CORAL: A Benchmark for Structure-aware and Brain-wide Neuron Reconstruction in Light Microscopy

arXiv:2608.30768 · cs.CV · Submitted 2026-08-31 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "CORAL: A Benchmark for Structure-aware and Brain-wide Neuron Reconstruction in Light Microscopy".

Jane: Here's the gist of the paper. Automatic neuron reconstruction from light microscopy images is a central problem in computational neuroanatomy where progress is limited by structural accuracy and scalability to whole-brain scales CORAL establishes two progressive tasks: blocklevel reconstruction and brainwide reconstruction To account for topological correctness beyond geometric distance similarity, a structureaware metric based on fiber prediction is introduced A brainwide neuron tracing framework extends arbitrary local reconstruction methods to the wholebrain scale through an iterative localto global process Using this benchmark, the results underscore the importance of structureaware evaluation and the need for more robust methods for complete neuron reconstruction.

Tom: Benchmark Construction and Tasks CORAL is built on a highquality wholebrain fMOST dataset with carefully curated neuronannotations It establishes two progressive tasks: blocklevel reconstruction, which evaluates reconstruction methods under limited spatialcontext, and brainwide reconstruction, which assesses complete neuronreconstruction at the wholebrain scale.

Jane: The blocklevel task involves inputting a 300x300x300 image cube and testing reconstruction methods on local cubes The brainwide task involves inputting the wholebrain volume and reconstructing the complete neuron across the brainwide using a wholebrain tracing framework.

Tom: Structureaware Evaluation Metrics To better assess reconstruction quality, structureaware metrics based on keypointprediction and fiberprediction are introduced These include the Keypoint metric which measures how accurately an algorithm predicts critical locations like bifurcations and leafs The Fiber metric directly measures the reconstruction performance of each neurite by calculating fiber intersectionoverunion (FIoU).

Paper summary: Tom: So, we're looking at this paper called "CORAL: A Benchmark for Structure-aware and Brain-wide Neuron Reconstruction in Light Microscopy". What's the big picture here?

Jane: Basically, it addresses a huge problem in computational neuroanatomy: automatically rebuilding neurons from light microscopy images. The paper claims that while some methods do well on small local blocks, we don't really know if that translates to reconstructing the whole brain accurately and at scale.

Lu: The core thesis is that existing benchmarks only test local image blocks or isolated structures, which misses the central challenge of whole-brain reconstruction—specifically, getting global topological correctness and preserving long-range continuity.

Meng: So, what's the claim? What exactly is CORAL doing to fix this gap?

Tom: They introduce two progressive tasks. First is blocklevel reconstruction, testing methods on small cubes, like three hundred by three hundred by three hundred image volumes. Then they have brainwide reconstruction, which tests the complete neuron across the whole brain using a wholebrain tracing framework.

Jane: That means they're testing how well methods handle both limited spatial context and the massive scale of the entire brain volume simultaneously. It seems like a necessary step up from just looking at small pieces.

Lalam: And to make sure we measure this better than just simple geometric distance, they introduce structureaware metrics based on fiber prediction. They have a Keypoint metric for critical locations like bifurcations and leaves, and a Fiber metric that measures performance by calculating fiber intersection over union, or FIoU.

Tom: The abstract says the Fiber metric is defined when its FIoU reaches a threshold eta. It’s not just about whether the prediction is spatially close; it’s about how accurately it traces the actual neurite structure itself.

Jane: That sounds crucial because, as the paper points out, many existing metrics are primarily geometric rather than structureaware, meaning they don't really capture if the reconstruction respects the underlying tree structure of a neuron.

Lu: The blocklevel findings show that segmentation and reconstruction methods are coupled; you can't optimize one without thinking about the other because small local errors can lead to big topological failures globally.

Meng: That sounds like a nightmare for deployment, actually. If you optimize for one part and it breaks the whole structure downstream, your model is brittle at scale.

Tom: Exactly what Meng means. Then they move to brainwide reconstruction and highlight three main error types: break error, merge error I which is a same neuron crossover, and merge error II which is an other neuron crossover.

Jane: That decomposition of errors—break versus merge—helps pinpoint exactly where the whole-brain tracing framework needs to improve its continuity preservation.

Lalam: Looking at the comparative performance insights, they found that learning based endtoend reconstruction algorithms have potential but often struggle with reconstructing longrange neurites, especially axonal structures.

Tom: They actually mention Kimimaro achieving better reconstruction quality for dendrites, but no method showed a significant advantage in axon reconstruction. So it's a split result there.

Jane: It seems the current methods are reasonably good at reconstructing dendrites, but when you look at those long axonal structures spanning across the whole brain, things get much harder for these existing models to handle structurally correctly.

Lu: And they admit a major limitation about this benchmark itself. They built it from a limited number of fully annotated wholebrain neurons, so it doesn't cover the full diversity of morphologies or imaging conditions yet.

Meng: That's a big caveat for anyone trying to use this as a production standard right now. It hasn't been tested on real, messy traffic data yet, which is where things usually break.

Tom: So, what does this all mean for the field? The conclusion of CORAL says it provides a standardized platform for developing and testing brainwide neuron reconstruction methods.

Jane: It sets up a rigorous testbed for evaluating these methods before they try to tackle the whole brain on actual imaging data. It forces researchers to look at structure awareness, not just pixel overlap.

Lu: The implication is that future work needs to focus on making those deep learning models robust enough to handle the propagation of small local structural errors across vast distances.

Meng: From my side, it means if we want this tech in the clinic or for large-scale analysis, we need methods that are explicitly tested against these kinds of long-range topological failures.

Tom: So, overall, CORAL is important because it pushes the community toward structureaware evaluation when rebuilding neurons at a whole brain scale. We'll keep an eye on how those metrics perform in future work.

Conclusion: Tom: So we’re wrapping up our look at CORAL, this paper by its authors on structure-aware and brain-wide neuron reconstruction from light microscopy images.

Jane: It really boils down to a standard test for figuring out if we can rebuild neurons across the entire brain accurately.

Tom: Yeah, the title itself is pretty direct, it’s not trying to be flashy; it just tells you what they built—a benchmark for structure aware and brain wide neuron reconstruction.

Jane: And what they did was set up two main tests: a smaller block level reconstruction and a full brain wide reconstruction test.

Lu: What I found interesting is how they moved from that small local context test to the whole brain tracing framework, which is where the real challenge lies.

Meng: From my side, it means if you're trying to deploy something for real applications, you need to know if it handles that scale and those errors across the entire structure.

Lalam: The big implication is that this benchmark gives researchers a standardized platform to actually test their methods before they try and tackle the whole brain on messy, real imaging data.

Tom: Exactly. It’s like giving everyone a common yardstick for how good their reconstruction method really is when you look at things at the whole brain level.

Jane: So, it’s not just about getting a high score on one test; it’s about understanding how well that method handles the structural details across the entire volume.

Tom: Right. It shows that geometry alone isn't enough anymore because small local mistakes can cause big problems globally, and this benchmark tests for that.

Jane: And they pointed out specific error types, like break errors and merge errors, which tells you exactly what kind of continuity the method is failing on at scale.

Lu: What’s next is seeing how these structure aware metrics actually perform when you compare different reconstruction algorithms against this rigorous test set.

Meng: That comparison will tell us if the learning based end to end algorithms they tested are actually robust enough for a production environment or if they still struggle with long range neurites.

Lalam: I'm curious to see the numbers on how those models handle those long range connections, because that’s where I’ve seen them falter before.

Department of Computer Science and Technology, Institute for Artificial Intelligence, Tsinghua University · Tsinghua Laboratory of Brain and Intelligence, Tsinghua University Laboratory of Brain and Intelligence · School of Basic Medical Sciences, Tsinghua University School of Basic Medical Sciences · Tsinghua-Peking Center for Life Sciences, Tsinghua-Peking Center for Life Sciences · Chinese Institute for Brain Research

cs.CV

Submitted: 2026-08-31

Updated: 2026-11-12

Code: https://github.com/yang-ze-kang/CORAL

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 90/100

The gist: Automatic neuron reconstruction from light microscopy images is a central problem in computational neuroanatomy where progress is limited by structural accuracy and scalability to whole-brain scales…

Key concepts

Blocklevel Reconstruction
This task tests how well a method can reconstruct neurons within small, localized areas of an image cube. It assesses performance under limited spatial context and helps determine if local errors can lead to global topological failures in the final reconstruction.
Brainwide Reconstruction
This involves reconstructing the entire neuron across a whole-brain volume using an iterative local-to-global tracing framework. It challenges methods to handle structural errors that propagate from small local mistakes into large, whole-brain inaccuracies.
Structureaware Metrics
These are new evaluation tools used to judge reconstruction quality by analyzing predicted keypoints (like bifurcations) and fiber predictions. They go beyond basic geometry by assessing how accurately the algorithm captures the actual physical structure of neurites.
Fiber IntersectionoverUnion (FIoU)
This metric directly measures a neuron's reconstruction performance by calculating how well the reconstructed fibers overlap with the ground truth fibers. A high FIoU score indicates that individual neurites have been reconstructed with high fidelity.

Terminology

Summary

Automatic neuron reconstruction from light microscopy images is a central problem in computational neuroanatomy where progress is limited by structural accuracy and scalability to whole-brain scales CORAL establishes two progressive tasks: blocklevel reconstruction and brainwide reconstruction To account for topological correctness beyond geometric distance similarity, a structureaware metric based on fiber prediction is introduced A brainwide neuron tracing framework extends arbitrary local reconstruction methods to the wholebrain scale through an iterative localto global process Using this benchmark, the results underscore the importance of structureaware evaluation and the need for more robust methods for complete neuron reconstruction. p. 0

Where the paper says it

To account for topological correctness beyond geometric distance similarity, we introduce a structure-aware metric based on fiber prediction. To further achieve complete neuron reconstruction across the entire brain, we develop a brain-wide neuron tracing framework that extends arbitrary local reconstruction methods to the whole-brain scale through an iterative localto-global process.

Benchmark Construction and Tasks

CORAL is built on a highquality wholebrain fMOST dataset with carefully curated neuronannotations It establishes two progressive tasks: blocklevel reconstruction, which evaluates reconstruction methods under limited spatialcontext, and brainwide reconstruction, which assesses complete neuronreconstruction at the wholebrain scale. The blocklevel task involves inputting a 300x300x300 image cube and testing reconstruction methods on local cubes The brainwide task involves inputting the wholebrain volume and reconstructing the complete neuron across the brainwide using a wholebrain tracing framework. p. 4

Where the paper says it

Moreover, we design the Fiber metric, a more comprehensive yet simple and intuitive metric, to evaluate the performance of different algorithms (Sec 4. (a) Benchmark Tasks Task 1: Block-Level Reconstruction (local morpgology) Input: 300×300×300 image cube Training 907 cubes Testing 1021 cubes local cube SWC trees Task 2: Brain-wide Reconstruction (global topology) Input: whole-brain volume Whole-Brain Tracing Framework Test set (right hemisphere): 14 neurons Start point: soma location (b) Whole-Brain Tracing Framework Heuristic Search candidate frontiers Local Tracing local trees Growing new subtree Global Tree Soma location local trees cache topology check Update Init (c) Evaluation Metrics Optimal matching matched unmatched FN TP d < τ FIoU ≥ η d ≥ τ FN FP FIoU < η ground truth prediction leaf dendrite Keypoint: Fiber: axon bifurcation Figure 3: Overview of the CORAL benchmark.

Structureaware Evaluation Metrics

To better assess reconstruction quality, structureaware metrics based on keypointprediction and fiberprediction are introduced These include the Keypoint metric which measures how accurately an algorithm predicts critical locations like bifurcations and leafs The Fiber metric directly measures the reconstruction performance of each neurite by calculating fiber intersectionoverunion (FIoU). For the Fiber metric, TPfiber is defined when its FIoU reaches a threshold η. p. 6

Where the paper says it

The fiber pair is considered correctly identified only if its FIoU exceeds the threshold η are considered true positives (TP): TP = X (fg,fp)∈M I FIoU(fg, fp) ≥ η , Conversely, matched fiber pairs with an FIoU below the threshold η are also counted as false positives and false negatives. Then we have the definitions of precision, recall, and F1 for Fiber metric: Precision = TP |Fpred| , Recall = TP |Fgt| , F1 = 2 ∗ Precision ∗ Recall Precision + Recall .

Blocklevel Reconstruction Findings

The experiments in the blocklevel task revealed that segmentation method and the reconstructionmethod are coupled, making algorithm optimization more challenging. Geometrybased metrics are insufficient to characterize finalreconstructionquality because minorlocalerrors can lead to globaltopologicalfailures. p. 1

Where the paper says it

First, in two-stage reconstruction methods, the segmentation method and the reconstruction method are coupled, making algorithm optimization more challenging. Second, geometry-based metrics are insufficient to characterize final reconstruction quality: minor local errors can lead to global topological failures, whereas structure-aware keypoint- and fiber-based metrics provide a substantially more comprehensive assessment of reconstruction quality.

Structureaware keypointandfiberbasedmetrics provide a substantiallymorecomprehensiveassessment of reconstructionquality. p. 0
Where the paper says it

To further achieve complete neuron reconstruction across the entire brain, we develop a brain-wide neuron tracing framework that extends arbitrary local reconstruction methods to the whole-brain scale through an iterative localto-global process. Using this benchmark, we provide the first structure-aware comparison of mainstream methods for local neuron reconstruction and further evaluate their performance in brain-wide reconstruction.

Brainwide Reconstruction Challenges

The brainwide reconstruction task highlighted that currentmethods remain far from achieving structurallycorrectreconstruction and fullyautomatedbrainwide reconstruction. Evensmalllocalstructuralerrors can propagate into largescaleerrors across thewholebrain, imposing higherrequirements on automatedreconstructionmethods. p. 1

Where the paper says it

Third, current methods remain far from achieving structurally correct reconstruction and fully automated brain-wide reconstruction. Even small local structural errors can propagate into large-scale errors across the whole brain, imposing higher requirements on automated reconstruction methods.

Errors in brainwide neuron reconstruction mainly fall into three categories: breakerror, mergeerror I (sameneuron crossover), and mergeerror II (otherneuron crossover). p. 8
Where the paper says it

Figure 7 shows the representative errors that occur in brain-wide reconstruction. As illustrated in the figure, errors in brain-wide neuron reconstruction mainly fall into three categories.

Comparative Performance Insights

The comparison of reconstruction methods showed that learningbased endtoendreconstructionalgorithms have greatpotential but often show limited ability to reconstruct longrange neurites. The results indicate that while existingmethods can reconstruct dendrites reasonably well, they still struggle to reconstructionlongrange axonal structures. p. 8

Where the paper says it

the substantial gap that still exists toward fully automated reconstruction. While existing methods can reconstruct dendrites reasonably well, they still struggle to reconstruction long-range axonal structures, highlighting the challenge of long-distance continuity in brain-wide reconstruction.

Kimimaro achieves betterreconstructionquality fordendrites whereas no reconstructionmethod demonstrates a significant advantage in axonreconstruction. p. 8
Where the paper says it

While existing methods can reconstruct dendrites reasonably well, they still struggle to reconstruction long-range axonal structures, highlighting the challenge of long-distance continuity in brain-wide reconstruction. Kimimaro achieves better reconstruction quality for dendrites, whereas no reconstruction method demonstrates a significant advantage in axon reconstruction.

Limitations of the Benchmark

Despite its strengths, CORAL has several limitations. The currentbenchmark is built from a limited number of fullyannotated wholebrainneurons and therefore does not yet cover the fulldiversity of neuronalmorphologies and imagingconditions. p. 0

Where the paper says it

We present CORAL, the first benchmark for structure-aware evaluation of automatic neuron reconstruction from light microscopy images at both local and whole-brain scales. Built on a high-quality whole-brain fMOST dataset with carefully curated annotations, CORAL establishes two progressive tasks: blocklevel reconstruction, which evaluates reconstruction methods under limited spatial context, and brain-wide reconstruction, which assesses complete neuron reconstruction at the whole-brain scale.

Moreover, alldata are derivedfrom mice and thebenchmark has not yet been evaluated on other species. p. 4
Where the paper says it

The diverse block-level images, together with the WBTF, enable reconstruction algorithms to be developed and trained on block images and then deployed for brain-wide tracing inference. Moreover, we design the Fiber metric, a more comprehensive yet simple and intuitive metric, to evaluate the performance of different algorithms (Sec 4.

Conclusion

CORAL provides a standardizedplatform for developing and testing brainwide neuronreconstructionmethods. The benchmark serves as a rigorous testbed for developing reconstructionmethods and evaluating brainwide neuronreconstruction. p. 1

Where the paper says it

Even small local structural errors can propagate into large-scale errors across the whole brain, imposing higher requirements on automated reconstruction methods. These findings demonstrated that CORAL provides a rigorous testbed for developing reconstruction methods and evaluating brain-wide neuron reconstruction.

The authors hope CORAL will serve as a efficientplatform for futureresearchonbrainwideneuronreconstruction. p. 9
Where the paper says it

We conducted extensive experiments on this benchmark and identified several key insights that may help guide future research in automated neuron reconstruction. We hope CORAL will serve as a efficient platform for future research on brain-wide neuron reconstruction.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Develop an iterative local to global tracing framework that extends arbitrary local reconstruction methods to whole-brain scale through a process involving heuristic search, local tracing, and topology-aware growing. This allows reconstruction algorithms to be developed on block images and then deployed for brain-wide tracing inference while ensuring the preservation of continuity, branching structure, and neuronal identity over long distances.

  2. Implement a structure aware metric based on fiber prediction to evaluate topological correctness beyond geometric distance similarity. This metric explicitly measures the recovery of bifurcation/termination events and root-to-leaf neuronal fibers which is crucial because root-to-leaf fibers are often more critical than local geometric similarity for assessing whether a reconstruction preserves the correct neuronal organization.

  3. Integrate keypoint metrics that separately evaluate bifurcations and leaves to assess structural correctness. This addresses the limitation where minor local errors can lead to global topological failures by measuring how accurately an algorithm predicts critical locations like bifurcations and leaves, as these affect structure through different patterns.

  4. Utilize the Fiber metric to directly measure the reconstruction performance of each neurite by calculating fiber intersection over union (FIoU). This provides a comprehensive evaluation by defining FIoU as Overlap(fg, fp) Length(fg) + Length(fp) − Overlap(fg, fp) and identifying true positives where its FIoU exceeds the threshold eta.

  5. Employ a whole-brain tracing framework that uses an iterative expansion process guided by structurally informative nodes like leafs and margin nodes to progressively explore the whole-brain volume. This strategy balances exploration and reliability by employing a breadth first search strategy instead of depth first search, which encourages more uniform spatial exploration and reduces the risk that early merging errors propagate along a single incorrect branch.

  6. Train segmentation models using a unified protocol involving random cropping and sliding window inference on 300x300x300 cubes to ensure fair comparison across different reconstruction methods. This standardization prevents bias by ensuring all segmentation models are trained under a unified protocol for fair comparison.

Abstract

Automatic neuron reconstruction from light microscopy images is a central problem in computational neuroanatomy. While recent methods have achieved encouraging results on local image blocks, it remains unclear whether such progress translates to reconstruction that is both structurally accurate and scalable to the whole-brain scale. We present CORAL, the first benchmark for structure-aware evaluation of automatic neuron reconstruction from light microscopy images at both local and whole-brain scales. Built on a high-quality whole-brain fMOST dataset with carefully curated annotations, CORAL establishes two progressive tasks: block-level reconstruction, which evaluates reconstruction methods under limited spatial context, and brain-wide reconstruction, which assesses complete neuron reconstruction at the whole-brain scale. To account for topological correctness beyond geometric distance similarity, we introduce a structure-aware metric based on fiber prediction. To further achieve complete neuron reconstruction across the entire brain, we develop a brain-wide neuron tracing framework that extends arbitrary local reconstruction methods to the whole-brain scale through an iterative local-to-global process. Using this benchmark, we provide the first structure-aware comparison of mainstream methods for local neuron reconstruction and further evaluate their performance in brain-wide reconstruction. Our results underscore the importance of structure-aware evaluation and the need for more robust methods for complete neuron reconstruction.

Sources

Related papers