[论文解读] POCOVID-Net: Automatic Detection of COVID-19 From a New Lung Ultrasound Imaging Dataset (POCUS)

Jannis Born, Gabriel Brändle+5 more|arXiv (Cornell University)|Apr 25, 2020

Ultrasound in Clinical Applications参考文献 66被引用 158

一句话总结

本论文提出 POCOVID-Net，一种在 1103-lung-POCUS 图像数据集上训练的 CNN，用于检测 COVID-19，达到逐帧 89% 的准确率和视频级 92% 的准确率，且对 COVID-19 的灵敏度为 96%，并提供一个用于数据贡献和预测的开放网页服务。

ABSTRACT

With the rapid development of COVID-19 into a global pandemic, there is an ever more urgent need for cheap, fast and reliable tools that can assist physicians in diagnosing COVID-19. Medical imaging such as CT can take a key role in complementing conventional diagnostic tools from molecular biology, and, using deep learning techniques, several automatic systems were demonstrated promising performances using CT or X-ray data. Here, we advocate a more prominent role of point-of-care ultrasound imaging to guide COVID-19 detection. Ultrasound is non-invasive and ubiquitous in medical facilities around the globe. Our contribution is threefold. First, we gather a lung ultrasound (POCUS) dataset consisting of 1103 images (654 COVID-19, 277 bacterial pneumonia and 172 healthy controls), sampled from 64 videos. This dataset was assembled from various online sources, processed specifically for deep learning models and is intended to serve as a starting point for an open-access initiative. Second, we train a deep convolutional neural network (POCOVID-Net) on this 3-class dataset and achieve an accuracy of 89% and, by a majority vote, a video accuracy of 92% . For detecting COVID-19 in particular, the model performs with a sensitivity of 0.96, a specificity of 0.79 and F1-score of 0.92 in a 5-fold cross validation. Third, we provide an open-access web service (POCOVIDScreen) that is available at: https://pocovidscreen.org. The website deploys the predictive model, allowing to perform predictions on ultrasound lung images. In addition, it grants medical staff the option to (bulk) upload their own screenings in order to contribute to the growing public database of pathological lung ultrasound images. Dataset and code are available from: https://github.com/jannisborn/covid19_pocus_ultrasound. NOTE: This preprint is superseded by our paper in Applied Sciences: https://doi.org/10.3390/app11020672

研究动机与目标

• Motivate the use of point-of-care ultrasound (POCUS) for rapid COVID-19 screening and supplement existing molecular tests.
• Provide a first open dataset of lung ultrasound recordings labeled for COVID-19, pneumonia, and healthy controls.
• Develop and validate a CNN (POCOVID-Net) for frame-level and video-level COVID-19 detection.
• Offer an open-access web service (POCOVIDScreen) to predict on ultrasound images and collect new data.

提出的方法

• Construct a 3-class lung POCUS image dataset (COVID-19, pneumonia, healthy) from 64 videos, resulting in 1103 images after processing.
• Use a VGG-16-based CNN (POCOVID-Net) with a single hidden layer (64 neurons), dropout 0.5, batch normalization, pretrained on ImageNet, and fine-tune last three layers.
• Train with cross-entropy loss using Adam optimizer (lr=1e-4) and apply data augmentation (rotations, flips, shifts).
• Evaluate with 5-fold cross-validation ensuring video-wise data split to keep train/test disjoint.
• Compare against COVID-Net and a ResNet-based alternative, and perform frame-to-video aggregation via majority vote or average probabilities.

实验结果

研究问题

• RQ1 Can a CNN trained on lung ultrasound frames distinguish COVID-19 from pneumonia and healthy tissue?
• RQ2 What is the frame-level and video-level performance of POCOVID-Net on a diverse, open ultrasound dataset?
• RQ3 How does POCOVID-Net compare to existing X-ray/CT-based COVID-19 classifiers when applied to ultrasound data?
• RQ4 What is the impact of data diversity and augmentation on model generalization for ultrasound-based COVID-19 detection?

主要发现

Class	Sensitivity	Specificity	Precision	F1-score	Frames	Videos/Images
POCOVID-Net	0.96	0.79	0.88	0.92	654	39
Pneumonia	0.93	0.98	0.95	0.94	277	14
Healthy	0.55	0.98	0.78	0.62	172	11
COVID-Net	0.98	0.57	0.77	0.86	654	39
Pneumonia	0.89	0.98	0.95	0.92	277	14
Healthy	0.01	1.00	0.20	0.01	172	11

• POCOVID-Net achieves 0.89 overall accuracy and 0.82 balanced accuracy on frame-level classification across three classes.
• COVID-19 detection sensitivity is 0.96, specificity 0.79, and F1-score 0.92 at the frame level.
• Pneumonia is detected with 0.93 sensitivity, 0.98 specificity, and 0.94 F1-score; healthy class has 0.55 sensitivity and 0.98 specificity.
• Video-level aggregation yields 92% accuracy (balanced 0.84) for classifying COVID-19, pneumonia, or healthy.
• Compared with COVID-Net (X-ray) on the same data, POCOVID-Net shows higher accuracy and balanced accuracy (0.82 vs 0.63); COVID-Net struggles with the healthy class (0% sensitivity).
• The authors provide an open web service (POCOVIDScreen) for predictions and data submission, fostering community data growth.