Author Affiliations
School of Information and Engineering, Xiangtan University, Xiangtan, Hunan 411105, Chinashow less
Fig. 1. Structure of residual block
Fig. 2. Flow chart of dermoscopy image classification method
Fig. 3. Examples of seven skin diseases
Fig. 4. Data set classification
Fig. 5. Example of augmented image. (a) Original image; (b) augmented image
Fig. 6. Dividing process of data set
Fig. 7. Examples of secondary data augmentation images. (a) Original image in basic train set; (b) images after secondary data augmentation
Fig. 8. Distribution of training set samples after secondary data augmentation
Fig. 9. Structure of FL-ResNet50 model
Fig. 10. Structure of two kinds of residuals blocks. (a) Identity block; (b) Conv block
Fig. 11. Confusion matrix of classification results
Fig. 12. Loss during training process
Fig. 13. Accuracy during training process
Category | Basic train set | Val set | Test set |
---|
Nv | 5380 | 437 | 888 | Mel | 1044 | 22 | 47 | Bkl | 967 | 43 | 89 | Bcc | 461 | 17 | 36 | Akiec | 282 | 14 | 31 | Vasc | 123 | 6 | 13 | Df | 103 | 3 | 9 | Total | 8360 | 542 | 1113 |
|
Table 1. Distribution of samples on three data sets
Network | F1-micro |
---|
ResNet50 | 0.85 | ResNet50 + augmentation | 0.87 |
|
Table 2. Classification performance comparison before and after data augmentation
Network | F1-micro |
---|
ResNet50 | 0.85 | FL-ResNet50 | 0.87 | FL-ResNet50 + augmentation | 0.88 | VGG19 | 0.83 |
|
Table 3. Comparative analysis on effectiveness of multi-classification Focal Loss function
Category | P | R | F1 |
---|
Akiec | 0.48 | 0.32 | 0.38 | Bcc | 0.74 | 0.69 | 0.71 | Bkl | 0.55 | 0.53 | 0.54 | Df | 1.00 | 0.11 | 0.20 | Mel | 0.55 | 0.55 | 0.55 | Nv | 0.93 | 0.96 | 0.95 | Vasc | 1.00 | 0.85 | 0.92 |
|
Table 4. Classification results of FL-ResNet 50 model using data augmentation method