Automatic repair of the public MICrONS segmentation, tested once on nine unseen regions

Dan Siroker, October 2026. Every number comes from the evaluation report's tables. The tables, the pre-registration file with its 72 hashes, the execution record, the per-neuron outputs and the scoring code are in one download: micrONS-repair-evaluation-2026-10-04.zip (1.5 MB).

Question

Starting from the released minnie65 segmentation (public MICrONS), how much expected run length can an automatic pipeline recover, trained on the sparse proofreading that exists, and does an image-based second stage add to it?

Method, frozen before the test

1. Baseline. Connected components within each padded evaluation crop: pieces that do not touch inside the crop are relabelled. This corrects the start state for a cropped evaluation and is not counted as repair.
2. Pair model. A 3D network over candidate joins between segments, pre-trained on cuts made in the segmentation's own segments (no proofread labels), then fine-tuned on 1,294 positive join examples labelled from the proofread state of four other regions.
3. Tip judge. "Does this tip continue?", retrained deterministically from the four training regions' tip tables, threshold 0.96.
4. Budget. 1,000 pair-model rows per region plus the judge's rows. This is J3.

J4 adds image-based joins from a dense affinity network on 16 x 16 x 40 nm voxels (pooled affinity at least 0.9 over at least 4,096 voxel faces). The candidate (J4), the comparator (J3), the success rule (gain at least 0.005 with the paired 95 percent lower bound above zero) and the hashes of 72 artefacts were written down before the nine regions were opened. Candidate-side steps read a label-free cache; only the evaluator read the truth.

Evaluation

Nine regions never used for training, tuning or selection, each at least 30 um from every training region: 1,754 proofread neurons (4,433 neuron-region records). Normalised expected run length (nERL) on TEASAR skeletons. Plain nERL counts a merger where a segment carries nodes of more than one proofread skeleton; strict nERL also counts a merger wherever a segment carries any node of another proofread cell, finished or not, so it is the harsher of the two. Both use our reference skeletons (388.9 mm of path). The audited axon-like cohort uses the PATHFINDER preprint's skeleton rules (Januszewski et al., bioRxiv 2025: 100 nm endpoint erosion, 300 nm nodes, 5-node splits, 10-node mergers, face exclusion for splits) on the synapse-defined axon-like cohort (345.5 mm, of which 82.8 mm axon-like), with mergers detected against every proofread skeleton. Intervals are 95 percent paired bootstraps grouped by neuron across regions.

Three models appear on these pages and they are different things. Proofreader is the 56M-parameter benchmark judge scored on ConnectomeBench2; it is not part of J3. J3's pair model is the 3D network over candidate joins described above, and its tip judge is a separate classifier; both are trained on proofread labels from the four training regions.

Result

J3 raises plain nERL by +0.040 over the baseline, positive in every one of the nine regions (leave-one-region-out range +0.035 to +0.042), and raises the audited axon-like cohort by +0.026. Breaks (splits) fall from 2,826 to 2,198 on the 388.9 mm of proofread path, 7.27 to 5.65 per mm, fewer in every region; wrong joins (mergers) go from 290 to 299, 0.75 to 0.77 per mm. The image-based second stage did not meet its pre-registered criterion: its gain sits on one region and ten neurons.

Quality and scope

Of the 20,578 join rows J3 selected across the nine regions, 269 are decidable against the finished-cell site truth: 238 correct and 31 wrong (precision 0.885 [0.841, 0.918]; pair-model rows 0.886, tip-judge rows 0.884). A further 148 rows are ambiguous (a side inside the site window with no decision, or a site not found); counting them as wrong gives 0.571. 410 rows join the same proofread but unfinished cell, 113 connect pieces of the same unproofread final root, and 19,638 rows (95.4 percent) connect objects outside the proofread-cell set. The evaluation therefore measures run length on proofread cells, which is what the truth supports; precision on the rest is not measured. Segment volume outside its largest annotated cell, summed over segments containing reference nodes, rises from 40,820 to 43,179 um3 under J3 (+5.8 percent), a proxy that includes other annotated cells and legitimate unannotated continuations.

Scope: nine regions of one heavily proofread dataset, scored against its own proofreading. A held-out-region estimate within minnie65, not evidence about other animals, and not an estimate of precision on unannotated tissue. Training used proofreading labels from four other regions; no evaluation truth was used during inference.

The candidate stage without proofread labels

The same 3D network trained three ways, scored unchanged on three held-out development regions (965 logged merges in all). Counts are logged merges recovered in the top 1,000 / 5,000 / 20,000 candidates per region, summed over the three regions. One seed per arm.

At 20,000 the cuts-only network recovers more logged merges than either labelled arm on every region (paired lower bounds above zero); at 5,000 the three are level; at 1,000, the budget the pipeline uses for its pair-model rows, the logged-merge network is ahead. This is MergeNet's construction run backwards: synthetic splits instead of synthetic merges, scored on the merges proofreaders actually made. The final pipeline fine-tunes on proofread labels.

Reproducibility

72 hashed artefacts (checkpoints, scripts, inputs) in a GO file written before scoring, and an execution record with the hash of every produced file and the machine that produced it. The pair model's validation scores reproduced bit for bit on a second GPU and driver stack, and to the top-1,000 ranking on a third; the join chain and the shape scorer reproduced byte for byte from the label-free cache. An independent review reproduced the intervals from the saved per-neuron outputs.

Two open questions

1. What independent annotation design would measure repaired breaks and newly introduced mergers among operations on currently unannotated tissue?
2. What training distribution or spatial context is needed to turn synthetic-split proposal recall into precise automatic axon repair?