Segmentation Learning Through a Hands-On Proxy

Diagram the FlyWire processing pipeline, distinguishing model training, segmentation inference, agglomeration, human proofreading, and connectome materialization.

Diagnose how image ambiguity can produce false-split and false-merge errors in example electron-microscopy regions.

Prepare a small pre-labeled electron-microscopy dataset as training and held-out arrays for a provided segmentation pipeline.

Train a lightweight baseline segmenter on the labeled image set.

Calculate intersection-over-union on the held-out images, inspect representative errors, and explain why this experiment illustrates—but does not retrain—FlyWire's production segmentation system.