README.md

Stain normalisation for domain generalisation

This project aims to analyze the impact of stain normal- ization on domain generalization in histopathological images. The analysis is carried out in two parts:

1. Qualitative assessment of how stain normalization affects the visual quality of the normalized images.
2. Evaluation of how stain normalization influences the performance of downstream tasks.
3. Analysis of the relationship between image quality post- normalization and the performance of downstream tasks.

Project Structure

• HierarchialFCOS

  • Adopted from paper Deep learning-based Subtyping of Atypical and Normal Mitoses using a Hierarchical Anchor-Free Object Detector
  • Changed from sub category detection to detect single category in different scales

• fcos_eval

  • Cluster scripts to train and evaluation mitotic detection using FCOS model

• multiGAN

  • Adopted from paper MultiPathGAN: Structure Preserving Stain Normalization using Unsupervised Multi-domain Adversarial Network with Perception Loss
  • Changes to train custom datasets in different settings

• multiGANEval

  • Scripts to run multiGAN inference on different datasets
  • Qualitative analysis of stain normalisation results using multiGAN and classical style transfer method

• utils

  • Helper function fro all of the above

Dependencies

 conda create --name stain_norm_eval python=3.12.3
 conda activate stain_norm_eval
 pip install -r requirements.txt

MultiPathGAN

To train multipathGAN on multiple domains in run:

 python -m multiGAN.train --is_cluster --mode 'train'