Fig. 1. Residual block
Fig. 2. Architecture of SK-ResNet
Fig. 3. Structure of discriminator
Fig. 4. Process of unsupervised pre-training
Fig. 5. Example of labeled areas of lung (areas 1, 2 denote pathology area)
Fig. 6. Image patch examples. (a) Source domain; (b) target domain
Fig. 7. Trend graph of classification results for training sets with different scales
Fig. 8. Misclassified examples
Scheme | Size of training set | Size of test set | Split manner |
---|
A | 4132 | 1031 | 5-fold | B | 2560 | 2479 | Case select | C | 11289 | 750 | Random select |
|
Table 1. Different dataset separation schemes
Network | Structure ofResNet block | favg |
---|
A | B | C |
---|
SK-ResNet 10 | [1,1,1,1] | 0.9586 | 0.9453 | 0.9651 | SK-ResNet 14 | [2,2,1,1] | 0.9609 | 0.9476 | 0.9640 | SK-ResNet 18 | [2,2,2,2] | 0.9533 | 0.9354 | 0.9553 | SK-ResNet 34 | [3,4,6,3] | 0.9497 | 0.9329 | 0.9549 |
|
Table 2. Comparison of classification results of SK-ResNet with different depths
Method | A | B | C |
---|
Without transfer | 0.9609 | 0.9476 | 0.9640 | DIM | 0.9799 | 0.9654 | 0.9707 | DIM+PM | 0.9818 | 0.9677 | 0.9756 |
|
Table 3. Comparison of results on SK-ResNet before and after using transfer learning
Ground truth | Predicted result |
---|
EM | FB | GG | NM | MN |
---|
EM | 0.91 | 0 | 0 | 0.08 | 0.01 | FB | 0 | 0.96 | 0.04 | 0 | 0 | GG | 0 | 0.02 | 0.98 | 0 | 0 | NM | 0 | 0 | 0.03 | 0.94 | 0.03 | MN | 0 | 0 | 0 | 0.06 | 0.94 |
|
Table 4. Classification confusion matrix of SK-ResNet
Ground truth | Predicted result |
---|
EM | FB | GG | NM | MN |
---|
EM | 0.97 | 0 | 0 | 0.03 | 0 | FB | 0 | 0.96 | 0.04 | 0 | 0 | GG | 0 | 0.02 | 0.98 | 0 | 0 | NM | 0 | 0 | 0.03 | 0.96 | 0.01 | MN | 0 | 0 | 0 | 0.03 | 0.97 |
|
Table 5. Classification confusion matrix of SK-ResNet using transfer learning
Method | A | B | C |
---|
Ref.[14] | 0.7725 | 0.7644 | 0.7987 | Ref. [16] | 0.9465 | 0.9215 | 0.9392 | Ref. [23] | 0.9483 | 0.9269 | 0.9554 | AlexNet[11] | 0.8962 | 0.8821 | 0.9226 | Pre-trained AlexNet[11] | 0.9471 | 0.9337 | 0.9609 | Pre-trained VGG-16[24] | 0.9603 | 0.9435 | 0.9664 | ResNet-18[17] | 0.9371 | 0.9306 | 0.9432 | Pre-trained ResNet-18 [17] | 0.9581 | 0.9443 | 0.9651 | Pre-trained DenseNet-121[25] | 0.9741 | 0.9579 | 0.9715 | SK-ResNet | 0.9609 | 0.9476 | 0.9640 | SK-ResNet (transfer) | 0.9818 | 0.9677 | 0.9756 |
|
Table 6. Comparison of classification performances of SK-ResNet and other methods