Carbon-nanotube sheets stop blurring into one blob: PLECTA reads only a binary axis mask and reaches mean F1 0.85 on synthetic scenes
Synopsis
Carbon-nanotube sheets, collagen and paper are built from overlapping strands, and in a micrograph a crossing merges distinct strands while a signal gap splits one strand into fragments. PLECTA takes only a binary mask marking strand centrelines and, at each junction, treats the possible continuations jointly as an exact maximum-weight partial matching, allowing strands to end and crossing pixels to belong to several strands. Across 128 synthetic scenes it reached a mean common-fragment pairwise F1 of 0.85 (95% bootstrap CI [0.84, 0.86]), falling from 0.91 to 0.77 as areal coverage rose from 20% to 60%.
Interpretation
PLECTA turns a binary mask that marks only strand centrelines into a set of overlap-aware single-strand instances, where a crossing pixel may belong to every strand passing through it. Earlier routes either skeletonised the mask, cut out junctions and paired arms greedily by direction with no later revision, or required a greyscale image or labelled training data. Across 128 synthetic scenes the mean common-fragment pairwise F1 was 0.85 (95% bootstrap CI [0.84, 0.86]), falling monotonically from 0.91 at 20% areal coverage to 0.77 at 60%.
The method converts tangent, chord and curvature differences between strand tips into link costs, then jointly selects non-conflicting links at junctions and across admissible gaps. Each tip takes at most one partner or stays unmatched at a penalty, so a strand can genuinely terminate at a junction of any degree instead of being forced onward. On 50 synthetic scenes the degraded axis mask gave mean F1 0.82 against 0.84 for the clean axis of the same scenes, a paired difference of 0.0224 (95% CI [0.0113, 0.0333]).
Matching runs over eight rounds, starting with lower unmatched penalties and tighter gap-angle limits and relaxing them to their configured values, with the final round fixing the output. Tangent and curvature estimates are updated along assembled chains before each round, so later decisions use longer strand geometry rather than only the short stubs near a junction. In the ablations, replacing joint junction solving with greedy pairing dropped F1 from 0.87 to 0.78, removing gap bridging dropped it to 0.73, and removing the junction chord-turn term dropped it to 0.68.
Perspective
The demonstrated scope is 128 synthetic scenes, 40 scenes from GraFT's released generator, and three annotated carbon-nanotube sheet SEM fields; the segmentation core reads only a binary axis mask, while depth ordering uses the matching greyscale image and no volumetric reconstruction is performed.
The method assumes directional continuity through crossings and remains sensitive to upstream mask errors, bridging short gaps but unable to recover omitted strands or reliably correct false axes; the automatic-axis mean F1 of 0.46 lies within the 0.36–0.53 range over which two annotators agree, so it should be read alongside ambiguity in strand identity at local occlusions.
