MetaHistoSeg: A Python Framework for Meta Learning in Histopathology Image Segmentation

A Python framework that implements unique scenarios in both meta learning and instance based transfer learning in Histopathology Image Segmentation with a benchmark dataset for training and validating models on out-of-distribution performance across a range of cancer types.

Paper Link

http://arxiv.org/abs/2109.14754

Getting started

conda create -n <your_virtual_env> python=3.7

Create a virtual environment, activate it and install the requirements in requirements.txt.

conda activate <your_virtual_env>
pip install -r requirements.txt

Requirements

• CUDA 9.0
• Python 3.7
• PyTorch 1.4.0
• Torchvision 0.5.0
• SimpleITK 2.1.1
• opencv-python 4.5.3.56
• bs4 0.0.1
• pathlib 1.0.1
• pytest-shutil 1.7.0
• scikit-image 0.18.3
• tensorboardX 2.4

Datasets

Download raw datasets by following the instructions on each host website.

Host	Datasets	Size	Anatomical sites	Num of Segments	Local folder
https://digestpath2019.grand-challenge.org/	digestPath	250	colon	2	grand-challenge/digestPath/tissue-train-pos-v1
https://gleason2019.grand-challenge.org/Register/	gleason 2019	244	prostate	6	grand-challenge/gleason2019
https://breastpathq.grand-challenge.org/Data/	BreastPathQ	154	breast	4	grand-challenge/BreastPathQ/breastpathq/datasets
https://monuseg.grand-challenge.org/Data/	MoNuSeg	30	Breast, kidney, Liver, Prostate, bladder, colon, stomach	2	grand-challenge/MoNuSeg Training Data
https://warwick.ac.uk/fac/sci/dcs/research/tia/glascontest/download	glandsegmentation	161	colon	2	grand-challenge/Warwick QU Dataset (Released 2016_07_08)

Usage

For meta learning scenario with episodic sampling, use train_episode.py. e.g. when setting gleason2019, BreastPathQ, and GlandSegmentation as the training tasks with equal probability and using digestpath as the validation task,

python train_episode.py /path/to/data --dataset gleason2019 BreastPathQ GlandSegmentation --dataset-prob 1/3 1/3 1/3 --validation-dataset digestpath --num-adaptations 1 --step-size 0.01 --meta-lr 0.0001 --max-iterations 200000 --batch-size 2 --batch-size-test 32 --meta-test-lr 0.001 --test-every 50 --use-cuda --num-workers 4 --device_ids "0,1,2,3" --output-folder /path/to/results

For transfer learning scenario with instance sampling, use train_batch.py. e.g. the same setup as above,

python train_batch.py /path/to/data --dataset gleason2019 BreastPathQ GlandSegmentation --dataset-prob 1/3 1/3 1/3 --validation-dataset digestpath --lr 0.0001 --max-iterations 200000 --batch-size 4 --batch-size-test 32 --meta-test-lr 0.001 --test-every 50 --use-cuda --num-workers 4 --device_ids "0,1,2,3" --output-folder /path/to/results

During inference, use refine.py. e.g. when new task is MoNuSeg,

python refine.py /path/to/data /export/medical_ai/ --dataset MoNuSeg --model-path /path/to/model --batch-size 32 --lr 0.001 --num-workers 4 --epochs 50 --device_ids "0,1,2,3" --output-folder /path/to/results --prediction-folder /path/to/visualization

Results

Meta-dataset task examples: Top to bottom: GlandSegmentation, MoNuSeg, digestpath and BreastPathQ. Left to right: original image, ground truth, segmentation by MAML and segmentation by instance based transfer learning.

References

This code is based on https://github.com/tristandeleu/pytorch-maml and https://github.com/tristandeleu/pytorch-meta

If you want to cite this paper

@article{MetaHistoSeg,
  author    = {Zheng Yuan and Andre Esteva and Ran Xu},
  title     = {MetaHistoSeg: A Python Framework for Meta Learning in Histopathology Image Segmentation},
  journal   = {DGM4MICCAI/DALI, LNCS 13003},
  year      = {2021}
}