Sarker M M K, Rashwan H A, Akram F, et al. Slsdeep: Skin lesion segmentation based on dilated residual and pyramid pooling networks[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2018: 21-29.
1. Overview
In this paper, it proposed SLSDeep
- encoder. dilated residual layer (DRN)
- decoder. pyramid pooling
- Negative Log Likelihood (NLL) + End Point Error (EPE)
2. Methods
- layer1~layer4. four pretrained DRN
2.1. Loss Function
α=0.5
- NLL
- EPE
u0, u1. first derivatives of u in x and y direction
3. Experiments
3.1. Dataset
- ISBI 2016. 900 train, 379 test; size range [542x718~2848x4288]
- ISBI 2017. 2000 train, 150 valid, 600 test
3.2. Metrics
- Specificity
- Sensitivity
- Jaccard Index
- Dice coefficient
- Accuracy
3.3. Details
- LR. decoder 0.01, encoder 0.001
- poly learning rate policy. 0.9
- batchsize. 16
- epoch. 100
- TITANX 12G, 20 hours
3.3.1. Data Augmentation
- random scale. 0.5~1.5
- random rotation. -10~10
- resize to 384x384 for training
3.4. Comparison
