Brain connectomics stops scanning every slice slowly: ZoomSEM rebuilds 16× detail from real low-resolution SEM scans

Serial electron microscopy gains a new route to throughput: fast low-resolution scans can be turned into high-resolution images instead of scanning every slice slowly. Cubic-millimeter-scale serial EM imaging previously took months to years, and deep-learning super-resolution methods were limited to small upscaling factors or relied on synthetic data, giving modest acceleration. Across datasets of diverse biological samples, ZoomSEM generalized robustly under realistic low-resolution acquisition and preserved the fine neuronal structures needed for connectomic analysis.

The model learns the resolution relationship between real scans, so low-resolution input still yields credible detail. The super-resolution method, which maps low-resolution images to high-resolution ones with a model, is trained on real low- and high-resolution SEM pairs rather than synthetic data. Training and evaluation cover 4×, 8× and up to 16× upscaling, with up to 30-fold effective acceleration.

The demonstrated scope is super-resolution reconstruction on datasets of diverse biological samples at 4×, 8× and up to 16× upscaling, with up to 30-fold effective acceleration, evaluated under realistic low-resolution acquisition.

The summary states robust generalization across datasets of diverse biological samples but gives no per-dataset quantitative error, and does not say whether 16× upscaling and 30-fold acceleration were achieved under the same condition.