1 Introduction
2 Related work
3 Proposed technique
3.1 Source domain model selection
3.2 Deep model augmentation
Network | Original | Input | Modified | Input | Activations |
---|---|---|---|---|---|
DenseNet201 | Input_1 | 224\(\times\)224\(\times\)3 | Input_grey | 448\(\times\)448\(\times\)1 | 224\(\times\)224\(\times\)3 |
Conv 7, 3, [2,2] | |||||
Batch-N, ReLU | |||||
ResNet50 | Input_1 | 224\(\times\)224\(\times\)3 | Input_grey | 448\(\times\)448\(\times\)1 | 224\(\times\)224\(\times\)3 |
Conv 7, 3, [2,2] | |||||
Batch-N, ReLU | |||||
Inception-V3 | Input_1 | 299\(\times\)299\(\times\)3 | Input_grey | 598\(\times\)598\(\times\)1 | 299\(\times\)299\(\times\)3 |
Conv 3, 3, [2,2] | |||||
Batch-N, ReLU | |||||
VGG-16 | Input | 224\(\times\)224\(\times\)3 | Input_grey | 448\(\times\)448\(\times\)1 | 224\(\times\)224\(\times\)3 |
Conv 3, 3, [2,2] | |||||
Batch-N, ReLU |
3.3 Deep model training
3.4 Representation augmentation with dictionaries
3.4.1 Representation computation with dictionaries
3.5 Classification
4 Experiments
Models | Spec.\(\%\) | Sens.\(\%\) | F1\(\%\) | Acc.\(\%\) | ERR\(\%\) | \(\dagger\)Acc.\(\%\) | \(\dagger\)Gain\(\%\) |
---|---|---|---|---|---|---|---|
Baseline (TL) | 89 | 80 | – | 83.33 | – | 16.76 | – |
DenseNet201 (Den.) | 94 | 55 | 46 | 89.65 | 37.91 | 48.27 | 188.0 |
Den+VGG | 94 | 55 | 47 | 90.00 | 3.38 | 50.03 | 3.65 |
Den+VGG+Res | 94 | 58 | 51 | 90.63 | 6.29 | 53.17 | 6.27 |
Den+VGG+Res+IV3 | 95 | 60 | 53 | 91.03 | 4.27 | 55.17 | 3.76 |
Proposed | 95 | 60 | 53 | 91.38 | 3.90 | 56.90 | 3.13 |
4.1 ChestX-ray14 data set
4.2 COVID-19 data set
4.3 Evaluation metrics
Models | Training time | Testing time (milliseconds) | ||
---|---|---|---|---|
Step1 | Step2 | Step3 | ||
DenseNet201 | 3 h 20 m 5 s | 6 h 15 m 20 s | 7 h 29 m 5 s | 6 |
VGG | 0 h 55 m 15 s | 1 h 51 m 25 s | 2 h 5 m 20 s | 8 |
ResNet | 0 h 47 m 20 s | 1 h 26 m 15 s | 1 h 35 m 10 s | 6 |
InceptionV3 | 1 h 14 m 25 s | 2 h 23 m 15 s | 2 h 31 m 15 s | 6 |
Class | Spec. | Sens. | F1-Score | Acc. |
---|---|---|---|---|
Atelectasis | 0.96/0.94 | 0.61/0.08 | 0.56/0.09 | 93.46/88.73 |
Cardiomegaly | 0.97/0.75 | 0.86/0.22 | 0.72/0.08 | 96.43/72.27 |
Effusion | 0.97/0.86 | 0.59/0.06 | 0.66/0.05 | 93.36/ 76.60 |
Infiltration | 0.93/0.86 | 0.36/0.09 | 0.42/0.09 | 85.13/74.43 |
Mass | 0.97/0.63 | 0.67/0.19 | 0.55/0.03 | 96.56/62.40 |
Nodule | 0.97/0.96 | 0.41/0.06 | 0.33/0.04 | 96.63/94.83 |
Pneumothorax | 0.94/0.97 | 0.7/0.02 | 0.66/0.04 | 92.10/86.90 |
Consolidation | 0.88/0.99 | 0.73/0.00 | 0.47/– | 87.30/92.2 |
Pleural thickening | 0.93/0.97 | 0.46/0.01 | 0.25/0.01 | 92.46/94.80 |
No finding | 0.95/0.99 | 0.52/0.00 | 0.64/– | 80.63/65.13 |
4.4 Results on chest X-ray14 data set
Model | Spec. \(\%\) | Sens. \(\%\) | F1 \(\%\) | Acc. without Dict.\(\%\) | Acc. with Dict. \(\%\) | \(\dagger\) Acc. without Dict.\(\%\) | \(\dagger\)Acc. with Dict.\(\%\) |
---|---|---|---|---|---|---|---|
Dense | 98.11 | 96.21 | 96.20 | 97.47 | 97.47 | 96.21 | 96.21 |
VGG | 98.86 | 97.73 | 97.74 | 96.97 | 98.48 | 95.45 | 97.73 |
Res | 97.73 | 95.45 | 95.48 | 94.44 | 96.97 | 91.67 | 95.45 |
IV3 | 96.59 | 93.18 | 93.08 | 95.45 | 95.45 | 93.18 | 93.18 |
Ensemble | 99.24 | 98.48 | 98.49 | 98.99 | \(\varvec{99.49}\) | 98.48 | \(\varvec{99.24}\) |
Class | Spec.\(\%\) | Sens.\(\%\) | F1\(\%\) | Acc.\(\%\) |
---|---|---|---|---|
Covid-19 | 100 | 100 | 100 | 100 |
Pneumonia | 98.86 | 100 | 98.88 | 99.24 |
Normal | 100 | 97.73 | 98.88 | 99.24 |
4.5 Results on COVID-19 data set
Study | No. of cases | Architecture | Data set | Accuracy\(\%\) |
---|---|---|---|---|
Ioannis et al. [5] | 224 COVID-19, 700 Pneumonia, 504 Healthy | VGG-19 | 98.75 | |
Wang et al. [61] | 53 COVID(+), 5526 COVID(−), 8066 Healthy | COVID-Net | 92.4 | |
Ozturk et al. [40] | 125 COVID-19, 500 Pneumonia, 500 No finding | Dark COVID-Net | [11] | 98 |
Asif et al. [25] | 1300 images of COVID-19, normal, pneumonia | CoroNet | 95 | |
Tougaccar et al. [55] | 295 COVID-19, 98 Pneumonia, 65 No findings | MobileNetV2 | 99.27 | |
Narin et al. [38] | 50 COVID-19, 50 No findings | ResNet-50 | [11] | 98 |
Hemaden et al. [18] | 25 COVID-19, 25 No findings | VGG-19, DenseNet-121 | [11] | 90 |
Sethy et al. [48] | 25 COVID-19, 25 No findings | ResNet-50 | [11] | 95.38 |
Toraman et al. [56] | 1050 COVID, 1050 No finding | CapsNet | 97.24 | |
Panwar et al. [41] | 192 COVID-19, 145 No findings | nCOVnet | [11] | 97.62 |
Ucar et al. [58] | 76 COVID-19, 1583 normal, 4290 pneumonia | Bayes-SqeezeNet | 98.3 | |
Proposed | 219 COVID-19, 219 Viral Pneumonia, 219 Normal | Augmented | \(\varvec{99.49}\) |