This implementation exhaustively finds biology differences within the image dataset using a self-supervised contrastive learning paradigm. When PCL is trained, the dataset is divided into image patches which are encoded into enriched patch embeddings. Image patch embeddings consists of n-dimensional vectors whose euclidean distances are lowest when referring to similar biology structures and the highest when they contain different biology structures.
Applications: They are especially useful when data is scarce and can be utilized in custom-made machine learning pipelines with very different objectives:
- Discovery of new biological structures
- Measurement of biological structures across subject-types (e.g., treated vs. control)
- Patient outcome prediction
- Subject regression analysis
- etc.
For more information about this tool please refer to this Paper.
Authors: Daniel Jiménez-Sánchez, Mikel Ariz, Hang Chang, Xavier Matias-Guiu, Carlos E. de Andrea, Carlos Ortiz-de-Solórzano.
To replicate the paper's experiments on a endometrial cancer 7-marker image dataset, first download the images following the link (download Example_POLE.zip).
When downloaded, add the images to the folder 'Examples/Example_POLE/'.
Run main.py