FPicker: Topology-Guided Evolution for Filament Tracing in Low-SNR Microscopy

Tingyin Zhao; Mingtao Huang; Yuan Shen

Source-linked AI summary

FPicker: Topology-Guided Evolution for Filament Tracing in Low-SNR Microscopy

Tingyin Zhao, Mingtao Huang, Yuan Shen

arXiv:2609.08305v1Sep 8, 2026cs.CVcs.AI

TL;DR

Cryo-EM filament tracing must recover flexible, intersecting structures from extremely low-SNR images, while existing paradigms suffer topological and geometric failures. FPicker combines center-endpoint perception with open-curve evolution and curriculum-based training to model filament topology explicitly. It achieves strong extreme-noise performance and state-of-the-art fine-tuned results on real-world EMPIAR data, while retaining scope limitations around branching and occasional over-merging.

Problem
Cryo-EM filament tracing is needed for helical parameter estimation and 3D reconstruction but is difficult under intersecting topologies and extremely low SNR.
Source: §1 · p. 2
Method
FPicker uses a center-endpoint representation, open-boundary graph evolution, and a bridge curriculum to learn filament geometry under extreme noise.
Source: §1 · p. 3, §3.5 · p. 8
Results
76.5% mSAP and a 9.8% gap rate under Extreme noise (-20 dB), with 82.9% fine-tuned mSAP on Custom-EMPIAR.
Source: §6 · p. 15
Takeaways & Limitations
The results support modeling intrinsic physical priors as a robust geometric backbone for signal-limited scientific imaging and sim-to-real transfer.
Source: §1 · p. 3, §6 · p. 15
FPicker does not provide unified modeling of true branching and can occasionally over-merge aligned filament fragments.
Source: §5 · p. 15

Abstract

from arXiv · show

Automating filament tracing in Cryo-Electron Microscopy (Cryo-EM) is essential for 3D helical reconstruction but challenged by intersecting topologies and extremely low Signal-to-Noise Ratios ($\text{SNR} = σ_s^2/σ_n^2$ < 0.1 or -10 dB). Existing paradigms fail: pixel-wise segmenters suffer from severe topological fracturing, box-based detectors face ghost center drift, sequential trackers derail due to error accumulation, and traditional active contours collapse under artificial closed-curve constraints. To resolve these bottlenecks, we present FPicker, the first topology-guided framework reconciling these incompatibilities. It unifies perception via a center-endpoint representation and an open-curve evolution module to explicitly model non-cyclic connectivity. On simulated benchmarks, FPicker outperforms top baselines by over $40\%$ relative gain in mean spatio-angular precision (mSAP) and reduces topological gap rates by over $60\%$ under extreme noise ($-20\text{ dB}$). By learning intrinsic physical geometry rather than local texture, FPicker demonstrates strong potential as a resilient geometric backbone. Its zero-shot performance on the real-world EMPIAR dataset exhibits robust topological resistance, achieving a state-of-the-art 82.9\% mSAP upon fine-tuning. Our results also suggest modeling physical priors is a highly robust path toward bridging the sim-to-real gap in signal-starved scientific imaging. The code is publicly available at: https://github.com/tomzhaosky/FPicker.

1 Introduction

Cryo-EM filament tracing is difficult because flexible, intersecting filaments must be recovered from extremely low-SNR images, where existing geometric paradigms break down. FPicker addresses this gap with center-endpoint perception, open-curve evolution, and training strategies designed for resilient geometric modeling.

Motivation

Cryo-EM filament picking remains a laborious bottleneck for helical parameter estimation and subsequent 3D helical reconstruction.Source: §1 · p. 2
At SNR < 0.1 or -10 dB, target signal is often locally indistinguishable from background noise.Source: §1 · p. 2

Existing limitations

Box-based detectors suffer ghost-center drift, while segmentation models can fracture topology when tracing flexible intersecting filaments.Source: §1 · p. 2
Sequential trackers accumulate errors into background, and cyclic active contours pull open-filament tips together.Source: §1 · pp. 2–3

FPicker

FPicker replaces box priors with a center-endpoint representation and uses open-boundary graph evolution to model non-cyclic filament topology.Source: §1 · p. 3
A bridge curriculum decouples local deformation-physics learning from global localization variance during extreme-noise end-to-end training.Source: §1 · p. 3

Results

Optimizing within Cryo-Sim yields effective zero-shot transfer and state-of-the-art fine-tuning performance on real-world EMPIAR micrographs.Source: §1 · p. 3

2 Related Works

Related filament-tracing methods use box linking, segmentation followed by skeletonization, or explicit topology modeling, but each has geometric weaknesses in low-SNR intersecting structures. These weaknesses motivate a representation that preserves instances and continuity without relying on ambiguous local evidence.

Box-based detection

Box-based detectors represent flexible 1D curves with 2D boxes, causing background-heavy proposals, anchor-size tuning, and ghost-center drift.Source: §2.1 · p. 3

Segmentation

Segment-then-skeletonize methods extract contours but can merge distinct filament instances at 2D projection crossovers.Source: §2.2 · p. 4

Topology modeling

Sequential trackers extend trajectories from local seeds but can irreversibly derail into background through Markovian error accumulation under extreme noise.Source: §2.3 · p. 4

3 Methods

FPicker uses a coarse-to-fine, two-stage pipeline that generates center-endpoint skeletal priors and refines them as open curves. Auxiliary geometric supervision, global-local feature fusion, and bridge curriculum training support topology-preserving deformation in extreme noise.

Overview

FPicker maps each Cryo-EM micrograph to open skeletal trajectories through a shared feature extractor, proposal generation, and topological refinement.The architecture is end-to-end trainable and follows a coarse-to-fine design.Source: §3.1 · pp. 4–5

Proposal Generation

Stage I detects topological centroids and endpoint displacements, interpolating them into linear skeletal priors that anchor subsequent curve evolution.Centroids are constrained to filament arc-length medians, while the priors define a rotation-aware capture range.Source: §3.1 · p. 5, §3.2 · pp. 5–6

Open-Curve Evolution

Stage II samples vertex features along each prior and iteratively deforms an open graph, preserving axial integrity without closed-curve tip retraction.The graph connects each internal vertex only to its immediate neighbors, and open-boundary aggregation sets tip tension to zero.Source: §3.3 · p. 6
Local OBC and global feature streams fuse to guide deformation, with data-dependent global descriptors broadcasting high-confidence structural cues to low-signal vertices.The global stream aggregates salient signals while auxiliary geometric regularization helps resist high-frequency noise.Source: Figure 3 · p. 6, §3.3 · p. 7

Training Strategy

The bridge curriculum gradually replaces ground-truth initialization with predicted priors, decoupling local deformation learning from early localization variance.Ground-truth injection is used first, then annealed toward autonomous inference as the detector stabilizes.Source: Figure 2 · p. 5, §3.1 · p. 5, §3.5 · p. 8

Loss Function

The evolution loss combines vertex fitting with edge-length uniformity to align curves with skeleton topology while maintaining physically plausible vertex spacing.This regularization is intended to prevent vertex clustering and bridge fragmented signals in extreme low-SNR conditions.Source: §3.6 · pp. 9–10

4 Experiments

FPicker is evaluated across simulated noise regimes, real-world transfer, data efficiency, and component ablations. It consistently addresses ghost-center collapse, topological fracturing, and endpoint shrinking while retaining strong performance under extreme noise.

Experimental setup

Cryo-Sim contains 20,000 simulated micrographs spanning High (0.1, -10 dB), Medium (0.05, -13 dB), and Extreme (0.01, -20 dB) noise regimes.Custom-EMPIAR contains 500 manually annotated real-world micrographs.Source: §4.1 · p. 10
FPicker is benchmarked against box-based detectors, pixel-wise segmenters, and closed active contours using mSAP, F1-score, gap rate, clDice, fp-clDice, and penalized tangent error.mSAP evaluates spatial Chamfer distance under an angular tolerance of Δθ < 15°.Source: §4.1 · p. 10, §4.2 · p. 10

Comparison with state-of-the-art

At Extreme noise (-20 dB), YOLOv8 and crYOLO fall to 28.5% and 25.4% mSAP, respectively, while YOLOv8 reaches a 45.2% gap rate.The results are attributed to bounding-box centers drifting into background signal.Source: §4.2 · pp. 10–11
FPicker achieves a 9.8% gap rate and 20.5° angle error under Extreme noise, indicating that open-curve evolution reduces endpoint tension while preserving continuity.Segmentation methods show high gap rates, and Deep Snake reaches 26.2° angle error at Medium SNR.Source: §4.2 · pp. 11–12

Sim-to-real transfer

FPicker reaches 82.9% mSAP after fine-tuning on Custom-EMPIAR, following zero-shot performance of 42.4% clDice and 28.6% mSAP.SAM 2 reaches 1.5% zero-shot mSAP, while established detectors reach 25.4% and 26.2% mSAP in the cited comparison.Source: §4.3 · pp. 12–13

Data efficiency

Fine-tuning FPicker with 10 images raises mSAP from 28.6% to 55.4%, and 50 images raise it to 70.1%.The reported adaptation uses sparse real-world annotations after Cryo-Sim training.Source: §4.4 · p. 13

Ablation studies

Ablations show that the center-endpoint representation raises EMPIAR mSAP to 52.5%, open-curve evolution lowers P-Ang from 45.2° to 19.4° at Medium SNR, and auxiliary fields raise Extreme-noise mSAP from 35.6% to 45.2%.The bridge curriculum further raises Extreme-noise mSAP to 76.5% and EMPIAR mSAP to 82.9%.Source: §4.5 · p. 14

Iteration refinement

Performance rises from 78.5% mSAP at T = 1 to 91.1% at T = 3, while inference speed remains between 38.3 and 36.9 FPS.The iteration analysis reports saturation around three deformation iterations.Source: §4.5 · p. 15

5 Limitations

FPicker remains bounded by its current topology and continuity assumptions. The paper identifies limitations in branching structures and occasional over-merging of aligned fragments.

Topology scope

FPicker does not jointly model branches such as Y-junctions, limiting its direct extension to vascular or neural branching.The authors state that true molecular branching is absent from the target Cryo-EM filaments and leave dynamic node degrees for future work.Source: §5 · p. 15

Continuity trade-off

FPicker occasionally over-merges aligned fragments despite a reported Over-Merging Rate of 7.96%.Explicitly repelling competing tips without sacrificing noise resistance remains unresolved.Source: §5 · p. 15

6 Conclusions

FPicker addresses filament-tracing incompatibilities by combining center-endpoint representation with open-boundary graph evolution. Under extreme noise, it achieves strong simulated performance and state-of-the-art fine-tuned performance on Custom-EMPIAR.

Conclusions

FPicker replaces rigid box priors with center-endpoint representation and unites holistic graph evolution with strictly open-boundary constraints.This design targets local tracking derailment and cyclic shrinking in filament tracing.Source: §6 · p. 15
76.5% mSAP and a 9.8% gap rate are achieved under Extreme noise (-20 dB).These results are reported alongside prevention of ghost center collapse and endpoint shrinking.Source: §6 · p. 15
82.9% fine-tuned mSAP is achieved on Custom-EMPIAR after training on Cryo-Sim.The reported training strategy also delivers effective zero-shot transfer to real-world data.Source: §6 · p. 15
Modeling intrinsic physical priors is presented as more robust than brute-force data scaling for signal-limited scientific imaging.The conclusion frames FPicker as a resilient geometric backbone for these regimes.Source: §6 · p. 15