Introduction
Motivation
Approach
Contribution
Outline
Related work
CNN and medical image classification
Pneumonia and COVID-19
COVID-19 diagnosis using chest CT scans
COVID-19 diagnosis using chest X-ray images
Summary
Related work | Technique | Dataset | Accuracy (%) |
---|---|---|---|
Wang et al. [21] | Inception transfer learning | Chest CT scan | 85.2 |
Sethy et al. [32] | Deep learning | Chest X-ray images | 95.38 |
Azemin et al. [20] | Deep learning | Chest X-ray images | 71.9 |
Oh et al. [24] | Statistical method | Chest X-ray images | 88.9 |
XCOVNet model | CNN | Chest X-ray images | 98.44 |
XCOVNet method
Data engineering
X-ray image type | X-ray front posteroanterior view | Dataset-1 image count [36] | Dataset-2 image count [37] |
---|---|---|---|
COVID-19 positive | \(\checkmark\) | 196 | – |
\(\times\) | 388 | – | |
COVID-19 negative | \(\checkmark\) | – | 1583 |
\(\times\) | – | – | |
Other disease | – | 366 | 4273 |
XCOVNet CNN architecture
Parameters | Value |
---|---|
Number of convolutional layers | 4 |
Maximum number of pooling layers | 3 |
Number of filters at convolutional layer | 32, 64, 128 |
Kernel window size at convolutional layer | \(3 \times 3\) |
Kernel window size at maximum pool layer | \(2 \times 2\) |
Number of strides at convolutional layer | 1 |
Number of strides at maximum pool layer | 1 |
Number of neurons in output layer | 2 |
Learning rate | 0.001 |
Number of epochs | 50 |
Number of iterations (per epoch) | 8 |
Input image size | \(224 \times 224 \times 3\) |
Number of input attributes | 150, 528 |
Number of output attributes | 2 |
Training data size | 294 |
Testing data size | 98 |
Training and testing data size ratio | 75% / 25% |
Activation function (before and after maximum pooling layer) | ReLU |
Batch size | 32 |
Optimizer | Adam |
Results and analysis
Evaluation metrics
Parameter | Predicted value | Actual value |
---|---|---|
True positive | Yes | Yes |
True negative | No | No |
False positive | Yes | No |
False negative | No | Yes |
Results
Training-testing data ratio (%) | Accuracy (%) | Precision (%) | Recall (%) | F1-Score (%) |
---|---|---|---|---|
CNN model-1 (Maximum pool size: \(2 \times 2\); stride: 1) | ||||
\(60-40\) | 98.36 | 83.69 | 99.48 | 90.90 |
\(70-30\) | 97.45 | 99.29 | 80.61 | 89.26 |
\(75-25\) | 98.44 | 98.45 | 97.44 | 97.94 |
\(80-20\) | 97.43 | 97.46 | 97.95 | 97.70 |
CNN model-2 (Maximum pool size: \(2 \times 2\); stride: 2) | ||||
\(60-40\) | 95.56 | 98.30 | 46.93 | 63.88 |
\(70-30\) | 94.91 | 95.52 | 97.95 | 96.72 |
\(75-25\) | 98.37 | 98.48 | 90.30 | 94.90 |
\(80-20\) | 98.71 | 97.95 | 69.89 | 82.28 |
CNN model-3 (Maximum pool size: \(3 \times 3\); stride: 3) | ||||
\(60-40\) | 97.46 | 96.50 | 98.46 | 97.47 |
\(70-30\) | 98.30 | 98.39 | 54.08 | 70.19 |
\(75-25\) | 97.95 | 92.78 | 98.46 | 95.54 |
\(80-20\) | 96.15 | 87.24 | 69.89 | 93.18 |
Related work comparison
Related work | Method | Image type | Images | Accuracy |
---|---|---|---|---|
Butt et al. [39] | 3-D CNN classification | Chest CT scan | Total: 528 \(\mathrm {COVID}^+\): 189 | 86.7% |
Wang et al. [21] | Inception transfer learning | Chest CT scan | Total: 1065 \(\text {COVID}^+\): 325 | 85.2% |
Sethy et al. [32] | Resnet50 and SVM classification | Chest X-ray | Total: 381 \(\mathrm {COVID}^+\): 127 | 95.38% |
Apostolopoulos and Mpesiana [33] | Deep CNN with ResNet and inception | Chest X-ray | Total: 1427 \(\mathrm {COVID}^+\): 224 | 97.82% |
XCOVNet | CNN classification | Chest X-ray | Total: 392 \(\mathrm {COVID}^+\): 196 | 98.44% |