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
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 α -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
α -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: α -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: α -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 ± 0.04
0.25 ± 0.04
0.25 ± 0.04
0.55 ± 0.02
0.55 ± 0.02
0.55 ± 0.02
0.74 ± 0.02
0.74 ± 0.02
0.74 ± 0.02
0.80 ± 0.02
0.79 ± 0.02
0.79 ± 0.02
0.82 ± 0.02
0.82 ± 0.01
0.82 ± 0.02
0.84 ± 0.01
0.90 ± 0.01
337 ± 4
176 ± 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 ± 0.04
0.13 ± 0.04
0.14 ± 0.04
0.45 ± 0.06
0.43 ± 0.05
0.44 ± 0.05
0.67 ± 0.05
0.64 ± 0.04
0.65 ± 0.04
0.74 ± 0.04
0.71 ± 0.03
0.72 ± 0.04
0.78 ± 0.04
0.75 ± 0.03
0.76 ± 0.03
0.81 ± 0.01
0.86 ± 0.01
1,474 ± 34
184 ± 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 ± 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 ± 0.01
0.03 ± 0.02
0.03 ± 0.02
0.18 ± 0.04
0.20 ± 0.04
0.19 ± 0.04
0.47 ± 0.05
0.51 ± 0.06
0.48 ± 0.05
0.62 ± 0.04
0.67 ± 0.04
0.63 ± 0.04
0.68 ± 0.03
0.73 ± 0.01
0.69 ± 0.02
0.75 ± 0.01
0.84 ± 0.01
278 ± 11
176 ± 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 α -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 ± 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
Regime
Experiment
training
distance
architecture
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
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 .
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
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
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 α -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 α -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: α -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: α -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.
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 (SNR=σs2/σn2 < 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 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.
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 (cryo-EM) determines the structures of proteins and macromolecular assemblies at near-atomic resolution, and the final 3D reconstruction depends on extracting a clean particle stack from noisy micrographs. This extraction decomposes into three sub-tasks, namely particle picking, contamination removal, and 2D class selection. Each of them, however, is trained and evaluated in isolation, and none is optimized for the reconstruction. We instead integrate the three sub-tasks into a single pipeline posed against downstream reconstruction quality. We instantiate the pipeline with a state-of-the-art component for each sub-task, CryoTransformer picking permissively, MicrographCleaner masking contamination, and CryoSift selecting 2D classes by a continuous quality score, and close the loop with a fine-tuning step that returns the surviving particles to the picker. The pipeline achieves a better 3D resolution than every picker we compare. We also show that the best 2D F1 is not the best resolution, so particle selection is better treated as one reconstruction-aware pipeline judged by the map it delivers.
Riku Itsuji, Yuanhao Wang, Xingjian Li +3
Keio University, Yokohama, Japan · Keio AI Research Center, Yokohama, Japan · Carnegie Mellon University, Pittsburgh, PA, USA
Cryo-electron microscopy (cryo-EM) enables single-particle analysis of biological macromolecules under strict low-dose imaging conditions, but the resulting micrographs often exhibit extremely low signal-to-noise ratios and weak particle visibility. Image denoising is therefore an important preprocessing step for downstream cryo-EM analysis, including particle picking, 2D classification, and 3D reconstruction. Existing cryo-EM denoising methods are commonly trained with pixel-wise or Noise2Noise-style objectives, which can improve visual quality but do not explicitly account for structural consistency required by downstream analysis. In this work, we propose a score-based denoising framework for cryo-EM that learns the clean-data score to recover particle signals while better preserving structural information. Building on this formulation, we further introduce a target-guided variant that incorporates reference-density guidance to stabilize score learning under weak and ambiguous signal conditions. Rather than simply amplifying particle-like responses, our framework better suppresses structured low-frequency background, which improves particle--background separability for downstream analysis. Experiments on multiple cryo-EM datasets show that our score-based methods consistently improve downstream particle picking and produce more structure-consistent 3D reconstructions. Experiments on multiple cryo-EM datasets show that our methods improve downstream particle picking and produce more structure-consistent reconstructions.