by Xiaoqing Guo, Zhen Chen, Yixuan Yuan.
This repository is for our ISBI2020 paper "Complementary Network with Adaptive Receptive Fields for Melanoma Segmentation", which aims to solve the hole (Fig. 1 (a)) and shrink (Fig. 1 (b)) problem in predictions. The relatively low contrast between melanoma and non-melanoma regions confuses the network and causes the appearance of holes. The fuzzy boundaries lead to the shrinking prediction and further decrease the sensitivity of prediction.
Fig. 1: Illustrations of (a) hole problem, (b) shrink problem. Each group includes the original image, ground truth and prediction of U-Net from left to right.
Tensorflow 1.4 Python 3.5
Clone the repository:
git clone https://github.com/Guo-Xiaoqing/Skin-Seg.git
cd Skin-Seg
sh ./script/train_dml_mobilenet_on_market.sh
sh ./script/evaluate_dml_mobilenet_on_market.sh
Each row includes the original image, dilated rate map, predictions and ground truth from left to right. Note that red in heat map denotes a larger receptive field.
Examples of complementary network results in comparison with other methods. The ground truth is denoted in black. Results of \cite{ronneberger2015u}, \cite{sarker2018slsdeep}, \cite{yuan2017improving} and ours are denoted in blue, cyan, green, and red, respectively.
To be updated
Please contact "xiaoqingguo1128@gmail.com"