Brief Review — Deep-COVID: Predicting COVID-19 from chest X-ray images using deep transfer learning

Deep-COVID, COVID-Xray-5k Dataset is Proposed

Deep-COVID: Predicting COVID-19 from chest X-ray images using deep transfer learning,
Deep-COVID, by Snap Inc., Isfahan University of Medical Sciences, The University of Iowa, and Isfahan University of Technology
2020 J. MEDIA, Over 650 Citations (Sik-Ho Tsang @ Medium)
Image Classification, Medical Image Analysis, Medical Imaging

• A dataset of 5000 Chest X-rays is prepared from the publicly available datasets. Images exhibiting COVID-19 disease presence were identified by board-certified radiologist.
• Transfer learning on a subset of 2000 radiograms was used to train four popular convolutional neural networks, including ResNet18, ResNet50, SqueezeNet, and DenseNet-121, to identify COVID-19 disease in the analyzed chest X-ray images.

Outline

1. COVID-Xray-5k Dataset
2. Model Architectures
3. Results

1. COVID-Xray-5k Dataset

Three sample COVID-19 images, and the corresponding marked areas by the radiologist.

The above shows 3 COVID-19 images. (To me, the characteristics is subtle.)

Sample images from COVID-Xray-5k dataset. The images in the first row show 4 COVID-19 images. The images in the second row are 4 sample images of no-finding category in Non-COVID images from CheXpert. The images in the third and fourth rows give 8 sample images from other sub-categotries in CheXpert.

Since the number of Non-Covid images was very small in the (https://github.com/ieee8023/covid-chestxray-dataset) dataset, additional images were employed from the CheXpert dataset.

Number of images per category in COVID-Xray-5k dataset.

Finally, COVID-Xray-5k dataset is formed, as above.

2. Model Architecture

• 4 model architectures are tried.

2.1. ResNet18 & ResNet50

ResNet-18

2.2. SqueezeNet

SqueezeNet Fire Module

2.3. DenseNet

DenseNet Dense Block

• (Please feel free to visit ResNet, SqueezeNet, and DenseNet if interested.)

3. Results

The ROC curve of four CNN architectures on COVID-19 test set.

All models have a similar performance according to the AUC with the SqueezeNet achieving a slightly higher AUC than the other models. It is worth mentioning that for highly imbalanced test sets.

Comparison of sensitivity and specificity of four state-of-the-art deep neural networks.

The confidence interval of specificity rates are small (around 1%), since we have around 3000 samples for this class, whereas for the sensitivity rate, a slightly higher confidence interval (around 2.7%) is obtained because of the limited number of samples.

COVID-19 infected regions detected by ResNet18 model, in six chest X-ray images from the test set. Vertical sets give the Original images (top row), COVID-19 region heatmap (2nd row), heatmap overlaid on the image (3rd row), and the independent standard of radiologist-marked COVID-19 disease regions (bottom row).

• The likely regions of COVID-19 disease marked by the board-certified radiologist are shown on the last row.

The generated heatmaps show a good agreement with the radiologist-determined regions of the COVID-19 disease.

A brief paper review today, Happy Boxing Day !!!

Reference

[2020 J. MEDIA] [Deep-COVID]
Deep-COVID: Predicting COVID-19 from chest X-ray images using deep transfer learning

4.1. Biomedical Image Classification

2017 [ChestX-ray8] 2019 [CheXpert] 2020 [VGGNet for COVID-19] [Dermatology] [Deep-COVID] 2021 [CheXternal] [CheXtransfer]

==== My Other Previous Paper Readings ====