Topology-aware few-shot segmentation improves rice leaf-cell boundary and count fidelity, while correction is risky under tested domain shifts
- 1School of Big Data and Artificial Intelligence, Nanning College of Technology
Abstract
Introduction: Plant-cell measurements require adjacent cells to remain separated, yet few-shot segmentation is usually assessed mainly by foreground overlap.
Methods: We developed domain-aware topology prompt adaptation (DTPA), which trains a small topology head over a frozen microscopy-specific Segment Anything backbone. We evaluated matched 10-shot segmentation on a leakage-free RiceCell split, conducted a secondary eight-image audit, and assessed label-free routing and selective correction on locked external tests.
Results: DTPA improved boundary F1 and reduced cell-count error relative to ConvNeXt, while reducing Dice. The boundary advantage persisted across one- to five-pixel tolerances and against two matched instance baselines. The secondary audit linked this trade-off to improved cell-count and typical-cell-area fidelity, but not to every image-derived measurement. The label-free router reduced severe under-counting on a locked PlantSeg root test. Selective correction was less reliable under the tested tissue shifts: 14 of 45 accepted Ovules retries were harmful, and the only accepted floral-meristem retry was harmful.
Discussion: Topology-aware prompting and routing are computational tools for plant microscopy, not evidence of biological discovery or universal model superiority. A correction-abstention rule derived after the external audit is only a deployment proposal and requires prospective target-domain validation.
Related
- Chemical and geometric representation fidelity improves drug--target affinity prediction
- Benchmarking long-read aligner-caller pipelines for structural variant detection across ten angiosperm species with PacBio high-fidelity sequencing
- Kinetics and Evolution-guided Design of High Fidelity CRISPR-Cas Endonucleases
- When Raman Spectral Fidelity and Bacterial Classification Accuracy Diverge
- Enhanced Fidelity and Sequence Generality of an RNA Polymerase Ribozyme
BibTeX
@article{wei2026topologyaware,
title = {{Topology-aware few-shot segmentation improves rice leaf-cell boundary and count fidelity, while correction is risky under tested domain shifts}},
author = {Gaixing Wei and Da Zeng and Shenglin Qin and Dezhen Zeng},
journal = {Frontiers in Plant Science},
year = {2026},
month = oct,
doi = {10.3389/fpls.2026.1950986},
url = {https://www.frontiersin.org/articles/10.3389/fpls.2026.1950986/full},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Topology-aware few-shot segmentation improves rice leaf-cell boundary and count fidelity, while correction is risky under tested domain shifts
AU - Gaixing Wei
AU - Da Zeng
AU - Shenglin Qin
AU - Dezhen Zeng
PY - 2026
DA - 2026/10/13
JO - Frontiers in Plant Science
DO - 10.3389/fpls.2026.1950986
UR - https://www.frontiersin.org/articles/10.3389/fpls.2026.1950986/full
AB - Introduction: Plant-cell measurements require adjacent cells to remain separated, yet few-shot segmentation is usually assessed mainly by foreground overlap. Methods: We developed domain-aware topology prompt adaptation (DTPA), which trains a small topology head over a frozen microscopy-specific Segment Anything backbone. We evaluated matched 10-shot segmentation on a leakage-free RiceCell split, conducted a secondary eight-image audit, and assessed label-free routing and selective correction on locked external tests. Results: DTPA improved boundary F1 and reduced cell-count error relative to ConvNeXt, while reducing Dice. The boundary advantage persisted across one- to five-pixel tolerances and against two matched instance baselines. The secondary audit linked this trade-off to improved cell-count and typical-cell-area fidelity, but not to every image-derived measurement. The label-free router reduced severe under-counting on a locked PlantSeg root test. Selective correction was less reliable under the tested tissue shifts: 14 of 45 accepted Ovules retries were harmful, and the only accepted floral-meristem retry was harmful. Discussion: Topology-aware prompting and routing are computational tools for plant microscopy, not evidence of biological discovery or universal model superiority. A correction-abstention rule derived after the external audit is only a deployment proposal and requires prospective target-domain validation.
SN - 1664-462X
ER -