← All projects
NeuroscienceArchived2026

MidasMap

Immunogold particle detection in electron microscopy

Finds gold particles in electron-microscope images of synapses, so nobody has to count them by hand.

Input patch, 6 nm heatmap, 12 nm heatmap and detections, on a synthetic image
0.943overall F1, leave-one-image-out
0.944F1 for 6 nm beads (AMPA)
0.909F1 for 12 nm beads (NMDA)

The problem

To see where receptors sit at a synapse, neuroscientists tag them with tiny gold beads and image the tissue with a transmission electron microscope. Then someone counts every bead by hand. They're a few pixels wide in a 2048×2048 image, on noisy tissue, and make up well under 0.1% of the pixels.

How it works

The backbone starts from CEM500K, self-supervised on about 500,000 electron microscopy images, instead of ImageNet photos. A stride-2 output and a BiFPN keep a 6 nm bead big enough for clean peak extraction, and focal loss with particle-centered patches handles the imbalance, where plain BCE collapses to all background.

Architecture and inference flow.
Architecture and inference flow.

Limitations

Ten annotated images, 453 hand-labeled particles, each fold testing on one held-out image.

  • Those ten images come from one lab; it hasn't been tested on other microscopes.
  • The F1 values are from the original evaluation, whose metrics file isn't in the repo. An earlier summary figure shows 0.928 on 429 particles.
  • The real images belong to the lab that provided them, so the figures here use a synthetic image.