Phikon
| Property | Value |
|---|---|
| Parameter Count | 86.4M |
| Model Type | Vision Transformer Base |
| Image Size | 224 x 224 x 3 |
| License | Owkin non-commercial license |
| Paper | View Paper |
What is phikon?
Phikon is a sophisticated self-supervised learning model specifically designed for histopathology image analysis. Developed by Owkin, it leverages the iBOT architecture and was trained on an extensive dataset of 40 million pan-cancer tiles from TGCA. This model represents a significant advancement in medical image analysis, particularly in the field of cancer research.
Implementation Details
The model is built on PyTorch 1.13.1 and was trained using NVIDIA V100 GPUs with 32GB RAM on the French Jean Zay cluster. It implements a Vision Transformer (ViT) architecture and processes images at 224x224x3 resolution, making it suitable for detailed histological analysis.
- Trained using self-supervised learning approach
- Utilizes masked image modeling technique
- Optimized for histopathology feature extraction
- Compatible with PyTorch framework
Core Capabilities
- Feature extraction from histology image tiles
- Cancer subtype classification
- Fine-tuning capability for specific cancer types
- Pan-cancer analysis support
Frequently Asked Questions
Q: What makes this model unique?
Phikon stands out due to its specialized training on an extensive histopathology dataset and its self-supervised learning approach, making it particularly effective for medical image analysis without requiring large amounts of labeled data.
Q: What are the recommended use cases?
The model is primarily designed for feature extraction from histology images and can be used for cancer classification across various subtypes. It's particularly valuable in research settings and can be fine-tuned for specific cancer analysis tasks.
