search

xiaoxuegao499 / LA-DNN-for-COVID-19-diagnosis

Online COVID-19 diagnosis with chest CT images: Lesion-attention deep neural networks
64 stars 18 forks source link

readme

LA-DNN for COVID-19 diagnosis

:fire: NEWS :fire:

1. Background

Chest (computed tomography) CT scanning is one of the most important technologies for COVID-19 diagnosis in the current clinical practice, which motivates more concerted efforts in developing AI-based diagnostic tools to alleviate the enormous burden on the medical system. We develop a lesion-attention deep neural network (LA-DNN) to predict COVID-19 positive or negative with a richly annotated chest CT image dataset. The CT image dataset contains 746 public chest CT images of COVID-19 patients collected from over 760 preprints, and the data annotations are accompanied with the textual radiology reports. We extract two types of important information from these annotations: One is the flag of whether an image indicates a positive or negative case of COVID-19, and the other is the description of five lesions on the CT images associated with the positive cases. The proposed data-driven LA-DNN model focuses on the primary task of binary classification for COVID-19 diagnosis, while an auxiliary multi-label learning task is implemented simultaneously to draw the model's attention to the five lesions of COVID-19 during the training. The joint task learning process makes it a highly sample-efficient deep model that can learn COVID-19 radiology features effectively with very limited samples. Our code is public in ./Code/.


Figure 1. The architecture of the proposed lesion-attention deep neural networks.

2. Data

3. Online Diagnosis System

An online system has been developed for fast online diagnoses using CT images at the web address https://www.covidct.cn/.

:satisfied: Welcome to visit!!! :satisfied:


Figure 2. Navigation bar of Online Diagnosis System.

4. Results

4.1 ROC & PRC


Figure 3. Performance of our proposed LA-DNN model for COVID-19 diagnosis in comparison with the baseline (Left: ROC curves, Right: Precision-recall curves; This is the latest result after adding new data).

4.2 Lesion attention map


Figure 4. Grad-CAM++ visualization for the baseline and our LA-DNN model with the backbone net of DenseNet-169 (Column 1 represents the original CT scans; Columns 2 and 3 are the class activation maps of the baseline; Columns 4 and 5 are the class activation maps of our LA-DNN model).

4.3 Visualization of the primary vs. auxiliary tasks


Figure 5. Plots of the pairwise relationships among the five lesions on making the final binary classification of COVID-19.

5. Dependence

6. More Information

We uploaded a video on YouTube to introduce our project. The web address is https://youtu.be/MripiZ-1pHU.

7. Citation

The details of our model can be found in this preprint: Online COVID-19 diagnosis with chest CT images: Lesion-attention deep neural networks.

Please cite our paper if you find this work useful:

    @inproceedings{liu2020fast,
      title={A fast online COVID-19 diagnostic system with chest CT scans},
      author={Liu, Bin and Gao, Xiaoxue and He, Mengshuang and Liu, Lin and Yin, Guosheng},
      booktitle={Proceedings of KDD},
      volume={2020},
      year={2020}
    }
    @article {Liu2020.05.11.20097907,
      author = {Liu, Bin and Gao, Xiaoxue and He, Mengshuang and Lv, Fengmao and Yin, Guosheng},
      title = {Online COVID-19 diagnosis with chest CT images: Lesion-attention deep neural networks},
      year = {2020},
      doi = {10.1101/2020.05.11.20097907},
      publisher = {Cold Spring Harbor Laboratory Press},
      URL = {https://www.medrxiv.org/content/early/2020/05/14/2020.05.11.20097907},
      eprint = {https://www.medrxiv.org/content/early/2020/05/14/2020.05.11.20097907.full.pdf},
      journal = {medRxiv},
    }