wiring oplabels2d × labels2d × table → tableimport fullseye as fs; fs.ledger.seg_synapse_nri(a, b, synapses, spacing=None, background=0) (実装を直接呼ぶなら import segcompare; segcompare.seg_synapse_nri(a, b, synapses, spacing=None, background=0)、台帳から引くなら opsemproof.get("seg_synapse_nri"))Neural reconstruction integrity (Reilly et al. 2018) of candidate b against truth a.
NRI scores the pairs of synapse ends that a reconstruction keeps together. Over the
2n ends (synapses as in :func:seg_synapse_partners; ends on background in
the truth are dropped, ends on background in the candidate are singletons), a pair of
ends that lies in one truth neuron and one candidate object is a true positive; a
pair in one truth neuron but different candidate objects is a false negative (a split);
a pair in one candidate object but different truth neurons is a false positive (a
merge). nri = 2TP / (2TP + FP + FN) with the pair precision and recall.
Identity (checked inside, fail-closed): with n_ij the contingency table of the ends,
TP = sum C(n_ij, 2), TP + FN = sum C(s_i, 2), TP + FP = sum C(t_j, 2), so NRI
is exactly 1 - adapted_rand_error of :func:seg_rand taken over the ends (the
- N form of the pair counts is 2 C(n, 2)). Both are computed and must agree to
1e-12. Returns nri, precision, recall, tp, fp, fn, n_ends.
Raises ValueError: as :func:seg_wiring_variation; fewer than 2 ends.
py -3.11 examples/poc_em_wiring_errors.pytable を入力に取れる)seg_synapse_partners · seg_wiring_variation · seg_wiring_exposure
wiring)seg_synapse_partners · seg_wiring_variation · seg_wiring_exposure
Provenance: segcompare.py — EMPROOF operator registry. この per-op ノートは tools/opdocs.py md が自動生成(手編集しない)。
© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.