I'm a writer and computational cancer scientist with expertise in computer vision, machine learning, and big data. I leverage computational pathology to better understand the biological mechanisms which inform disease prognosis. I encourage you to test the models below on your own datasets or expand upon them with your own creativity! Feel free to contact me for implementation assistance.
You can find a complete list of my publications on Google Scholar.
Due to hardware limitations and because of the massive size of a whole slide image (WSI), survival modelling on WSI datasets is typically done using a two-stage approach: encoding and aggregation. First, a model is trained on WSI tiles or patches, and then a second method is used to aggregate information generated from the output of the first stage to output a final prediction. EPIC-Survival bridges encoding and aggregation into an end-to-end survival modelling approach, while introducing a new loss term, stratification boosting, to encourage the model also to discriminate between risk groups for subtyping.
Before the rise of self-supervised learning, convolutional autoencoders were the standard for converting WSIs into low-dimensional feature vectors. However, a simple MSE-loss does not encourage the model to group together similar features in the embedding space. This model combines a clustering loss with MSE. When applied to WSIs, each cluster produced within the embedding space can be interpretted as a morpholical feature of histology phenotypes across a dataset. These features can then be used as covariates in downstream prognostic modelling tasks with success.