Author Affiliations
1School of Electrical Engineering, Xinjiang University, Urumqi, Xinjiang 830047, China2Network and Information Technology Center, Xinjiang University, Urumqi, Xinjiang 830046, Chinashow less
Fig. 1. Network structure of our algorithm
Fig. 2. ResNet50_32×4d residual structure
Fig. 3. Efficient dual-channel attention mechanism module
Fig. 4. Atrous spatial pyramid pooling module
Fig. 5. Feature decoding block
Fig. 6. ISIC 2017 dermoscopy image dataset
Fig. 7. Data preprocessing
Fig. 8. Test results of each algorithm segmentation index on the ISIC 2017 dataset. (a) Accuracy curve of verification set; (b) loss curve of verification set; (c) Dice_Coefficient curve of test set; (d) Jaccard_Index curve of test set
Fig. 9. Test results of the speed and stability of each algorithm on the ISIC 2017 dataset. (a) Speed test results; (b) stability test results
Fig. 10. Comparison of the segmentation results of each algorithm on the ISIC 2017 dataset and the real label, in which the smooth curves represent the real labels, and the zigzag curves represent the segmentation results
Method | Val_Accuracy /% | Dice_Coefficient /% | Jaccard_Index /% | Specificity /% | Error /% |
---|
Our | 95.00±0.04 | 88.74±0.06 | 81.55±0.04 | 89.54±0.30 | 4.99±0.04 | DeepLab V3 Plus | 93.83±0.07 | 85.59±0.03 | 78.16±0.07 | 85.63±0.70 | 6.16±0.07 | DeepLab V3 | 93.10±0.30 | 85.52±0.15 | 76.90±0.10 | 86.38±0.80 | 6.80±0.30 | CE-Net | 92.47±0.05 | 83.54±0.20 | 74.76±0.14 | 86.31±0.60 | 7.52±0.05 | U-Net | 91.52±0.04 | 78.81±0.10 | 69.32±0.07 | 73.99±0.80 | 8.48±0.04 |
|
Table 1. Test results of each algorithm segmentation index on the ISIC 2017 dataset
Method | Val_Accuracy /% | Dice_Coefficient /% | Jaccard_Index /% | Specificity /% |
---|
Our | 95.00 | 88.74 | 81.55 | 89.54 | Goyal, et al[21] | — | 87.14 | 79.34 | — | Tang, et al[22] | 93.58 | 85.83 | 77.75 | — | Singh, et al[23] | 94.95 | 87.90 | 76.65 | 97.05 |
|
Table 2. Comparison with other advanced methods on the ISIC 2017 dataset
Algorithm | Val_Accuracy /% | Dice_Coefficient /% | Jaccard_Index /% | Specificity /% | Error /% |
---|
With ASPP & EDAM | 95.00±0.04 | 88.74±0.06 | 81.55±0.04 | 89.54±0.30 | 4.99±0.04 | With EDAM | 92.47±0.20 | 85.20±0.04 | 76.92±0.20 | 86.31±0.10 | 7.52±0.20 | With ASPP | 94.32±0.10 | 86.52±0.08 | 78.50±0.10 | 87.91±0.80 | 7.50±0.10 | Without ASPP and EDAM | 90.43±0.18 | 82.08±0.06 | 73.38±0.08 | 88.54±0.60 | 9.56±0.18 |
|
Table 3. Test results of our algorithm and its ablation module on ISIC 2017 dataset