Sample-Efficient Deep Learning for COVID-19 Diagnosis Based on CT Scans

  1. Xuehai He
  2. Xingyi Yang
  3. Shanghang Zhang
  4. Jinyu Zhao
  5. Yichen Zhang
  6. Eric Xing
  7. Pengtao Xie

Reviewed by

Read the full article

Listed in

Log in to save this article

Abstract

Coronavirus disease 2019 (COVID-19) has infected more than 1.3 million individuals all over the world and caused more than 106,000 deaths. One major hurdle in controlling the spreading of this disease is the inefficiency and shortage of medical tests. There have been increasing efforts on developing deep learning methods to diagnose COVID-19 based on CT scans. However, these works are difficult to reproduce and adopt since the CT data used in their studies are not publicly available. Besides, these works require a large number of CTs to train accurate diagnosis models, which are difficult to obtain. In this paper, we aim to address these two problems. We build a publicly-available dataset containing hundreds of CT scans positive for COVID-19 and develop sample-efficient deep learning methods that can achieve high diagnosis accuracy of COVID-19 from CT scans even when the number of training CT images are limited. Specifically, we propose a Self-Trans approach, which synergistically integrates contrastive self-supervised learning with transfer learning to learn powerful and unbiased feature representations for reducing the risk of overfitting. Extensive experiments demonstrate the superior performance of our proposed Self-Trans approach compared with several state-of-the-art baselines. Our approach achieves an F1 of 0.85 and an AUC of 0.94 in diagnosing COVID-19 from CT scans, even though the number of training CTs is just a few hundred.

Article activity feed

  1. ScreenIT

    SciScore for 10.1101/2020.04.13.20063941: (What is this?)

    Please note, not all rigor criteria are appropriate for all manuscripts.

    Table 1: Rigor

    NIH rigor criteria are not applicable to paper type.

    Table 2: Resources

    Experimental Models: Organisms/Strains
    SentencesResources
    To study the second factor — neural architectures, we experiment different backbone networks, including VGG16 [47], ResNet18 [46], ResNet50 [46], DenseNet-121 [48], DenseNet-169 [48], EfficientNet-b0 [49], and EfficientNet-b1 [49].
    VGG16
    suggested: None
    DenseNet-169
    suggested: None
    Software and Algorithms
    SentencesResources
    In our case, we can take a classic neural architecture such as ResNet [46] and its weights pretrained on large-scale image classification datasets such as ImageNet, then fine-tune it on the COVID19-CT dataset, with the goal of transferring the images and classes labels in ImageNet into our task for mitigate the deficiency of COVID-19 CTs.
    ResNet
    suggested: (RESNET, RRID:SCR_002121)

    Results from OddPub: Thank you for sharing your code and data.

    Results from LimitationRecognizer: An explicit section about the limitations of the techniques employed in this study was not found. We encourage authors to address study limitations.

    Results from TrialIdentifier: No clinical trial numbers were referenced.

    Results from Barzooka: We did not find any issues relating to the usage of bar graphs.

    Results from JetFighter: Please consider improving the rainbow (“jet”) colormap(s) used on page 9. At least one figure is not accessible to readers with colorblindness and/or is not true to the data, i.e. not perceptually uniform.

    Results from rtransparent:
    • Thank you for including a conflict of interest statement. Authors are encouraged to include this statement when submitting to a journal.
    • Thank you for including a funding statement. Authors are encouraged to include this statement when submitting to a journal.
    • No protocol registration statement was detected.

    About SciScore

    SciScore is an automated tool that is designed to assist expert reviewers by finding and presenting formulaic information scattered throughout a paper in a standard, easy to digest format. SciScore checks for the presence and correctness of RRIDs (research resource identifiers), and for rigor criteria such as sex and investigator blinding. For details on the theoretical underpinning of rigor criteria and the tools shown here, including references cited, please follow this link.

  2. Version published to 10.1101/2020.04.13.20063941v1 on medRxiv