- The paper presents a novel training-free neuron tracing framework that combines heuristic spatial search with dimension-aware semantic verification via NeuroSAM 2.
- It achieves superior reconstruction fidelity with recall 0.753 and reduces manual proofreading time by over 33% compared to supervised methods.
- The integration into interactive human-in-loop workflows demonstrates scalable, high-precision mapping essential for advancing connectomics.
Probe-EM: Targeted Neuron Tracing via Training-Free Semantic Verification
Motivation and Context
The exponential growth in volume electron microscopy (EM) datasets, often spanning terabytes to petabytes, has profoundly advanced the field of connectomics, enabling circuit-to-whole-brain reconstructions [microns] [h01] [flywire]. Automated neuron segmentation and reconstruction pipelines, relying on deep learning and image processing techniques [mala] [ffn] [superhuman] [lsd] [embedded], are essential for initial annotation. Nonetheless, they suffer from persistent over-segmentation errors, requiring extensive manual proofreading—sometimes tens of person-years for a complete Drosophila brain [tenyear1] [tenyear2]. Existing frameworks for error correction predominately adopt rigid validation paradigms and heavily depend on large-scale labeled datasets, impeding scalability, flexibility, and generalizability in practical connectomics workflows.
Methodological Innovations
Heuristic Spatial Search (HSS)
Probe-EM introduces a training-free neuron tracing paradigm, formalized as a heuristic spatial search (HSS) process. This approach reframes tracing as a seed-driven retrieval task, utilizing skeleton-based geometric priors to guide local-to-global reconstruction. By operating on the spatial topology extracted from skeletonization, the HSS module prioritizes regions adjacent to endpoints, circumventing the inefficiency of exhaustive global scanning. The candidate set for further verification is restricted to proximity neighborhoods based on explicit geometric criteria, enabling targeted tracing of specific neuronal populations with minimal computational overhead.
Error propagation is mitigated by a topological pruning mechanism within HSS, which halts expansion upon detection of morphological anomalies such as excessive endpoint density or uncharacteristic complexity—typical signatures of glial misconnections or segmentation mergers. This leads to robust expansion of neuronal trees without divergence.
Dimension-Aware Semantic Verification (DASV) with NeuroSAM 2
Semantic verification is accomplished via the Dimension-Aware Semantic Verification (DASV) module anchored by NeuroSAM 2—a foundation model specifically fine-tuned for EM data [sam2] [segneuron]. Segment pairs are classified as intra-slice or inter-slice splits, activating specialized strategies:
- Planar Ensemble Consensus (PEC): For intra-slice splits, connectivity verification is conducted using geometric prompt sampling (anchor, probe, and rear points) and bidirectional cross-prediction. An ensemble consensus protocol (multiple randomized trials with strict dual-threshold criteria on overlap ratios) ensures only genuine continuity is accepted, effectively suppressing false positives from stochastic mask expansion.
- Axial Spatio-Temporal Propagation (ASP): For inter-slice splits, NeuroSAM 2 performs memory-augmented propagation across Z-axis slices, treating them as a pseudo-video series. Connectivity is validated via maximum volumetric occupancy over the propagation window, surpassing conventional static IoU matching in robustness to slice misalignment and morphological variance.
Integration and Workflow
Probe-EM seamlessly integrates HSS and DASV into the Neuroglancer interactive visualization platform, constructing a human-in-the-loop annotation and proofreading environment. This provides efficient support for iterative, expert-driven tracing and validation workflows, especially in large-scale EM datasets.
Experimental Results
On the mouse suprachiasmatic nucleus (SCN) dataset [SCN], with high-volume and high-resolution characteristics, Probe-EM demonstrates unequivocal superiority over supervised learning baselines:
- Axon Fascicle Tracing: Probe-EM with NeuroSAM 2 achieves recall 0.753, precision 0.626, and F1 0.595, outperforming voxel-wise CNN (F1 0.571), geometric PointNet (F1 0.288), and multimodal fusion (F1 0.436).
- Soma-seeded Tracing: Probe-EM achieves recall 0.705, precision 0.565, and F1 0.544, substantially exceeding baseline approaches.
Supervised methods require extensive, dataset-specific annotation to capture morphological diversity; Probe-EM, leveraging foundation model generalization, operates entirely training-free, maintaining high-fidelity and adaptability across novel domains.
Human Proofreading Efficiency
User studies involving expert annotators reveal Probe-EM reduces manual proofreading times by an average of 33.4%, with concurrent improvement in F1 measure (+6.8%). Integration into interactive workflows assists annotators in discovering missed branches and verifying continuity, mitigating labor burden and enhancing accuracy.
Ablation Studies
Core ablation experiments underscore the necessity of ASP (loss in recall when omitted), PEC (drop in F1), and geometric prompt sampling (degraded precision and recall without it). Evaluation of axial propagation window size identifies trade-offs between bridging gaps and accumulation of semantic drift, establishing optimal hyperparameters for maximal tracing fidelity.
Theoretical and Practical Implications
Probe-EM embodies a shift toward targeted, training-free approaches in large-scale connectomics, breaking reliance on rigid, annotation-dependent supervised paradigms. The skeleton-guided spatial search and foundation model-based verification generalize across diverse dataset characteristics, supporting flexible, user-driven tracing and rapid validation. This is pivotal for scaling connectomic mapping from local circuits to whole-brain volumes, and will catalyze further development of interactive human-in-the-loop systems and foundation model-driven analysis in biomedical imaging.
From a theoretical perspective, the rigid binary inference paradigm in traditional supervised frameworks is superseded by adaptive, topology-driven search paired with expressive zero-shot semantic verification. The systematic handling of intra- and inter-slice splits establishes new standards for robust neuron reconstruction.
Anticipated future developments include further expansion of foundation models for EM imagery, integration with multi-modal connectomic pipelines, and deployment in fully automated annotation loops with minimal human intervention. The methodological principles demonstrated by Probe-EM are applicable beyond neuroscience, in any domain requiring targeted, annotation-independent mapping of complex structures in large-scale volumetric data.
Conclusion
Probe-EM presents a training-free, targeted neuron tracing framework leveraging skeleton-guided heuristic search and dimension-aware semantic verification via NeuroSAM 2. The method eliminates computational redundancy, achieves state-of-the-art performance in reconstruction fidelity, reduces annotation time by one-third, and supports scalable, high-precision connectomics workflows. Its adoption will substantially impact automated neuron mapping, large-volume proofreading, and the advancement of foundation model integration in biomedical research.