This interactive tool allows you to explore how cAItomorph classifies peripheral blood smears, and visualize the impact of each individual cell.
For further insight, access our paper, code and data (link pending).
Using the drop-down menu at the top left, cAItomorph predictions for a patient from three datasets can be loaded. The following components are visualized:
Patient-level UMAP shows individual patient embeddings and how they cluster according to their disease.
Disease predictions visualize output probabilities for all classes as returned by our algorithm. The highest prediction is considered the final “algorithm diagnosis”. Hovering over datapoints in the UMAP or swarmplot shows how cAItomorph classifies the corresponding cells.
Cell-level UMAP presents cell embeddings provided by the DinoBloom hematology foundation model.
Attention swarmplot (bottom) illustrates which cells are relevant for diagnosis. The algorithm learns to assign higher attention to diagnostically relevant cells.
Hovering over datapoints displays the location of individual cells, their single-cell classifications, and the corresponding cell images.
This figure was developed at the Institute of AI for Health by Fatih Ă–zlĂĽgedik and Furkan DaĹźdelen.
