The next steps for this project are directed at making it more accessible and usable for pathologists. This includes finalizing the uncertainty quantifications for all of our models, testing our models on a secondary white blood cell dataset from researchers in Barcelona (Acevado et al., 2020), and exploring the possibility of transfer learning for the models we’ve built. Finally, we would love to hear any feedback about our project; feel free to contact us via the form on our team page below:
A prototype of our product can be seen below. Users can upload a single cell image of a white blood cell and select a confidence level they want for the model’s prediction. The model then provides a cell classification along with a set of cells that ensures, with the provided confidence level, that the correct label is in the set as well as providing the certainty of those predictions. For example in the image below a researcher submitted an image of a bilobed promyelocyte and said that they want a prediction set with 95% confidence. The model then outputs its best guess, “Bilobed Promyelocyte”, as well as the other top predictions that reach the 95% interval as well as the confidence of each prediction.