In this repository, we read a dataset of X-ray images, which includes classes: Covid, Normal, Lung Opacity and Viral Pneumonia. Then we classify the data using different networks such as Cnn, EffiecientNet_V2_S, Swin Transformer and CnnTransformer. EffiecientNet_V2_S and Swin Transformer networks are trained by fine tune or transfer learning method and Cnn and CnnTransformer networks are trained with initial weights. The EfficientTransformer network consists of a part of the EffiecientNet_V2_S network and the Encoder layer of the Transformer network, and this network is trained only on the COVID-19 Radiography Database.
The dataset that has been trained and evaluated with that model is the dataset of x-ray images. This dataset contains 25,103 images and includes four classes: Covid, Normal, Lung Opacity and Viral Pneumonia.
21,165 images are related to the COVID-19 Radiography Database, and 317 images are related to the Covid-19 Image dataset, and 3,621 images are also related to the Curated Chest X-Ray Image Dataset for COVID-19.
The reason why images from different datasets were used is to improve the data distribution of each class to avoid biasing the model or network on one of the classes.
| Train | Validation | Test | |
|---|---|---|---|
| CnnTransformer | 0.95 | 0.93 | 0.92 |
| Cnn | 0.92 | 0.91 | 0.89 |
| EffiecientNet_V2_S | 0.99 | 0.96 | 0.95 |
| Swin Transformer | 0.97 | 0.95 | 0.95 |
| EfficientTransformer | 0.98 | 0.96 | 0.96 |
| Train | Validation | Test | |
|---|---|---|---|
| CnnTransformer | 0.14 | 0.19 | 0.23 |
| Cnn | 0.24 | 0.25 | 0.3 |
| EffiecientNet_V2_S | 0.05 | 0.14 | 0.17 |
| Swin Transformer | 0.9 | 0.15 | 0.17 |
| EfficientTransformer | 0.04 | 0.11 | 0.11 |
| CnnTransformer | Cnn | EffiecientNet | Swin Transformer | EfficientTransformer | |
|---|---|---|---|---|---|
| Params | 11,276,804 | 1,554,948 | 20,182,612 | 27,522,430 | 6,116,648 |