Open JohnTigue opened 4 years ago
This is unsupervised learning low-D representation. Which one would then neighborhooder (umap/tsne) to visualize. The training method is triplet-loss. They pick a point in a 3D neuron. They take two pictures of it from similar angles. They take a third picture randomly from somewhere else on the neuron. Those three are the triplet for triplet-loss training.
Ok, then for brightfield just take locally 3D cropped MinIPs. Three of those to triplet-loss train on. In this brightfield case (not original EM domain) the lighting direction is fixed along the optical axis. EM is free from. Ergo, using the MinIP because all images will be align along the light direction so can learn specialized to that environment/context.
Learning cellular morphology with neural networks