This is the official pytorch implementation of the MB-DCNN model:

Paper: A Mutual Bootstrapping Model for Automated Skin Lesion Segmentation and Classification. (https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8990108)

Python 3.7
Torch==1.4.0
Torchvision==0.5.0
Apex==0.1
CUDA 10.0

• Clone this repo

git clone https://github.com/YtongXie/MB-DCNN.git
cd MB-DCNN

• Download ISIC2017 dataset, Extra 1320 images(The ID information of the extra 1320 images is listed in ./dataset/Extra_images.csv)
  

• Put the data under ./dataset/data/

• Run python ./dataset/extractPatch_cls_train.py and python ./dataset/extractPatch_cls_val_test.py to obtain the cropped training, validation and testing patches for classification task.

• Run python ./dataset/extractPatch_seg_train.py to obtain the resized training patches for segmentation task.

• Run python ./dataset/list_cls.py and python ./dataset/list_seg.py to generate the data lists.

• Download pretrained weights from Deeplabv3+, Xception and put them under ./models/ .

• Run python train_CoarseSN.py to train the coarse segmentation network for roughly segmenting skin lesions.

• The segmentation network structure is defined in ./net/models.py, and the hybrid loss is defined in ./net/loss.py.

• Run python generate_Coarse_mask.py to obtain the coarse masks for mask-CN and enhanced-SN.

• Run python train_MaskCN.py to train the mask-guided classification network for skin lesion classification.

• The classification network structure is defined in ./net/models.py, and the loss is cross-entropy loss.

• Run python train_EnhancedSN.py to train the enhanced segmentation network for more accurate skin lesion segmentation.

• Run python eval_MaskCN.py and python eval_EnhancedSN.py to start the evaluation.

If this code is helpful for your study, please cite:

@ARTICLE{8990108,
  author={Yutong Xie and Jianpeng Zhang and Yong Xia and Chunhua Shen},
  journal={IEEE Transactions on Medical Imaging}, 
  title={A Mutual Bootstrapping Model for Automated Skin Lesion Segmentation and Classification}, 
  year={2020},
  volume={39},
  number={7},
  pages={2482-2493},}

The codes for Deeplabv3+ network and Xception network are reused from the YudeWang and Cadene.
Thanks to YudeWang and Cadene for the pretrained weights for Deeplabv3+ network and Xception network.

Yutong Xie (yutong.xie678@gmail.com)