COVID-19 Prediction Models

Our COVID-19 prediction models are composed of a novel coronavirus pneumonia (NCP) recognition model trained by a large number of COVID-19 patient CT image data sets. It can accurately identify important clinical markers related to the characteristics of NCP lesions and help the radiologists and clinicians quickly diagnose and identify COVID-19 patients for early intervention and appropriate allocation of resources.

The diseases caused by the 2019 novel coronavirus (SARS-CoV-2) are collectively referred to as COVID-19. These diseases have a high mortality rate, which can cause fever, cough and other flu-like symptoms. Many affected patients will develop novel coronavirus pneumonia (NCP) and rapidly develop severe acute respiratory failure. Chest computed tomography (CT) radiography is an important tool for diagnosing lung diseases including NCP. Our COVID-19 prediction models based on deep learning algorithms use a large amount of clinical data and CT to establish NCP recognition models, which helps to distinguish NCP and common influenza or other types of pneumonia (such as viral pneumonia and bacterial pneumonia) can help radiologists and clinicians quickly diagnose and identify COVID-19 patients, so that clinicians can plan for early monitoring and management of these patients.

Model Components

  • Lung Lesion Segmentation Model
    Lung lesion segmentation model is a segmentation framework based on U-Net deep neural network, which solves the problem of semantic level and realizes end-to-end segmentation. In addition, this model is trained with more than 5,000 CT slices from NCP patients and common pneumonia patients. During the training process, Dice Coefficient and 10-fold cross-validation are used to evaluate and adjust the model. The accuracy of the final model can reach more than 97%, which can accurately segment the lung parenchyma part of the original lung CT image and identify the normal area and the lesion area.

    Figure 1 Lung Lesion Segmentation Model

  • Diagnosis Analysis Model
    The Diagnosis Analysis Model takes the lung lesion map generated by the segmentation network in the Lung Lesion Segmentation Model as input, and normalizes the CT image through data preprocessing for further diagnosis and prediction. Usually the original CT images from different sources may be collected by different imaging devices, so there is a lack of a unified standard. We standardize it through data preprocessing, which not only reduces the prediction error of the Diagnosis Analysis Model, but also provides better usability. Through internal verification, our Diagnosis Analysis Mode has an accuracy of 92%, a sensitivity of 95%, and a deliberate degree of 91%. It can accurately and quickly distinguish NCP from other common pneumonias.

    Figure 2 Diagnosis Analysis Model

Results and Delivery

  • Accurate
    Our COVID-19 prediction models are trained from more than 5,000 CT slices from nearly 300 pneumonia patients (including about 80 NCP patients). A large amount of training data and robust algorithms make our model's accuracy rate as high as 90% (sensitivity 92%; specificity 86%), which is significantly higher than manual recognition accuracy.
  • Efficient
    Supported by high-performance servers and robust algorithms, the COVID-19 prediction models can diagnose a large number of CT slices in a short period of time and keep working 24/7, which can greatly reduce the workload of doctors and improve their diagnosis effectiveness.
  • Applicable
    Our COVID-19 prediction model can accurately and quickly complete NPC screening, and make decision assistance based on AI screening results. According to the test results, our COVID-19 prediction model can help junior radiologists to improve the quality of diagnosis, so that their diagnostic level can be improved to be equivalent to that of senior radiologists in a short time.

CD ComputaBio provides AI-driven solutions for COVID-19 prediction and diagnosis according to clients' detailed requirements.

Related Solutions:

Online Inquiry

CD ComputaBio

Copyright © 2024 CD ComputaBio Inc. All Rights Reserved.