HeLP Challenge 2019 Breast Cancer 1st place solution

This repository is 1st place solution to the Breast Cancer Classification Task of HeLP Challenge 2019.

Model

Stage 1

• Preprocessing: ROI extraction, Rescale, Vahadane Stain Normalization
• Pixel-wise Segmentation: Feature Pyramid Network(FPN)

Stage 2

• Feature extraction from probability heatmap
• Prediction final probability and major axis based on features

And also, please click this link to see the detailed model description.

Dependencies

• keras
• segmentation_models
• openslide
• staintools
• numpy
• pandas
• sklearn
• skimage

Usage

Dataset

data
  └── train
     ├── level4
     │  ├── Image
     │  │  ├── slide_001.png
     │  │  ├── ...
     │  │  └── slide_#.png
     │  └── Mask
     │     ├── mask_001.png
     │	   ├── ...
     │	   └── mask_#.png
     └── label.csv
            
========= After training, the directories are created as below. =========

  ├── volume
  │  ├── dataset
  │  │  └── level4 
  │  │     ├── img
  │  │	   │  ├── slide001_patch001.png
  │  │ 	   │  ├── ...
  │  │     │  └── slide#_patch#.png
  │  │	   └── mask
  │  │	      ├── mask001_patch001.png
  │  │        ├── ...
  │  │        └── mask#_patch#.png
  │  └── model
  │       └── fpn_weights.h5
  └── heatmap
      ...

Train

Run the train.py.

$ python train.py

Inference

Run the inference.sh.

$ sh inference.sh

Authors

• Daeyoung Kim / @cyc1am3n
• Taewoo Kim / @Taeu
• Jonghyun Choi / @ExcelsiorCJH