README.md

metadata

license: mit

Snuffy

Snuffy is a state-of-the-art framework for whole-slide image (WSI) classification, introduced in the paper Snuffy: Efficient Whole Slide Image Classifier by Hossein Jafarinia et al. from Sharif University of Technology. Tested on the TCGA Lung Cancer and CAMELYON16 datasets, it consists of two main components:

1. Self-Supervised Continual Pre-Training with PEFT: Uses Parameter Efficient Fine Tuning (PEFT) with AdaptFormer for effective training in the pathology domain.
2. Snuffy MIL-Pooling Architecture: A novel pooling architecture designed for sparse transformers, tailored to the complexity of cancer biology and tissue microenvironments.

Snuffy addresses the challenge of balancing computational power and performance in WSI classification, offering two versions:

• Efficient Snuffy: Pre-trained on natural images and fine-tuned with PEFT on WSIs.
• Exhaustive Snuffy: Trained entirely from scratch on WSIs.

Both versions use the Snuffy MIL-pooling architecture.

Usage

The code and documentation for Snuffy is available at: https://github.com/jafarinia/snuffy
This repository includes weights for the embedder, embeddings, and aggregator models as described in the paper.

Available models include:

1. Snuffy SimCLR from scratch (Aggregator provided here)
2. Snuffy DINO from scratch (All components provided here)
3. Snuffy DINO with Adapter (All components provided here)
4. Snuffy MAE with Adapter (All components provided here)

BibTeX entry and citation info

Paper page

@misc{jafarinia2024snuffyefficientslideimage,
      title={Snuffy: Efficient Whole Slide Image Classifier}, 
      author={Hossein Jafarinia and Alireza Alipanah and Danial Hamdi and Saeed Razavi and Nahal Mirzaie and Mohammad Hossein Rohban},
      year={2024},
      eprint={2408.08258},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2408.08258}, 
}