This is our ongoing PyTorch implementation for Semantic Segmentation with Squeeze-and-Excitation Block: Application to Infarct Segmentation in DWI, Medical Imaging Meets NIPS workshop, NIPS 2017

Results of our proposed method and baseline networks. (a) b1000 image (b)ADC map (c) DenseNet (d) SE-DenseNet (e) UNet (f) SE-UNet (g) Ground truth.

We replaced fully connected layer with 1x1 convolution layer in original excitation operation and observed that 16 reduction ratio of excitation is better.

You can refer to Squeeze-and-Excitation Networks

• Python 2.7
• PyTorch >= 0.2.0
• Visdom
• NumPy

You have to construct suitable dataloader for your dataset. For example, we generate our dataloader for below data structure.

data |---ADC

    |---train
    
    |---val
    
    |---test
    
|---b1000

    |---train
    
    |---val
    
    |---test
    
|---mask

    |---train
    
    |---val
    
    |---test

We only open a part of our test data because of patient information.

U-Net: Convolutional Networks for Biomedical Image Segmentation can be divided into three blocks, e.g. encoding, bottleneck, and decoding blocks

Instead, SE-Unet is built on SE-encoding, SE-bottleneck, and SE-decoding blocks that stack SE block behind each block, respectively. If you want to use SE-Unet, you have to set reduction ratio.

• SE-Unet

python train.py --root_dir ./data --save_dir ./weight --model se_unet --reduction_ratio 16 --network_depth 4 --bottleneck 5 --init_features 32 --gpu_ids 0,1 --batch_size 20

• Unet

python train.py --root_dir ./data --save_dir ./weight --model unet --network_depth 4 --bottleneck 5 --init_features 32 --batch_size 20 --gpu_ids 0,1

Denseblock is key component of Densenet and is dependent on growth rate. Details of denseblock are illustrated by The One Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic Segmentation

SE-Densenet is constructed by stacking SE-denseblocks and they are generated by adding SE block to original denseblock. If you want to use SE-Densenet, you have to set reduction ratio.

• SE-Densenet

python train.py --root_dir ./data --save_dir ./weight --model se_densenet --down_blocks 4,5,7,10,12 --up_blocks 12,10,7,5,4 --growth_rate 16 --reduction_ratio 16 --batch_size 2 --bottleneck 15 --init_features 48 --gpu_ids 0,1

• Densenet

python train.py --root_dir ./data --save_dir ./weight --model densenet --down_blocks 4,5,7,10,12 --up_blocks 12,10,7,5,4 --growth_rate 16 --batch_size 2 --bottleneck 15 --init_features 48 --gpu_ids 0,1

We upload pretrained weight of SE-Unet. You have to generate below model architecture and use two gpus in order to load a pretrained weight.

python test.py --root_dir ./data --weight_dir ./weight/se_unet/se_unet.pth --save_dir ./results --model se_unet --reduction_ratio 16 --network_depth 4 --bottleneck 5 --init_features 32 --gpu_ids 0,1

If you want to continue training from a previously saved model, you set argument --weight_dir.

You can obtain information of argument by running help.

python train.py -h

• To view training results and loss plots, you should run visdom before training

python -m visdom.server

click the URL http://localhost:8097

Some codes are from

• https://github.com/bodokaiser/piwise
• https://github.com/ycszen/pytorch-seg
• https://github.com/meetshah1995/pytorch-semseg
• https://github.com/moskomule/senet.pytorch

This work was supported by Kakao Corp. and Kakao Brain Corp.