CryoTracer: A Synthetic-to-Real System for Filament Instance Tracing in Cryo-EM

Radosław Kuczbański; Wojciech Kozłowski; Mateusz Grzesiuk; Konrad Karanowski; Lukas Frey; Andrzej Rzepiela; Maciej Zięba

cs.CVOct 13, 2026

CryoTracer: A Synthetic-to-Real System for Filament Instance Tracing in Cryo-EM

Authors: Radosław Kuczbański, Wojciech Kozłowski, Mateusz Grzesiuk, Konrad Karanowski, Lukas Frey, Andrzej Rzepiela, Maciej Zięba

Organizations: WUST, Poland · ETH Zurich, Switzerland · Tooploox, Poland

Abstract

Particle picking in cryo-electron microscopy is commonly formulated as an object-detection problem in which individual particles are represented by bounding boxes, a natural representation for globular particles but a poor match for filaments that can extend across an entire micrograph. Existing methods reconstruct filament trajectories by linking local predictions in a separate post-processing stage, decoupling detection from instance formation and creating potential ambiguities at gaps and crossings. Directly preserving filament identity is particularly useful for polymorphic specimens, where segments originating from the same filament should remain associated. However, learning such instance-level representations is challenging because expert centerline annotations are scarce and costly to obtain. We introduce CryoTracer, a synthetic-to-real framework that addresses this limitation through procedural micrograph generation, transformer-based polyline prediction, and specimen-specific fine-tuning. Procedural pretraining exposes the model to diverse filament geometries, intersections, and imaging conditions before adaptation using a limited number of fully annotated experimental micrographs. CryoTracer represents filaments as ordered centerline polylines and optimizes their geometry using an orientation-invariant discrete Fréchet objective, enabling consistent learning regardless of tracing direction. Evaluation on established and newly expert-annotated datasets shows that CryoTracer improves instance-level filament recovery over existing particle-detection approaches. These results establish end-to-end polyline prediction as a practical alternative to detection-and-linking pipelines for recovering complete filament instances with limited micrograph annotation budgets. Code and data are available at https://github.com/cryotracer/cryotracer

Figures & tables

Panel of Figure 1 — Figure 1: Polyline prediction and iterative refinement in CryoTracer. (a) The $Q=50$ highest-scoring encoder proposals provide initial reference polylines with $K=5$ ordered points. (b) Cross-attention updates each query using encoder features around its current polyline (orange). Purple markers indicate sampling locations, with marker area proportional to attention weight. (c) The updated query refines the polyline, and the resulting curve (teal) guides attention in the next decoder layer. This process progressively refines the predictions.

Method	mF1	mRecall	mPrecision	Box IoU	Coverage	Pick count	Mean pick length (nm)
Standard regime: ground truth						347	181
CryoTracer (ours)	0.63	0.63	0.63	0.84	0.90	337	176
Topaz	0.08	0.13	0.06	0.82	0.88	947	56
crYOLO	0.03	0.03	0.03	0.66	0.69	329	122
Practical regime: ground truth						1,578	190
CryoTracer (ours)	0.54	0.53	0.56	0.81	0.86	1,474	184

Table 1: Filament detection on EMPIAR-10022 in the standard and practical regimes. mF1, mRecall, and mPrecision average per-micrograph scores over five distance tolerances (Section 4.1 ). Box IoU measures spatial overlap; coverage is the fraction of annotated area recovered. Pick count and mean polyline length (nm) summarize filament traces, with ground truth as a reference. CryoTracer values are means over five training seeds. Bold marks the highest unrounded score per metric and regime; pick statistics are descriptive.

Method	mF1	mRecall	mPrecision	Box IoU	Coverage	Pick count	Mean pick length (nm)
Standard regime: ground truth						272	203
CryoTracer (ours)	0.40	0.43	0.39	0.75	0.84	278	176
Topaz	0.01	0.04	0.01	0.59	0.80	1,659	33
crYOLO	0.01	0.01	0.01	0.60	0.66	540	60
RELION Amyloid	0.11	0.11	0.11	0.59	0.63	342	94
Practical regime: ground truth						1,364	203

Table 2: Filament detection on

\alpha

-synuclein in the standard and practical regimes. mF1, mRecall, and mPrecision average per-micrograph scores over five distance tolerances (Section 4.1 ). Box IoU measures spatial overlap; coverage is the fraction of annotated area recovered. Pick count and mean polyline length (nm) summarize filament traces, with ground truth as a reference. CryoTracer values are means over five training seeds. Bold marks the highest unrounded score per metric and regime; pick statistics are descriptive.

					EMPIAR-10022		$\alpha$ -synuclein
Experiment	Training	Distance	Points	Pyramid levels	mF1	Box IoU	mF1	Box IoU
Standard regime
Main	Synthetic + real	Fréchet	5	4	0.62	0.85	0.44	0.75
Training strategy	Synthetic only	Fréchet	5	4	0.00	0.05	0.00	0.01
	Real only (5k)	Fréchet	5	4	0.16	0.49	0.18	0.51
	Real only (25k)	Fréchet	5	4	0.59	0.83	0.39	0.71

Table 3: CryoTracer ablations of synthetic initialization, curve supervision, and architecture. Synthetic pretraining uses up to 20,000 steps; real-data training uses up to 5,000 steps unless marked 25k. Bold labels identify the main model configuration. Higher mF1 and Box IoU are better. Training and architecture details appear in Appendix B.1 .

Figure 2: EMPIAR-10022, standard regime. One selected micrograph from the corresponding evaluation set, with method predictions and ground truth labeled in the panels. Images are low-pass filtered and contrast adjusted for display. Colors distinguish curves within a panel and do not indicate matched identities across methods.

Figure 3: $\alpha$ -synuclein, standard regime. One selected micrograph from the corresponding evaluation set, with method predictions and ground truth labeled in the panels. Images are low-pass filtered and contrast adjusted for display. Colors distinguish curves within a panel and do not indicate matched identities across methods.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: EMPIAR-10022 dataset examples. Each row pairs an image (left) with a low-pass-filtered image and its ground-truth centerlines (right). Colors distinguish annotated filament instances.

Figure 5: $\alpha$ -synuclein dataset examples. Each row pairs an image (left) with a low-pass-filtered image and its ground-truth centerlines (right). Colors distinguish annotated filament instances.

Figure 6: Synthetic dataset examples. Six synthetic micrographs are shown with their ground-truth centerlines. Colors distinguish annotated filament instances.

		P@25	R@25	F1@25	P@50	R@50	F1@50	P@100	R@100	F1@100	P@150	R@150	F1@150	P@200	R@200	F1@200	Box IoU	Coverage	Pick count	Mean pick length (nm)
Regime	Method
standard	CryoTracer (ours)	0.25 $\pm$ 0.04	0.25 $\pm$ 0.04	0.25 $\pm$ 0.04	0.55 $\pm$ 0.02	0.55 $\pm$ 0.02	0.55 $\pm$ 0.02	0.74 $\pm$ 0.02	0.74 $\pm$ 0.02	0.74 $\pm$ 0.02	0.80 $\pm$ 0.02	0.79 $\pm$ 0.02	0.79 $\pm$ 0.02	0.82 $\pm$ 0.02	0.82 $\pm$ 0.01	0.82 $\pm$ 0.02	0.84 $\pm$ 0.01	0.90 $\pm$ 0.01	337 $\pm$ 4	176 $\pm$ 3
	Topaz	0.01	0.02	0.01	0.02	0.04	0.03	0.06	0.13	0.08	0.09	0.22	0.13	0.11	0.25	0.15	0.82	0.88	947	56
	crYOLO	0.01	0.01	0.01	0.01	0.01	0.01	0.02	0.02	0.02	0.04	0.03	0.04	0.06	0.05	0.06	0.66	0.69	329	122
	Ground truth	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	347	181
practical	CryoTracer (ours)	0.14 $\pm$ 0.04	0.13 $\pm$ 0.04	0.14 $\pm$ 0.04	0.45 $\pm$ 0.06	0.43 $\pm$ 0.05	0.44 $\pm$ 0.05	0.67 $\pm$ 0.05	0.64 $\pm$ 0.04	0.65 $\pm$ 0.04	0.74 $\pm$ 0.04	0.71 $\pm$ 0.03	0.72 $\pm$ 0.04	0.78 $\pm$ 0.04	0.75 $\pm$ 0.03	0.76 $\pm$ 0.03	0.81 $\pm$ 0.01	0.86 $\pm$ 0.01	1,474 $\pm$ 34	184 $\pm$ 5

Appendix

Table 4: Detailed filament-detection results on EMPIAR-10022. Precision (P), recall (R), and F1 are averaged over micrographs at each distance tolerance (in Å). Box IoU and coverage measure spatial overlap. Pick count and mean polyline length (nm) summarize filament traces, with ground truth as a reference. CryoTracer values are mean

\pm

sample standard deviation over five seeds, with statistics computed separately per seed. Bold marks the highest unrounded score per metric and regime. Pick statistics are descriptive. Regimes use different evaluation sets.

		P@25	R@25	F1@25	P@50	R@50	F1@50	P@100	R@100	F1@100	P@150	R@150	F1@150	P@200	R@200	F1@200	Box IoU	Coverage	Pick count	Mean pick length (nm)
Regime	Method
standard	CryoTracer (ours)	0.02 $\pm$ 0.01	0.03 $\pm$ 0.02	0.03 $\pm$ 0.02	0.18 $\pm$ 0.04	0.20 $\pm$ 0.04	0.19 $\pm$ 0.04	0.47 $\pm$ 0.05	0.51 $\pm$ 0.06	0.48 $\pm$ 0.05	0.62 $\pm$ 0.04	0.67 $\pm$ 0.04	0.63 $\pm$ 0.04	0.68 $\pm$ 0.03	0.73 $\pm$ 0.01	0.69 $\pm$ 0.02	0.75 $\pm$ 0.01	0.84 $\pm$ 0.01	278 $\pm$ 11	176 $\pm$ 7
	Topaz	0.00	0.00	0.00	0.00	0.01	0.00	0.01	0.05	0.01	0.01	0.07	0.02	0.01	0.08	0.02	0.59	0.80	1,659	33
	crYOLO	0.00	0.00	0.00	0.01	0.01	0.01	0.01	0.01	0.01	0.01	0.01	0.01	0.02	0.02	0.02	0.60	0.66	540	60
	RELION Amyloid	0.00	0.00	0.00	0.00	0.00	0.00	0.13	0.12	0.12	0.20	0.20	0.20	0.23	0.23	0.22	0.59	0.63	342	94
	Ground truth	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	272	203

Appendix

Table 5: Detailed filament-detection results on

\alpha

-synuclein. Precision (P), recall (R), and F1 are averaged over micrographs at each distance tolerance (in Å). Box IoU and coverage measure spatial overlap. Pick count and mean polyline length (nm) summarize filament traces, with ground truth as a reference. CryoTracer values are mean

\pm

standard	Main	Synthetic + real	Fréchet	5 points, 4 levels	0.23	0.23	0.23	0.54	0.54	0.54	0.74	0.73	0.73	0.79	0.79	0.79	0.83	0.82	0.82	0.85	0.90
					P@25	R@25	F1@25	P@50	R@50	F1@50	P@100	R@100	F1@100	P@150	R@150	F1@150	P@200	R@200	F1@200	Box IoU	Coverage
Regime	Experiment	training	distance	architecture
	Training strategy	Synthetic only	Fréchet	5 points, 4 levels	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.05	0.05
		Real only	Fréchet	5 points, 4 levels	0.00	0.00	0.00	0.03	0.02	0.02	0.16	0.12	0.13	0.33	0.23	0.27	0.45	0.31	0.37	0.49	0.56
		Real only (25k)	Fréchet	5 points, 4 levels	0.25	0.24	0.24	0.51	0.50	0.50	0.70	0.68	0.69	0.77	0.74	0.75	0.80	0.78	0.79	0.83	0.88
	Distance metric	Real only	Chamfer	5 points, 4 levels	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.42	0.60

Appendix

Table 6: CryoTracer ablations on EMPIAR-10022. Precision (P), recall (R), and F1 are averaged over micrographs at each distance tolerance (in Å). Box IoU and coverage measure spatial overlap. Synthetic pretraining uses up to 20,000 steps. Real-data training uses up to 5,000 steps unless marked 25k. Synthetic-only rows evaluate the same frozen checkpoint in both regimes. Bold labels identify the main model configuration. Bold scores mark the highest unrounded value per metric and regime. Higher scores are better. Training and architecture details appear in Appendix B.1 .

standard	Main	Synthetic + real	Fréchet	5 points, 4 levels	0.02	0.03	0.02	0.22	0.24	0.23	0.54	0.56	0.54	0.64	0.67	0.64	0.75	0.77	0.75	0.75	0.84
					P@25	R@25	F1@25	P@50	R@50	F1@50	P@100	R@100	F1@100	P@150	R@150	F1@150	P@200	R@200	F1@200	Box IoU	Coverage
Regime	Experiment	training	distance	architecture
	Training strategy	Synthetic only	Fréchet	5 points, 4 levels	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.01	0.02
		Real only	Fréchet	5 points, 4 levels	0.00	0.00	0.00	0.04	0.04	0.04	0.15	0.14	0.15	0.29	0.28	0.27	0.48	0.46	0.45	0.51	0.63
		Real only (25k)	Fréchet	5 points, 4 levels	0.02	0.02	0.02	0.15	0.16	0.15	0.48	0.49	0.48	0.62	0.62	0.61	0.70	0.69	0.68	0.71	0.81
	Distance metric	Real only	Chamfer	5 points, 4 levels	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.35	0.48

Appendix

Table 7: CryoTracer ablations on

\alpha

-synuclein. Precision (P), recall (R), and F1 are averaged over micrographs at each distance tolerance (in Å). Box IoU and coverage measure spatial overlap. Synthetic pretraining uses up to 20,000 steps. Real-data training uses up to 5,000 steps unless marked 25k. Synthetic-only rows evaluate the same frozen checkpoint in both regimes. Bold labels identify the main model configuration. Bold scores mark the highest unrounded value per metric and regime. Higher scores are better. Training and architecture details appear in Appendix B.1 .

Figure 7: Picked-filament length distributions for EMPIAR-10022 (top) and

\alpha

-synuclein (bottom), under the standard (left) and practical (right) evaluation regimes. Length is measured along each polyline in nanometres. Histograms share 40 equal-width bins and show the percentage of each method’s picks in each bin, summing to

100\%

per method within a panel. The dashed curve denotes ground truth. Legend counts give the total number of picks.

Figure 8: EMPIAR-10022, standard regime. Eight selected micrographs from the corresponding evaluation set, with method predictions and ground truth labeled in the panels. Images are low-pass filtered and contrast adjusted for display. Colors distinguish curves within a panel and do not indicate matched identities across methods.

Figure 9: EMPIAR-10022, practical regime. Eight selected micrographs from the corresponding evaluation set, with method predictions and ground truth labeled in the panels. Images are low-pass filtered and contrast adjusted for display. Colors distinguish curves within a panel and do not indicate matched identities across methods.

Figure 10: $\alpha$ -synuclein, standard regime. Eight selected micrographs from the corresponding evaluation set, with method predictions and ground truth labeled in the panels. Images are low-pass filtered and contrast adjusted for display. Colors distinguish curves within a panel and do not indicate matched identities across methods.

Figure 11: $\alpha$ -synuclein, practical regime. Eight selected micrographs from the corresponding evaluation set, with method predictions and ground truth labeled in the panels. Images are low-pass filtered and contrast adjusted for display. Colors distinguish curves within a panel and do not indicate matched identities across methods.

Electron Microscopy

See all 36 papers

FPicker: Topology-Guided Evolution for Filament Tracing in Low-SNR Microscopy
Sep 8, 2026Tingyin Zhao, Mingtao Huang, Yuan ShenCryo-Electron Microscopy Microscopy Image Analysis
Department of Electronic Engineering, Tsinghua University, Beijing, China · Beijing National Research Center for Information Science and Technology, Beijing, China
Reconstruction-Aware Cryo-EM Particle Picking
Aug 28, 2026Riku Itsuji, Yuanhao Wang, Xingjian Li +3Cryo-Electron Microscopy Electron Microscopy
Keio University, Yokohama, Japan · Keio AI Research Center, Yokohama, Japan · Carnegie Mellon University, Pittsburgh, PA, USA
Score-Based Matching with Target Guidance for Cryo-EM Denoising
Apr 20, 2026Xiaoqi Wu, Xueying Zhan, Wen Li +3Cryo-Electron Microscopy Score-Based Generative Modeling
Carnegie Mellon University · Towson University

Sep 8, 2026J/K move · Enter open · S save

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

Tingyin Zhao, Mingtao Huang, Yuan Shen

Department of Electronic Engineering, Tsinghua University, Beijing, China · Beijing National Research Center for Information Science and Technology, Beijing, China

Cryo-Electron Microscopy Microscopy Image Analysis Electron Microscopy

Open paper AskPDF arXiv

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.

Abstract

Electron Microscopy

Explore similar work

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

Reconstruction-Aware Cryo-EM Particle Picking

Score-Based Matching with Target Guidance for Cryo-EM Denoising