Skip to content

Latest commit

 

History

History
133 lines (88 loc) · 6.57 KB

README.md

File metadata and controls

133 lines (88 loc) · 6.57 KB

Segment Really Anything - Even Your Whole Slide Images on a Laptop

Pathology Finetuning Framework Banner

Pathology Slide Finetuning Framework

🧬 Laptop-friendly finetuning of segmentation models for arbitrary structures on WSIs

👋 Introduction

So—you want to segment some arbitrary structure on pathology slides.
Sounds simple, right? After all, AI is everywhere! 🤖✨

But once you start searching for tools, reality hits fast:

  • 💸 Many solutions require expensive commercial platforms
  • 🧠 Others feel like you need a PhD in computer vision just to run them
  • 🧪 Most pretrained models were trained on different stainings, making them useless for your specific task

If you’ve ever thought:

“I’m too lazy or simply not capable of getting these existing methods to work…”

…then this project is exactly for you.
You’re not alone—and it really shouldn’t be this hard.

🎯 What This Project Gives You

This framework provides a simple, practical, laptop-friendly pipeline that lets you finetune a segmentation model for your structures on your slides:

🔧 The Workflow

  1. 📝 Annotate a few examples of the target structure
    Use QuPath to draw annotations on a small set of slides.

  2. 🧠 Finetune a segmentation model
    Train a model specifically on your annotations and your staining.

  3. 🗂 Generate & export segmentation masks
    Run inference on the rest of your slides and export masks in a QuPath-readable format.

That’s it. No GPU clusters, no server processes, no 200-line configs you don’t understand.

📌 Why Not Just Use Existing Tools?

Most existing solutions for finetuning pathology models are either too heavy, too outdated, or don’t play nicely with real-world pathology workflows.

Limitations of Existing Solutions

Method / Ecosystem Strengths Major Limitations (Motivation for This Project)
MONAI Label Server Good integration with MONAI pipelines; semi-automatic labeling Requires dedicated GPU server, difficult to run on laptops; complex setup; produces cryptic runtime errors; not optimized for WSI-scale finetuning
Napari Pathology Plugins Plugin-based architecture; simple UI Many plugins are outdated, unstable, or unmaintained; poor handling of large WSIs; limited or no finetuning support; segmentation tools often unreliable
MedSAM / SAM Finetuning Strong pretrained segmentation capabilities Finetuning typically produces only binary/semantic segmentation masks; output masks cannot be viewed or edited directly in QuPath; workflows are often non-transparent and hard to adapt
Generic PyTorch/TensorFlow Scripts Full flexibility; research-friendly Not optimized for WSIs; require extensive coding; missing built-in patch extraction, slide metadata handling, QuPath mask support, or pathology-specific evaluation tools
Commercial Annotation Tools Polished UI, enterprise support Limited customizability; expensive; finetuning often not transparent; export formats frequently incompatible with open-source pathology tools

✨ Project Goals

  • ✅ Accessible – runs on a laptop (CPU, optional GPU)
  • ✅ Transparent – no black-box servers or mysterious Docker stacks
  • ✅ Pathology-first – built for WSIs, not toy image datasets
  • ✅ QuPath-friendly – annotations in, masks out, no hacks
  • ✅ Extensible – plug in your own models and backbones

🚀 Quick Start (Sketch)

# 1. Clone the repo
conda create -n medsam_patho python=3.10
git clone <YOUR-REPO-URL>.git
cd <YOUR-REPO-NAME>
pip install -e .


# 3. Run a minimal finetuning example
python train.py --config configs/finetune_example.yaml

✏️ Provide GT segmentations

  1. Annotate as much slides as you want (of course the amount of annotated slides impacts the performance)

  2. Load your images to QuPath

Create sample annotations

  1. For fast annotations, we recommedn using the magic brush from the tools bar in QuPath

  2. Select > File > Export objects as geojson > Export All objects

  3. Name your slide and geojson slide_0.svs + slide_0.geojson. THE NAMING MUST BEGIN WITH 0!

  4. Convert your geojson into numpy

python3 ConvertGeoJsonToNumpy.py --wsi /Path/slide_4.svs --geojson /Path/slide_4.geojson --out /Path/slide_4_mask.npy --level 0

  1. Repeat this for as many slides as you want, continue naming them slide_2.svs + slide_2.geojson

  2. Download the initial MedSam checkpoint from https://drive.google.com/file/d/1UAmWL88roYR7wKlnApw5Bcuzf2iQgk6_/view?usp=drive_link and make sure you give the Lab from MedSam proper credit when you use it https://github.com/bowang-lab/MedSAM

  3. Start finetuning of the segmentation model on your annotated tiles

python3 finetuning.py --task <describe_here_your_task (e.g. GlomeruliSegmentationOnMassonsTrichrome)> --MEDSAM_CHECKPOINT_PATH <define here the path to the checkpoint that you have previously downloaded> --WSI_DATA_DIR <where-your-data-is-located> --TILE_SIZE <WSIs need to be patched into smaller images, define here the size (depends on your computer RAM)> --MAGNIFICATION_LEVEL <usually 20 or 40> --EPOCHS <10 is recommended> --BATCH_SIZE <set to 1 or two for your laptop> --ANNOTATED_SLIDES <How many slides have you sampled> --LEARNING_RATE <1e-5> --DEVICE <if your computer has a GPU then set to cuda, if you have a MAC with M2 processor or later then mps, for all others set simply to cpu>

  1. After training is finished, adjust in the Inference script the path to the checkpoint after your finetuning and the path to the slides and then run python3 Inference.py

  2. Visualize your AI generated segmentations in QuPath:

Create sample annotations

License

This project is made avialble under the CreativeCommons License. See the license file for more details.

Acknowledgement

Please cite the release of this repo via zenodo https://doi.org/10.5281/zenodo.17720737 if you find it useful.

Copyright

Segment Really Anything - Even Your Whole Slide Images on a Laptop is developed and maintained by the Division of Medical Image Computing at the German Cancer Research Center (DKFZ). Create sample annotations