AI-guided electron microscopy for accelerated brain mapping

Electron microscopy sits at the core of connectomics as the most reliable technique for resolving synapses and wiring at the nanometer scale. However, acquiring such vast datasets is time-consuming and expensive which limits the practice primarily to large specialized centers. We identify that the root cause is that traditional workflows use the electron beam blindly. They treat imaging as a pixel-perfect copying process and expend the same precious beam time on featureless resin as on dense synaptic neuropil. This uniform fidelity creates massive inefficiency because segmentation success relies on resolving a minority of complex features rather than maintaining equal quality everywhere.

We present SmartEM to make connectomics adoptable in standard labs. Its goal is a biological answer rather than a flawless digital replica. Mimicking human vision, SmartEM integrates goal-directed logic into the microscope’s control loop. The process begins with a fast, low-fidelity "scout scan." A neural network evaluates this image in real time to predict where higher quality is needed by identifying regions of high segmentation uncertainty or concentrated biological value. The microscope selectively revisits these targets using rapid beam deflections. This eliminates time-consuming mechanical steps and the need for on-the-fly image alignment. Because brain tissue is heterogeneous, this adaptive refinement yields a sevenfold reduction in beam time across nematode, mouse, and human samples while preserving segmentation accuracy. For an adult C. elegans, this drops imaging time from roughly two months to eight days. SmartEM represents a shift from exhaustive imaging to adaptive acquisition. It upgrades widely available single-beam instruments into high-throughput systems and points toward a “thinking microscope” that uses context to decide what it needs to see next.

Learning Objectives:

1. Build a basic understanding of the use of SEM for connectomics.

2. Discuss the SmartEM pipeline and how it integrates AI into the microscope's control loop.

3. Review the strategy for training a neural network to detect errors by comparing fast and slow imaging data.