2D Skin layers and Keratinocytes Segmentation

This project descibes the codes for 2D skin layers (air gap, SC, epidermis, dermis) and keratinocytes segmentation model skin images.

The segementation model is based on the U-Net architecture with applying deep supvision and our proposed deep feature sharing method.

Dataset Format

The file name of each image and its corresponding layer and cell labeling should be the same.

./[your own path]/dataset
----/image   % Image files [png]
--------/1_20180129094629_cheek.png
--------/1_20180209120922_normal.png
--------/1_20180326_111332_foot.png
----/layer   % skin layers labeling [png]
--------/1_20180129094629_cheek.png
--------/1_20180209120922_normal.png
--------/1_20180326_111332_foot.png
----/cell    % cell nuclei labeling [png]
--------/1_20180129094629_cheek.png
--------/1_20180209120922_normal.png
--------/1_20180326_111332_foot.png

Available Model

• U-Net
• U-Net with Deep Supervision
• U-Net with Deep Feature Sharing
• U-Net with Deep Supervision and Deep Feature Sharing

Installation

• Clone this repo:

git clone https://github.com/tomohiroliu22/skin-segmentation-with-DS-and-DFS
cd skin-segmentation-with-DS-and-DFS

Model Comparison and Evaluation

• Training all types of the models with 5 folds cross-validation by running

bash train.sh

• Testing all types of the models with 5 folds cross-validation by running

bash test.sh

Model Training

Training for U-Net with DS and DFS

python main.py --dataroot ./[your own path]/datasets --name [your experiment name] --model DFS_w_DS --phase train --lr 0.001 --step 10 --epoch 25

Training for U-Net with DFS

python main.py --dataroot ./[your own path]/datasets --name [your experiment name] --model DFS --phase train --lr 0.001 --step 10 --epoch 25

Training for U-Net with DS

python main.py --dataroot ./[your own path]/datasets --name [your experiment name] --model DS --phase train --lr 0.001 --step 10 --epoch 25

Training for U-Net

python main.py --dataroot ./[your own path]/datasets --name [your experiment name] --model NONE --phase train --lr 0.001 --step 10 --epoch 25

Model Testing

Testing for U-Net with DS and DFS

python main.py --dataroot ./[your own path]/datasets --name [your experiment name] --model DFS_w_DS --phase test --modelpath ./[path to your model]

Testing for U-Net with DFS

python main.py --dataroot ./[your own path]/datasets --name [your experiment name] --model DFS --phase test --modelpath ./[path to your model]

Testing for U-Net with DS

python main.py --dataroot ./[your own path]/datasets --name [your experiment name] --model DDS --phase test --modelpath ./[path to your model]

Testing for U-Net

python main.py --dataroot ./[your own path]/datasets --name [your experiment name] --model NONE --phase test --modelpath ./[path to your model]

Model Performance

The mean DICE coefficient comparison for 4 types of models

Model	U-Net	U-Net+DS	U-Net+DFS	U-Net+DS+DFS
Cell	0.7125	0.7138	0.7185	0.7199
Layer	0.8883	0.8904	0.8915	0.8928