This repo contains code for the Capstone Project of the MS Heath Data Science at Harvard University. We developed an interactive viewer to help pathologists view images through embeddings generated from deep learning models.

Camelyon 2016 dataset. This can be applied to any pathology images.

1. Divide images into tiles
2. Generate embeddings of tiles using a model pre-trained on pathology images
3. Conduct K-means to classify tiles based on their embeddings
4. Run the xml parser to parse the cancerous polygons annotated on the image
5. Integrate the above info into the Napari viewer

Run pipeline by running python main.py

--image_path [path of the input wsi]

--model_path [path of the model to generate embedding]

--prefix [path to store the media data]

--tile_size [tile size]

--level [perform embedding at which level of image. the higher the level, the quicker the preprocess]

--xml_path [the path of label xml file]

--tissue_ratio [filter out tiles whose tissue ratio is below this value]

If there is no media data in the prefix directory, program will run preprocessing from very begining and could take a long time. Otherwise, preprocessed data will be loaded.

Aside from packages listed in requirment.txt, the program also needs pathml. Please refer to their homepage for instruction of installation.

Example xml file used: test_046.xml from the Camelyon 2016 dataset. Download image from this link: https://drive.google.com/drive/folders/0BzsdkU4jWx9Bb19WNndQTlUwb2M?resourcekey=0-FREBAxB4QK4bt9Zch_g5Mg

https://youtu.be/MdAcIYptNvE