARDS, while ARDS patients constituted about 5% of the total number of ICU mechanically ventilated patients suggesting that they constituted an extremely high disease burden in the ICU patient population. In view of the fact that the diagnosis standard of ARDS (such as increased pulmonary vascular permeability and diffuse alveolar damage) is difficult to determine in clinical practice, it is considered that accurate CT imaging segmentation can advance ARDS intrapulmonary assessment standardization to a new step and combine clinical parameters, so that early and accurate early warning of ARDS is achieved, and the method has become a key link for optimizing clinical treatment strategies and reducing mortality. At present, the research focus of ARDS is gradually changed from simple support treatment to early identification and accurate intervention of high-risk patients, ARDS and the severe tendency thereof are identified as early as possible and accurately in ICU inpatients, decision basis can be provided for early implementation of lung protective ventilation strategies and conservative liquid management schemes, and the method has important clinical significance. Existing studies have demonstrated that a variety of laboratory biomarkers and imaging quantitative features can be used to construct ARDS predictive models. For example, the feasibility of monitoring ARDS from CT images based on radiology or traditional quantitative analysis methods has been verified, and studies have indicated that the area under the curve (Area Under the Curve, AUC) of the upper lung affected area at the time of admission to the hospital is statistically significant in predicting whether ARDS patients require extracorporeal membrane oxygenation. In addition, a calculation method based on CT image data and gas exchange parameters has been used for deducing the mechanical ventilation requirement in the ARDS patient lung function recovery process, a nationwide study is performed to construct a deep learning (DEEP LEARNING, DL) model by quantitative CT features and initial clinical indexes for predicting the severe degree of COVID-19, the result shows that the model has higher accuracy and stability in disease classification, and a rapid and accurate pneumonia inflammation segmentation method is provided by studying and providing a rapid and accurate pneumonia inflammation segmentation method by decomposing a 3D segmentation problem into three 2D problems, so that the segmentation precision is greatly improved while the model calculation complexity is remarkably reduced. The ARDS prediction research based on imaging features and DL has preliminarily achieved effects, and experimental results show that the accuracy of a prediction model constructed based on image features can reach 82.4%, the severity of ARDS can be effectively estimated, and reliable intelligent auxiliary support is provided for clinical diagnosis and treatment decision. However, at present, research on artificial intelligence (ARTIFICIAL INTELLIGENCE, AI) based on CT imaging at home and abroad is focused on focus segmentation and diagnosis, and a multi-dimensional comprehensive prediction model is not fully constructed by combining clinical indexes and biological markers, so that the application value of the AI technology in clinical early warning and prospective intervention is limited to a certain extent. Most of the existing early warning models belong to static diagnosis or s