This is my personal implementation(with Torch) of HookNet. I also summarize HookNet paper in my blog.

We propose HookNet, a semantic segmentation model for histopathology whole-slide images, which combines context and details via multiple branches of encoder-decoder convolutional neural networks. Concentric patches at multiple resolutions with different fields of view are used to feed different branches of HookNet, and intermediate representations are combined via a hooking mechanism. - Mart van Rijthoven. (2020)

Also, I tried to make Quad-scale HookNet, which is a new model that transformed HookNet.

This model uses the same hooking mechanism as hooknet, but performs concat from 3 context structures with different mpps to 1 target structure.

pip install -r requirements.txt

Above, I install python 3.6 with CUDA 11.4

• model/hooknet.py: main HookNet and Quad-scale HookNet model script
• model/hooknet_se_resnext101_32x4d.py: main HookNet_SE-ResNeXt101_32x4d model script
• datagen.py: the data dataloader and augmentation script
• functional.py: naming a weight of model and converting outputs to images script
• test_view.py: visualizing outputs script
• train.py: main training script

HookNet_Data
    ├ slide_num_1
    |       ├ input_x1
    |       ├ input_x2
    |       ├ input_x4
    |       ├ input_x8
    |       └ input_y1
    .
    .
    .
    └ slide_num_n
            ├ input_x1
            ├ input_x2
            ├ input_x4
            ├ input_x8
            └ input_y1    

• input_x1: mpp=1 image patches(284x284x3) directory
• input_x2: mpp=2 image patches(284x284x3) directory
• input_x4: mpp=4 image patches(284x284x3) directory
• input_x8: mpp=8 image patches(284x284x3) directory
• input_y1: mpp=1 mask patches(284x284) directory (The shape of the mask patches will be 70x70 by cropping.)

python train.py --BASE_PATH './HookNet_Data/*/input_y1/*.png' --INPUT_SHAPE 284 --CLASSES 3  --TARGET_INDICE 4 --CONTEXT_INDICE 2 --MODEL 'hooknet'--ENCODER 'normal' --LOSS_FUNCTION 'dice_loss' --DESCRIPTION 'HookNet_test'

• --BASE_PATH: The path of input_y1 mask patches
• --BATCH_SIZE: The batch size of training model.
• --INPUT_SHAPE: The input shape of the patch.
• --CLASSES: The number of output classes.
• --TARGET_INDICE: The number of target indice of Hooknet. It should be 5 > target > context > 1.
• --CONTEXT_INDICE: The number of context indice of HookNet. It should be 5 > target > context > 1.
• --EPOCHS: The epochs batch size of training model.
• --MODEL: Choose the model either hooknet or quad_scale_hooknet
• --ENCODER: Choose the encoder model either normal or se_resnext101_32x4d
• --LOSS_FUNCTION: Choose the loss function either celoss or diceloss
• --DESCRIPTION: Add the name of a training model weight.

• Model Structure
• Encoder model structure

• Public implementation(with Tensorflow)
• SE-ResNext101_32x4d
• Training