Vision Mamba for ADPv2: Hierarchical Histological Tissue Type Classification

This project provides an implementation of the Vision Mamba model for training and evaluation on the Atlas of Digital Pathology v2 (ADPv2) dataset, as described in the paper:

ADPv2: A Hierarchical Histological Tissue Type-Annotated Dataset for potential Biomarker Discovery of Colorectal Disease ([paper link])

The ADPv2 dataset comprises 20,000 image patches from healthy colon tissue, each annotated according to a highly detailed hierarchical tissue taxonomy spanning 32 distinct tissue types across three levels of tissue specificity. For a comprehensive description of the dataset, refer to Section 3 ("Dataset") of the paper.

---

Structure

• Pretraining Script: Self-supervised pretraining of Vision Mamba using Barlow Twins on unlabelled histology patches.
  This pretraining phase is optional but recommended to enhance downstream classification performance.

• Finetuning Script: Supervised training on labelled ADPv2 tissue types for multi-label classification.
  This step uses the labelled patches as shown in the taxonomy below.

• SLURM Scripts: Sample job scripts for distributed GPU training on clusters like Compute Canada.

• Taxonomy Reference: See the table below for the hierarchical taxonomy of tissue types used in this project.
  Green labels indicate the tissue types included in the current multi-label model; red labels are defined in the ontology but do not appear in the present dataset.

---

Hierarchical Taxonomy of Histological Tissue Types

Table 1: Hierarchical taxonomy of histological tissue types in ADPv2.

• Red = tissue types without any presence in the current dataset.
• Green = tissue types included in multi-label representation model training.

For more information on each tissue class, refer to the paper and the image above.

---

🚀 Getting Started

1. Clone & Environment

git clone https://github.com/AtlasAnalyticsLab/ADPv2.git
cd ADPv2

# Create and activate environment
python -m venv VmambaEnv
source VmambaEnv/bin/activate

# Install dependencies
pip install -r requirements.txt

---

📂 Folder Structure

Download annotated dataset at the folowing links: Part 1: https://zenodo.org/records/15307021 Part 2: https://zenodo.org/records/15312384 Part 3: https://zenodo.org/records/15312792 Simply place the downloaded dataset images under the same training folder. Likewise place corresponding ground truth annotation csv files into a single unified csv file. For specific training setup, refer to the bash scripts found in classifications/training_scripts/. configure the filepaths for the training images and ground truth files to your specific paths.

📄 Citation If you use this code, please cite:

[citation]

🤝 Contributing Fork the repo

Create a feature branch (git checkout -b feat/awesome)

Commit your changes (git commit -m "Add awesome feature")

Open a Pull Request