/MRN

Primary LanguagePython

MRN (Multi-Resolution Network)

This is my personal implementation of MRN. I also summarize MRN paper in my blog.

Structure

Two novel multi-resolution networks are proposed to learn from input patches extracted at multiple levels. These patches share the same centroid and shape (size in pixels), but with an octave based increase of the pixel size, micrometers per pixel (mpp). Only the central high resolution patch is segmented at the output. - Feng Gu et al. (2018)

Environment

pip install -r requirements.txt

Above, I install python 3.6 with CUDA 11.4

Description

Repository Structure

  • model/mrn.py: main MRN model script
  • model/mrn_se_resnext101_32x4d.py: main MRN 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

Training

Data Preparation
MRN_Data
    ├ slide_num_1
    |       ├ input_x1
    |       ├ input_x2
    |       └ input_y1
    .
    .
    .
    └ slide_num_n
            ├ input_x1
            ├ input_x2
            └ input_y1
  • input_x1: mpp=1 image patches(512x512) directory
  • input_x2: mpp=2 image patches(512x512) directory
  • input_y1: mpp=1 mask patches(512x512) directory

You can get this data structure by using util_multi.py

Train Example
python train.py --BASE_PATH './MRN_Data/*/input_y1/*.png' --BACKBONE 'seresnext101' --CLASSES 4 --LOSS_FUNCTION 'diceloss' --DESCRIPTION 'MRN_Test'
Train Option
  • --BASE_PATH: The path of input_y1 mask patches
  • --BACKBONE: The backbond model of MRN model. You can choose vgg16 or seresnext101
  • --BATCH_SIZE: The batch size of training model.
  • --CLASSES: The number of output classes.
  • --MULTIPLE: If you want to setting input_x2 mpp=4 with input_x1 mpp=1, you can add this option 2.
  • --EPOCHS: The epochs batch size of training model.
  • --LOSS_FUNCTION: Choose the loss function either celoss or diceloss.
  • --DESCRIPTION: Add the name of a training model weight.

Reference

paper

code