acceptodds
Under review as a conference paper at ICLR 2027

End-to-End Neuron Tree Tracing with Historical Trajectories

Abstract

Reconstructing neuronal trees from 3-D microscopy requires recovering branching connectivity from weak and ambiguous local image evidence. A false junction or missed branch can corrupt an entire downstream arbor. We introduce TreeTracer, an end-to-end autoregressive model that expands a point-and-direction prompt into a neuronal subtree. An elegant end-to-end design enables the joint optimization of neuronal tree geometry and topology predictions through gradient-based learning. A causal decoder encodes historical trajectory features for more accurate topology prediction and employs local–global cross-attention to extract visual evidence more efficiently. On block-level subtree reconstruction, TreeTracer achieved Fiber F1 scores of 58.4% on CWMBS and 66.4% on CORAL, compared with and for the strongest competing methods, respectively. On brain-wide neuron reconstruction, it reached 41.6%, compared with 35.1% for the strongest competing method.

Primary area:
applications to neuroscience & cognitive science
Keywords:
neuron resonstruction, computer vision, autogressive tracing, neuron segmentation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_gcdROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes an end-to-end 3D neuron reconstruction algorithm for optical neuron images. It formulates 3D neuron reconstruction as starting from a voxel crop and a point-direction hint (p0,d0)(p_0, d_0), and at each step predicting the direction to the next node, the node type, and a bounded position residual. Data results in the paper show that the proposed method achieves favorable reconstruction results on CWMBS and CORAL.

Critical strengths

The guided autoregressive subtree generation employed in the paper constitutes a meaningful contribution. In the process of neuron reconstruction, branch detection is transformed from a post-hoc connection rule into a supervised, in-the-loop decision-making process, and the per-node targets (type + outgoing direction + residual) can be transferred to other trackers.

Critical weaknesses

The methods section of this paper is difficult to understand, especially Section 3.2. Many of the mathematical symbols lack rigorous definitions, and the relationships between symbols are not sufficiently explained, making it hard for me to accurately understand the method proposed by the authors. How the method in Section 3.2 is used in conjunction with the model architecture in Section 3.3 is also unclear, as there is no high-level description of this.

Questions for authors

Q1. Why does the experimental section of the paper lack experiments on the BigNeuron dataset, which is commonly used for neuron reconstruction?

Q2. What is the prompt inference cost of the sandwich variant of TreeTracer? Table 13 only provides the results for the vanilla version.

9 comments
AuthorsOpenReview ↗

Part [1/2] Response to W1: clarity of Sections 3.2 and 3.3

Thank you for the specific feedback on presentation. We agree that Section 3.2 introduces many symbols without a high-level picture and that its link to Section 3.3 is not stated. In the revision we will open Section 3 with the overview below and a step-by-step algorithm, add a notation table, and remove overloaded symbols.

Overview. TreeTracer reconstructs the subtree selected by a prompt (p0,d0)(\mathbf{p}_0,\mathbf{d}_0), a start point and an initial direction (Sec. 3.1). It keeps a set of active paths, each ending at a current node tt, and repeats the following for every active path.

  1. State. The path's history Ht\mathcal{H}_t (Eq. 4) lists, for every node jj from the root to tt, its position p^j\widehat{\mathbf{p}}_j, the direction u^j\widehat{\mathbf{u}}_j along which this path leaves it (one of K=192K=192 unit vectors in a fixed codebook U\mathcal{U}), and its type c^j\widehat{c}_j: continuation C, branch B, or leaf L.
  2. Network fθf_\theta (Sec. 3.3). The image encoder turns the crop into dense features F+\mathbf{F}^{+} and a few global tokens G\mathbf{G}. Each history entry becomes one token. The causal decoder attends to the earlier tokens, to F+\mathbf{F}^{+} around the coarse next position p‾t+1\overline{\mathbf{p}}_{t+1} from step 3, and to G\mathbf{G}. Three heads then output, for the next node, a distribution over the KK directions π^t+1d\widehat{\boldsymbol{\pi}}^{d}_{t+1}, a distribution over the three types π^t+1c\widehat{\boldsymbol{\pi}}^{c}_{t+1}, and an offset o~t+1\widetilde{\mathbf{o}}_{t+1} (Eq. 5).
  3. Advance (Eq. 6). The next node is placed one step of length ss along the current direction, p‾t+1=p^t+su^t\overline{\mathbf{p}}_{t+1}=\widehat{\mathbf{p}}_t+s\widehat{\mathbf{u}}_t, and then shifted by the bounded offset δt+1=sρtanh⁡(o~t+1)\boldsymbol{\delta}_{t+1}=s\rho\tanh(\widetilde{\mathbf{o}}_{t+1}).
  4. Expand (Eqs. 7–8). The predicted type sets the number of children. C gives one child along the most probable direction; B gives the most probable directions whose total probability reaches α\alpha, at most two; L ends the path. Each child starts a new active path whose history is the current history plus the new node and that child's direction.
  5. Output (Eq. 10). When no path remains active, the union of all paths is the predicted subtree.

Only step 2 has learned parameters; steps 3–5 are fixed rules that turn the network outputs into a tree. Training (Sec. 3.4) supervises the three outputs of step 2 along annotated paths. For images larger than the 32×96×9632\times96\times96 crop, a path that reaches the crop boundary continues in a new crop together with its recent history (Appendix A.7, Algorithm 2).

Notation table (to be added to Section 3)

SymbolMeaningWhere it comes from
(p0,d0)(\mathbf{p}_0,\mathbf{d}_0)prompt: start point and initial directioninput (Sec. 3.1)
U=ukk=1K\mathcal{U}=\\{\mathbf{u}_k\\}_{k=1}^{K}direction codebook, K=192K=192 unit vectorsfixed (Appendix A.1)
Ht\mathcal{H}_thistory of one path, from the root to node ttgrows in step 4
p^j, u^j, c^j\widehat{\mathbf{p}}_j,\ \widehat{\mathbf{u}}_j,\ \widehat{c}_jposition, outgoing codeword, and type of node jjentries of Ht\mathcal{H}_t
π^t+1d, π^t+1c\widehat{\boldsymbol{\pi}}^{d}_{t+1},\ \widehat{\boldsymbol{\pi}}^{c}_{t+1}predicted direction and type distributionsheads gdg_d, gcg_c (Sec. 3.3)
o~t+1, δt+1\widetilde{\mathbf{o}}_{t+1},\ \boldsymbol{\delta}_{t+1}raw and bounded position offsethead gog_o; Eq. 6
s, ρs,\ \rhostep size and offset rangehyperparameters
p‾t+1\overline{\mathbf{p}}_{t+1}coarse next position; center of local attentionEq. 6; decoder
α, Dt+1(α)\alpha,\ \mathcal{D}^{(\alpha)}_{t+1}cumulative threshold and candidate child directionsEq. 7
At+1\mathcal{A}_{t+1}directions of the children createdEq. 8
F+, G, ht\mathbf{F}^{+},\ \mathbf{G},\ \mathbf{h}_tdense features, global tokens, decoder outputSec. 3.3
AuthorsOpenReview ↗

Part [2/2] Response to Q1 & Q2

Response to Q1: BigNeuron

We did not evaluate on BigNeuron, and we will list the single-domain evaluation as a limitation. TreeTracer targets brain-wide reconstruction of long-projecting neurons in mouse whole-brain light microscopy. CWMBS and CORAL come from this setting (blocks cropped from whole mouse-brain images; CORAL is fMOST) and provide both training subtrees and test blocks. The brain-wide experiment then traces 14 complete neurons through volumes exceeding 10410^4 voxels per side, which tests continuation across many windows. BigNeuron's Gold166 covers many species and imaging modalities in small image stacks. It is a good test of cross-domain generalization, but each modality has only a few annotated stacks for training a learned tracer, and the small stacks do not exercise long-range continuation.

Response to Q2: inference cost of the sandwich variant

Sandwich runs the same tracer from the supplied prompt and from up to 32 accepted automatic seeds, each traced in two directions, and then fuses and refines the traces (Algorithm 1). The table extends Table 13 (CWMBS test set, 483 prompts from 37 images, RTX 3090, same timing scope; the Trexplorer and vanilla rows match Part 1 of our response to Reviewer c9bY). Generated length rewards spurious paths, so we report correctly traced length per second.

MethodCorrect length (voxels)Time (s)Time per prompt (s)Correct length per second ↑
APP2(SegMamba)23,018.81467.400.9749.25
Kimimaro(SegMamba)15,824.06240.430.5065.82
neuTube(SegMamba)29,885.47751.381.5639.77
APP2(GBP-Net)30,974.881,396.512.8922.18
Kimimaro(GBP-Net)59,672.82954.421.9862.52
neuTube(GBP-Net)51,990.332,038.374.2225.51
Trexplorer18,578.935,630.7611.663.30
Trexplorer Super50,112.713,293.986.8215.21
TreeTracer (vanilla)50,465.62630.061.3080.10
TreeTracer (sandwich)62,831.505,646.7111.6911.13

Sandwich is the most expensive configuration: 11.7 s per prompt, 9.0× vanilla and 5.9× Kimimaro(GBP-Net), and about the same as Trexplorer. The extra time is spent tracing the automatic seeds and fusing and refining the traces. In return it traces the most correct length, 5.3% more than Kimimaro(GBP-Net) and 24.5% more than vanilla, and reaches the highest Fiber F1. Vanilla is the fastest per correctly traced voxel. We will add these rows to Table 13.

Dear Reviewer,

Thank you again for your thoughtful feedback. We would like to follow up on our two-part response to your comments.

In Part [1/2], we provided a step-by-step overview connecting the tree-generation procedure in Section 3.2 to the network architecture in Section 3.3, together with a notation table. In Part [2/2], we explained our dataset selection, and reported the inference cost of TreeTracer (sandwich).

We would greatly appreciate your feedback on whether these responses address your concerns, particularly regarding the clarity of the method and the connection between Sections 3.2 and 3.3. If any aspects remain unclear, we would be happy to provide further explanations.

Thank you for your time and consideration.

Best regards,

The Authors

Reviewer_1sagOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces TreeTracer, an autoregressive model designed to reconstruct neuronal tree structures from 3D microscopy images, guided by point-and-direction prompts. The model employs a multi-scale image encoder and a causal decoder, integrating local and global image features with trajectory tokens to predict positional offsets, outgoing directions, and node types. The authors evaluate both a vanilla version of the tracer and a sandwich pipeline—the latter augmenting the base model with image-based seed generation and structural refinement. Experimental results on two local datasets demonstrate that the sandwich pipeline outperforms current state-of-the-art baselines. The brain-wide experiment also shows a higher overall Fiber F1, with the main improvement on axons.

Critical strengths
  1. Promising reconstruction performance, particularly on non-branching targets. Vanilla substantially outperforms the strongest reported baselines on non-branching subtrees in both datasets. The whole-brain experiments provide compelling evidence beyond local subtree analysis: among all evaluated methods, TreeTracer achieves the highest overall Fiber F1 and axonal Fiber F1scores, although Kimimaro performs better for dendrites.
  2. A reasonably transparent implementation and a strong positional-refinement ablation. The appendix specifies the multi-seed pipeline, window continuation, data splits, and Fiber evaluation. Combining offset prediction with perturbation-based training yields substantial improvements in reconstruction, validating the effectiveness of the model’s correction component.
Critical weaknesses
  1. The branch-event claim lacks direct evaluation. Fiber F1 measures recovery of complete paths and their connectivity, but branch-subset scores also depend on path localization, termination, and coverage. However, the paper does not report bifurcation detection precision/recall or a tolerance-based Bifurcation-F1. On CORAL, even sandwich is below the strongest two-stage method on branching subtrees (59.1% vs. 63.1%); on CWMBS, the two are nearly tied. The overall gains therefore do not establish superior branch-event prediction. Appendix G's explanation that annotation noise is the main cause of low bifurcation recall is not directly quantified.

  2. The source of the overall performance improvement remains unclear. Vanilla exceeds the strongest overall baseline by only 0.9 percentage points on each local dataset; sandwich gains a further 5.7 and 4.5 points after adding seeds, trace fusion, connection, and pruning. In the attention ablation, local image attention alone nearly matches the full decoder, while the independent benefit of historical self-attention is uncertain. The ablation study regarding direction-tokens fails to isolate the contribution of long-history aggregation. In contrast, the ablation study on positional-refinement yields much clearer results. A comparison with matched seed, postprocessing, and compute budgets, plus a controlled history ablation, would clarify the contribution of learned subtree expansion.

Questions for authors
  1. Could you report bifurcation and leaf Keypoint precision, recall, and F1 on the same CORAL predictions, with a stated localization tolerance, for vanilla, sandwich, and the strongest two-stage baselines? Furthermore, providing a detailed breakdown of missed branches, false junctions, and premature stops would help clarify which specific topological errors are mitigated by the node-type head.
  2. How was each baseline's output converted into the prompted, directed target subtree, including root snapping, direction selection, stopping, and pruning? Can a strong single-path tracer, ideally including DeepBranchTracer if feasible, be compared under the same seed, fusion, connection, pruning, and compute budgets as sandwich?
  3. Can you isolate historical aggregation from the current direction token using a history-length control? For the brain-wide comparison, please provide paired per-neuron results or a neuron-level confidence interval; micro-averaging can give unequal weight to differently sized neurons. It is currently unclear whether the improvement is distributed across the majority of cells or concentrated within a small number of large, highly branched cells.
5 comments
AuthorsOpenReview ↗

Part [1/3] Response to Q1 & W1

Thank you for the careful review.

Regarding the statement, “Fiber F1 measures recovery of complete paths and their connectivity, but branch-subset scores also depend on path localization, termination, and coverage,” could you please clarify the intended distinction? Our understanding is that Fiber F1 inherently depends on path localization, termination, and coverage through FIoU-based matching.

For keypoint metrics (bifurcation and leaf points) and topology error metrics, please refer to Parts 3–4 of our response to Reviewer c9bY (Q3).

Appendix G. We visually inspected image regions around 100 randomly selected branch points in the training set and found that more than half of the annotations deviated from the ideal branch point locations.

AuthorsOpenReview ↗

Part [2/3] Response to Q2: subtree conversion, matched comparison, and DeepBranchTracer

Conversion to the prompted, directed subtree. All reconstruction outputs, including those of TreeTracer (sandwich), undergo the same directed-subtree extraction procedure described in Appendix A.4. Given the prompt (p0,d0)(\mathbf{p}_0,\mathbf{d}_0), we identify the nearest reconstructed node of p0\mathbf{p}_0. We reroot its component at this node, anchor it to p0\mathbf{p}_0, and retain the root child whose outgoing direction best aligns with d0\mathbf{d}_0, together with all its descendants. We will make this shared procedure explicit and separately document each pipeline’s stopping, pruning, and computational settings.

Matched seeds, fusion, pruning, and compute. All baselines run with their native initialization and post-processing: SPE-DNR and NETracer seed from local maxima of their predicted maps; DRL traces one path from the prompt point and direction; APP2 and neuTube use their own seed detection, tracing, connection, and pruning; and Kimimaro skeletonizes the whole mask. For the two-stage pipelines, "the same seeds and post-processing" is not well defined. Kimimaro takes no seeds, APP2 and neuTube's seeding and pruning are part of the algorithm, so replacing them would produce hybrids that are no longer the published methods. These pipelines also already cover the whole block, since segmentation and skeletonization process all foreground; the additional seeds give a prompt-driven tracer comparable coverage. Our seed proposal is adapted from neuTube's strategy, so neuTube already runs with comparable seeding.

Strong tracing baselines and DeepBranchTracer. Our evaluation already includes NETracer, a subsequent tracing method from the same research group as DeepBranchTracer. We have also added comparisons with Trexplorer and Trexplorer Super; please see Part 1 of our response to Reviewer c9bY (Q1 & W1). DeepBranchTracer needs node-radius supervision. Its authors told us that their radius targets were estimated with neuTube; these estimates were noisy on our data, and we could not train the model reliably.

AuthorsOpenReview ↗

Part [3/3] Response to Q3 & W2: history control, source of the gain, and per-neuron results

History beyond the current direction token. We ran a history-length control; the setup and table are in Part 5 of our response to Reviewer c9bY (Q5). With H = 1 the model keeps the immediate predecessor token, including the current heading, but no earlier history, while the weights and image features stay fixed. Longer history improves node-type prediction (macro-F1 +9.18 and +13.99 points from H = 1 to H = 32 on CWMBS and CORAL; paired 95% CIs exclude zero) but not direction prediction (+1.92 and +1.12 points; CIs include zero). We therefore claim a history benefit for topology decisions, not for geometry.

Source of the overall gain. With paired source-image cluster bootstrap CIs (10,000 resamples), vanilla's margin over Kimimaro(GBP-Net) is +0.95 points on CWMBS ([−5.16, 6.69]) and +0.85 on CORAL ([−3.30, 5.03]), which is not significant; the sandwich margins of +6.63 ([2.40, 10.54]) and +5.40 ([1.38, 9.53]) are. The further sandwich gain is concentrated in branching subtrees, where branch-subset Fiber F1 rises by 13.1 points on both datasets (paired 95% CIs [7.5, 18.9] and [7.5, 18.7]) and recall rises from 27.3/34.0 to 46.1/58.9 on CWMBS/CORAL, consistent with extra entry points for missed branches. On the model side, the clearest contribution is positional refinement (Table 6: Fiber F1 27.5 without offset and perturbation, 52.7 with both).

Per-neuron brain-wide results. We use the TreeTracer results of Table 3 (Fiber micro-F1 41.56%), obtained with the brain-wide pipeline described in the paper. We computed macro-averaged F1 over the same 14 neurons, with equal weight per neuron, and paired differences with 95% percentile CIs from 10,000 neuron-level bootstrap resamples.

The macro-averaged results were:

MethodDendrite macro-F1 (%)Axon macro-F1 (%)Fiber macro-F1 (%)
Kimimaro (SegMamba)69.1123.0735.13
APP2 (SwinUNETR)62.2020.9734.04
neuTube (MedNeXt)56.4822.3932.70
TreeTracer64.7928.9940.25

The paired Fiber F1 comparisons were:

BaselineNeurons with higher Fiber F1Mean gain, pp [95% CI]Median gain, pp
Kimimaro (SegMamba)9/14 (64.3%)+5.11 [−1.51, 12.16]+6.02
APP2 (SwinUNETR)9/14 (64.3%)+6.21 [−3.31, 15.39]+5.74
neuTube (MedNeXt)10/14 (71.4%)+7.55 [−1.10, 16.31]+7.98

TreeTracer has the highest Fiber and axon macro-F1, and its Fiber F1 is higher on most neurons against each baseline; its dendrite F1 is below Kimimaro's. With 14 neurons, all three intervals include zero, so these results show a consistent direction rather than an established gain. Excluding the three longest neurons, which also have the most GT branch points, the mean gains are +5.03, +7.89, and +8.63 points (higher on 7/11, 8/11, and 9/11 neurons), so the gains are not driven by the largest cells; they are larger in the half of the cohort with more branch points. We will add per-neuron plots to the revision.

Reviewer_c9bYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TreeTracer grows neuronal subtrees from 3-D light-microscopy images and an oriented seed, iteratively predicting a position offset, an outgoing direction, and a node type (continuation, branch or leaf). A "sandwich" variant wraps the pretrained model with image-derived seeds plus heuristic fusion and pruning. It is evaluated on CWMBS and CORAL crops locally and on 14 neurons brain-wide.

Critical strengths

Worthwhile problem: A single false junction or missed branch can corrupt an entire downstream arbor.

Strong single-path tracing: On non-branching subtrees, the base model reaches 70.1/75.6 F1, versus 55.4/65.8 for the best baseline.

Broad evaluation: Evaluations across crops and whole-brain volumes, with an axon gain.

Critical weaknesses
  1. Unclear novelty. [1] already grows a tree from a root by node classification and spawning queries at branches. [2] adds cross-patch trajectory memory. Neither is cited. And [3] is also closer than the paper suggests, which uses history-aware node/edge decoding.

  2. Reliance on heuristics. Greedy decoding cannot revise errors, so missed branches require rule-based reseeding. Tracing beyond one 32×96×96 crop, including for Tables 1–2, relies on hand-crafted stitching and pruning in Algorithm 2. Motivating errors, such as false junctions and missed branches, are not quantified.

  3. Under-specified topology decoding. The number of children is set by a 0.5 probability threshold, angular suppression and a two-child cap rather than predicted. Training also perturbs only positions while keeping ground-truth types and directions in the history, so the model never learns to handle its own topological errors.

  4. Overstated topology gains. The abstract reports the rule-based sandwich scores without labeling them, while the base model beats the strongest baseline by only 0.9 F1, less than one bootstrap standard deviation. On branching subtrees, it trails the best two-stage baselines by 12.8 points on CWMBS and 17.1 on CORAL, contradicting the claimed improvement in topology prediction.

Reference:

[1] Trexplorer: Recurrent detr for topologically correct tree centerline tracking. In MICCAI 2024.

[2] Trexplorer super: Topologically correct centerline tree tracking of tubular objects in ct volumes. In MICCAI 2025.

[3] Netracer: A topology-aware iterative tracing approach for tubular structure extraction. In ICCV 2025.

Questions for authors
  1. What does TreeTracer exactly add over related works, and can a controlled comparison or targeted ablation show that these components matter?

  2. How sensitive are branch results to the 0.5 threshold, angular suppression and the two-child cap, and how often do the type and direction heads disagree? Have you tried predicting the number of children directly, or perturbing type and direction inputs during training?

  3. Can you report false junctions, missed branches and fiber switches directly, together with branch-point precision and recall within a spatial tolerance?

  4. If the strongest baselines receive the same extra seeds and post-processing, how much of the sandwich gain remains?

  5. With the current heading and image features held fixed, does a longer history improve F1 consistently across several training seeds?

8 comments
AuthorsOpenReview ↗

Part[1/6] Response to Q1 & W1: relation to Trexplorer, Trexplorer Super, and NETracer

Thank you for the detailed review. We added the requested comparisons and analyses and will revise several claims.

Positioning. We agree that Trexplorer [1] and Trexplorer Super [2] are closely related and should have been discussed; we will cite both and expand the comparison with NETracer [3]. Rooted tree growth, learned branching, and the use of past trajectory information all appear in prior work, and we do not claim them individually. Our contribution is a specific formulation of these ideas, which differs as follows.

  • TreeTracer’s autoregressive causal decoder represents trajectory history within and across windows in a unified format. In contrast, Trexplorer and Trexplorer Super handle these two cases separately: within each window, they reuse the object queries from the previous step as inputs to the current step; across windows, they encode historical information using a separate MLP and add the resulting embeddings to the object queries. Both methods use Hungarian matching during training. TreeTracer conditions each prediction step on an ordered sequence of ancestor states, each comprising a position, an outgoing direction, and a node type. Each ancestor state is represented as a separate token and processed through causal self-attention. Supervision is provided by aligned next-step targets along annotated paths, eliminating the need for query-to-node matching.
  • Direction-guided advance with a learned offset. At each step, the current outgoing direction codeword defines a coarse position proposal, which determines where local image features are sampled. A bounded learned offset then provides a fine-grained correction, yielding the final next-step position.
  • NETracer only traces one path at a time. From many seeds it follows single paths, guided by predicted future nodes and edges, and stops at already traced paths, so its reconstruction is assembled from path fragments. TreeTracer instead predicts branching as a node type during tracing and expands all children from the shared history, so one prompt yields one connected tree.

Comparison with Trexplorer and Super. We adapted and trained Trexplorer and Trexplorer Super using their official implementations and evaluated them on our neuron-reconstruction tasks using the Fiber metrics. We report vanilla TreeTracer, without sandwich multi-seed augmentation.

MethodCWMBSCORAL
No-branch F1Branch F1All F1No-branch F1Branch F1All F1
Trexplorer7.43.17.4_{3.1}3.91.33.9_{1.3}5.81.75.8_{1.7}62.02.662.0_{2.6}27.52.327.5_{2.3}46.82.146.8_{2.1}
Trexplorer Super70.44.1\mathbf{70.4}_{4.1}31.8‾2.4\underline{31.8}_{2.4}50.6‾2.8\underline{50.6}_{2.8}75.3‾2.2\underline{75.3}_{2.2}31.6‾2.5\underline{31.6}_{2.5}55.7‾2.2\underline{55.7}_{2.2}
TreeTracer (vanilla)70.1‾3.8\underline{70.1}_{3.8}37.82.4\mathbf{37.8}_{2.4}52.72.4\mathbf{52.7}_{2.4}75.62.3\mathbf{75.6}_{2.3}46.02.8\mathbf{46.0}_{2.8}61.92.0\mathbf{61.9}_{2.0}

TreeTracer improves overall F1 over Trexplorer by 46.9 and 15.1 percentage points on CWMBS and CORAL, respectively. Against the stronger Trexplorer Super, the improvements are 2.1 and 6.2 points. The comparison with Super is particularly informative: no-branch performance is similar—70.1 versus 70.4 on CWMBS and 75.6 versus 75.3 on CORAL—while TreeTracer improves branch-subset F1 by 6.0 and 14.4 points, respectively. Thus, the main performance advantage over Super occurs on subtrees containing branches.

Model size and computational efficiency. We also measured the following implementation-level costs. Throughput is correctly traced length (voxels) per second on the 483 CWMBS test prompts (RTX 3090, Table 13 timing scope); the full timing table, including sandwich, is in Part 2 of our response to Reviewer gcdR.

MethodParametersTraining updates to validation-loss plateau on CORALCorrect length per second
Trexplorer51.22M400003.30
Trexplorer Super55.28M8200015.21
TreeTracer (vanilla)7.38M500080.10

TreeTracer uses 85.6% and 86.6% fewer parameters than Trexplorer and Super, respectively. Its measured tracing throughput is 24× and 5.3× higher. With the same batch size and learning rate in the reported CORAL experiments, its validation loss reached a plateau after fewer optimization updates. We will include the adaptation, training, and timing details in the revision.

References

[1] R. Naeem et al. Trexplorer: Recurrent DETR for Topologically Correct Tree Centerline Tracking. MICCAI 2024.

[2] R. Naeem et al. Trexplorer Super: Topologically Correct Centerline Tree Tracking of Tubular Objects in CT Volumes. MICCAI 2025.

[3] C. Liu, Y. Jiang, N. Zheng. NETracer: A Topology-Aware Iterative Tracing Approach for Tubular Structure Extraction. ICCV 2025.

AuthorsOpenReview ↗

Part[2/6] Response to Q2 & W3: topology decoding

Child count. We do not explicitly predict the number of children. Instead, we assign all nodes with multiple children to a single bifurcation class and cap the number of predicted children at two. This design is motivated by two considerations.

  1. Branch nodes are rare, and almost all are bifurcations. In the CWMBS training subtrees, only 0.64% of nodes are branch nodes (2,018 of 314,795), and 98.6% of these have exactly two children; only 28 have three or more. Splitting the branch class by child count would therefore provide extremely limited supervision for higher-order branches.
  2. Annotated multifurcations may reflect closely spaced bifurcations. In neuron annotations, some multifurcations may correspond to several nearby bifurcations rather than a single higher-order junction. TreeTracer can represent such configurations by predicting successive bifurcation nodes, allowing it to recover the associated branches while retaining the two-child cap.

Sensitivity to threshold α and the child cap. We evaluated 8 decoding configurations on the CORAL test dataset, using a fixed checkpoint and holding all other inference settings constant. The threshold α controls the cumulative probability mass of direction candidates; the type head uses argmax. Branch-point F1 and Fiber F1 are computed over all 500 subtrees after the full sandwich pipeline.

Cumulative threshold αChild capBranch-point F1Fiber F1
0.3231.4067.38
0.4231.4866.91
0.5 (default)231.8866.42
0.6233.3367.11
0.7232.3666.51
0.9229.9565.07
0.5Unlimited31.3865.99
0.9Unlimited16.4447.39
  1. Threshold α. Around the default (α=0.3–0.7, cap=2), branch-point F1 spans 31.40–33.33% and Fiber F1 spans 66.42–67.38%. Raising α to 0.9 reduces these scores by 1.93 and 1.35 percentage points relative to the default. These results indicate modest sensitivity near the default for the final pipeline.

  2. Child cap. At α = 0.5, removing the two-child cap introduces slightly more false-positive branches than the additional true-positive branches it recovers, leading to decreases of only 0.50 and 0.43 points in Branch-point F1 and Fiber F1, respectively. At α = 0.9, the increase in false-positive branches substantially outweighs the gain in true positives, resulting in much larger drops of 13.51 and 17.68 points in Branch-point F1 and Fiber F1, respectively. These results indicate that the two-child cap has little impact on performance under the default setting of α = 0.5.

For angular suppression, the minimum pairwise separation in the 192-direction codebook is 12.81°. Consequently, the default 6° rule cannot suppress distinct direction tokens.

Regarding head disagreement, the two heads predict complementary quantities: the type head predicts the upcoming topological event, while the direction head estimates a distribution over continuation directions. Distribution modes do not correspond one-to-one with node types: directional uncertainty can produce multiple plausible directions at a continuation, while a bifurcation need not produce clearly separated peaks. This ambiguity motivates a dedicated type head supervised to predict topology directly.

Regarding perturbations to type and direction inputs during training, we perturb point coordinates while retaining their original type and direction tokens. The model therefore encounters displaced tracing states paired with reference cues that may no longer match the local structure at the perturbed locations. This augmentation introduces potential inconsistencies between position and the accompanying type/direction information.

AuthorsOpenReview ↗

Part [3/6] Response to Q3: keypoint and topology-error metrics

Keypoint metric. Following the CORAL evaluation protocol, we report branch and leaf precision (P), recall (R), and F1, using a spatial tolerance of 5 voxels on both datasets.

All values are in % (higher is better).

MethodCWMBSCORAL
Branch PBranch RBranch F1Leaf PLeaf RLeaf F1Branch PBranch RBranch F1Leaf PLeaf RLeaf F1
Segmentation + reconstruction
neuTube (raw)4.41.44.4_{1.4}37.75.737.7_{5.7}7.92.17.9_{2.1}10.72.910.7_{2.9}33.62.533.6_{2.5}16.23.016.2_{3.0}4.11.24.1_{1.2}48.44.048.4_{4.0}7.52.17.5_{2.1}13.93.713.9_{3.7}45.61.945.6_{1.9}21.34.121.3_{4.1}
neuTube(V-Net)12.91.912.9_{1.9}45.03.345.0_{3.3}20.12.320.1_{2.3}27.83.227.8_{3.2}41.12.041.1_{2.0}33.22.233.2_{2.2}21.56.221.5_{6.2}49.83.849.8_{3.8}30.05.830.0_{5.8}52.67.552.6_{7.5}55.21.955.2_{1.9}53.94.153.9_{4.1}
neuTube(DSCNet)10.81.410.8_{1.4}51.34.2\mathbf{51.3}_{4.2}17.81.917.8_{1.9}23.02.623.0_{2.6}44.32.344.3_{2.3}30.32.230.3_{2.2}25.15.025.1_{5.0}52.03.9\mathbf{52.0}_{3.9}33.94.6\mathbf{33.9}_{4.6}56.44.956.4_{4.9}57.21.857.2_{1.8}56.82.756.8_{2.7}
neuTube(SegMamba)11.81.511.8_{1.5}47.2‾4.9\underline{47.2}_{4.9}18.91.918.9_{1.9}24.73.124.7_{3.1}40.52.140.5_{2.1}30.72.230.7_{2.2}21.67.321.6_{7.3}50.2‾3.6\underline{50.2}_{3.6}30.26.530.2_{6.5}54.68.754.6_{8.7}57.12.057.1_{2.0}55.84.555.8_{4.5}
neuTube(GBP-Net)13.61.113.6_{1.1}39.14.639.1_{4.6}20.21.320.2_{1.3}30.62.630.6_{2.6}37.82.637.8_{2.6}33.81.533.8_{1.5}20.34.320.3_{4.3}52.03.5\mathbf{52.0}_{3.5}29.24.529.2_{4.5}51.65.451.6_{5.4}57.61.957.6_{1.9}54.43.154.4_{3.1}
APP2(V-Net)7.51.57.5_{1.5}22.53.122.5_{3.1}11.31.711.3_{1.7}28.73.828.7_{3.8}37.92.637.9_{2.6}32.72.432.7_{2.4}21.56.721.5_{6.7}28.94.428.9_{4.4}24.74.224.7_{4.2}49.96.849.9_{6.8}32.12.032.1_{2.0}39.12.439.1_{2.4}
APP2(DSCNet)7.11.47.1_{1.4}25.93.625.9_{3.6}11.11.811.1_{1.8}24.93.524.9_{3.5}37.92.437.9_{2.4}30.02.530.0_{2.5}20.86.920.8_{6.9}26.43.426.4_{3.4}23.24.523.2_{4.5}50.16.550.1_{6.5}31.61.831.6_{1.8}38.82.538.8_{2.5}
APP2(SegMamba)7.51.47.5_{1.4}23.23.523.2_{3.5}11.31.711.3_{1.7}27.63.627.6_{3.6}37.62.537.6_{2.5}31.82.631.8_{2.6}24.45.224.4_{5.2}27.84.027.8_{4.0}26.03.626.0_{3.6}54.54.654.5_{4.6}32.62.032.6_{2.0}40.81.940.8_{1.9}
APP2(GBP-Net)12.34.412.3_{4.4}15.62.615.6_{2.6}13.82.913.8_{2.9}41.76.941.7_{6.9}30.22.430.2_{2.4}35.03.035.0_{3.0}0.20.10.2_{0.1}17.02.917.0_{2.9}0.30.10.3_{0.1}1.20.31.2_{0.3}26.41.826.4_{1.8}2.20.62.2_{0.6}
Kimimaro(V-Net)29.2‾4.1\underline{29.2}_{4.1}22.53.722.5_{3.7}25.43.425.4_{3.4}62.52.862.5_{2.8}46.03.046.0_{3.0}53.02.653.0_{2.6}41.87.141.8_{7.1}13.72.213.7_{2.2}20.73.120.7_{3.1}81.61.781.6_{1.7}50.61.950.6_{1.9}62.41.762.4_{1.7}
Kimimaro(DSCNet)25.03.625.0_{3.6}30.33.630.3_{3.6}27.42.527.4_{2.5}59.83.659.8_{3.6}54.82.754.8_{2.7}57.22.057.2_{2.0}48.86.4\mathbf{48.8}_{6.4}22.03.022.0_{3.0}30.33.530.3_{3.5}81.8‾1.9\underline{81.8}_{1.9}53.62.053.6_{2.0}64.81.764.8_{1.7}
Kimimaro(SegMamba)19.02.919.0_{2.9}31.84.631.8_{4.6}23.82.223.8_{2.2}50.64.850.6_{4.8}51.93.451.9_{3.4}51.22.951.2_{2.9}48.1‾6.1\underline{48.1}_{6.1}23.52.823.5_{2.8}31.63.331.6_{3.3}80.11.780.1_{1.7}52.72.052.7_{2.0}63.61.763.6_{1.7}
Kimimaro(GBP-Net)31.23.2\mathbf{31.2}_{3.2}32.83.632.8_{3.6}31.92.6\mathbf{31.9}_{2.6}61.52.561.5_{2.5}52.72.152.7_{2.1}56.81.756.8_{1.7}26.13.326.1_{3.3}32.13.132.1_{3.1}28.82.828.8_{2.8}70.42.270.4_{2.2}63.71.863.7_{1.8}66.91.566.9_{1.5}
Tracing
DRL0.00.00.0_{0.0}0.00.00.0_{0.0}0.00.00.0_{0.0}54.91.054.9_{1.0}35.31.335.3_{1.3}43.01.043.0_{1.0}0.00.00.0_{0.0}0.00.00.0_{0.0}0.00.00.0_{0.0}53.90.553.9_{0.5}39.50.939.5_{0.9}45.60.645.6_{0.6}
SPE-DNR9.72.79.7_{2.7}19.83.419.8_{3.4}13.12.713.1_{2.7}28.53.428.5_{3.4}35.81.335.8_{1.3}31.72.131.7_{2.1}12.52.112.5_{2.1}15.52.515.5_{2.5}13.92.113.9_{2.1}38.81.338.8_{1.3}36.41.136.4_{1.1}37.61.037.6_{1.0}
NETracer0.00.00.0_{0.0}0.00.00.0_{0.0}0.00.00.0_{0.0}57.40.857.4_{0.8}35.41.635.4_{1.6}43.81.443.8_{1.4}0.00.00.0_{0.0}0.00.00.0_{0.0}0.00.00.0_{0.0}53.10.653.1_{0.6}35.61.135.6_{1.1}42.60.942.6_{0.9}
TreeTracer (vanilla)27.45.927.4_{5.9}7.81.87.8_{1.8}12.22.512.2_{2.5}75.72.0\mathbf{75.7}_{2.0}59.0‾2.1\underline{59.0}_{2.1}66.31.8\mathbf{66.3}_{1.8}29.76.029.7_{6.0}9.71.79.7_{1.7}14.72.414.7_{2.4}82.21.5\mathbf{82.2}_{1.5}68.2‾1.6\underline{68.2}_{1.6}74.51.4\mathbf{74.5}_{1.4}
TreeTracer (sandwich)28.63.228.6_{3.2}31.33.631.3_{3.6}29.9‾2.9\underline{29.9}_{2.9}64.7‾2.1\underline{64.7}_{2.1}66.32.1\mathbf{66.3}_{2.1}65.5‾1.7\underline{65.5}_{1.7}27.32.927.3_{2.9}38.33.538.3_{3.5}31.9‾2.8\underline{31.9}_{2.8}72.21.972.2_{1.9}76.51.6\mathbf{76.5}_{1.6}74.3‾1.4\underline{74.3}_{1.4}

Comments on the paper

2 comments

Thank you for your response. We are pleased that our explanations have addressed most of your concerns.

We acknowledge that the ablation in Part [6/6] of our response evaluated the prediction accuracy of the type and direction heads using ground-truth (GT) histories of varying lengths. In our follow-up response to Reviewer 1sag, we provided additional experiments evaluating the effect of history length on reconstruction performance during inference, when the model conditions on its own previous predictions. We reproduce these results below for convenience.

We evaluated H∈{1,2,4,8,16}H \in \{1, 2, 4, 8, 16\} on CORAL and CWMBS using three metrics: Fiber F1, Bifurcation F1, and Leaf F1. Subscripts denote standard deviations across 1,000 bootstrap resamples at the source-image level.

CORAL

MethodMetricH=1H=2H=4H=8H=16
VanillaFiber F146.175.57\mathbf{46.17}_{\mathbf{5.57}}53.444.71\mathbf{53.44}_{\mathbf{4.71}}60.602.81\mathbf{60.60}_{\mathbf{2.81}}63.492.26\mathbf{63.49}_{\mathbf{2.26}}62.812.18\mathbf{62.81}_{\mathbf{2.18}}
VanillaBifurcation F19.312.629.31_{2.62}9.772.419.77_{2.41}10.732.1510.73_{2.15}13.692.3313.69_{2.33}12.221.9512.22_{1.95}
VanillaLeaf F162.584.6662.58_{4.66}68.243.5968.24_{3.59}73.791.9373.79_{1.93}75.811.4675.81_{1.46}75.571.5975.57_{1.59}
SandwichFiber F157.083.3257.08_{3.32}62.802.4962.80_{2.49}63.892.1663.89_{2.16}66.952.3966.95_{2.39}67.432.0167.43_{2.01}
SandwichBifurcation F123.332.3423.33_{2.34}27.872.2627.87_{2.26}26.602.2426.60_{2.24}29.442.4229.44_{2.42}28.952.1628.95_{2.16}
SandwichLeaf F167.982.4267.98_{2.42}70.961.9470.96_{1.94}72.251.6272.25_{1.62}73.211.7473.21_{1.74}74.901.4874.90_{1.48}

CWMBS

MethodMetricH=1H=2H=4H=8H=16
VanillaFiber F130.655.53\mathbf{30.65}_{\mathbf{5.53}}43.333.39\mathbf{43.33}_{\mathbf{3.39}}48.774.22\mathbf{48.77}_{\mathbf{4.22}}49.364.01\mathbf{49.36}_{\mathbf{4.01}}49.123.65\mathbf{49.12}_{\mathbf{3.65}}
VanillaBifurcation F15.982.075.98_{2.07}10.711.7110.71_{1.71}12.272.0212.27_{2.02}11.972.5111.97_{2.51}10.192.1210.19_{2.12}
VanillaLeaf F148.055.5248.05_{5.52}61.262.6961.26_{2.69}64.562.8764.56_{2.87}65.072.8565.07_{2.85}65.542.6665.54_{2.66}
SandwichFiber F150.063.8450.06_{3.84}52.853.7052.85_{3.70}53.504.1153.50_{4.11}53.713.8053.71_{3.80}55.573.0955.57_{3.09}
SandwichBifurcation F120.653.2220.65_{3.22}22.133.4422.13_{3.44}23.043.9823.04_{3.98}23.953.7823.95_{3.78}27.163.1127.16_{3.11}
SandwichLeaf F161.542.7961.54_{2.79}62.612.5262.61_{2.52}63.502.7263.50_{2.72}65.222.6765.22_{2.67}65.872.3765.87_{2.37}

These results show that longer history generally improves reconstruction performance when the model uses its own predictions as history:

  • Single-prompt tracing (vanilla): Increasing the history length from H=1H=1 to H=8H=8 improves Fiber F1 by 17.32 points on CORAL and 18.71 points on CWMBS.
  • Multi-seed tracing (sandwich): The same increase in history length improves Fiber F1 by 9.87 points on CORAL and 3.65 points on CWMBS.

The accompanying improvements in Bifurcation F1 and Leaf F1 indicate that longer history also benefits the detection of branch points and leaf nodes. These findings suggest that improved prediction of topological structure may contribute to the gains in final fiber reconstruction performance.

AuthorsOpenReview ↗

Summary of the rebuttal and additional evidence for TreeTracer

Dear Area Chair,

TreeTracer addresses a difficult decision in 3-D neuron reconstruction: when a tracer reaches a possible junction or weak signal, should it continue, branch, or stop? From a point and an initial direction, TreeTracer grows a connected subtree. It represents each ancestor along the current path as an ordered token. At each step, the current heading identifies where to look in the image, and the network predicts a position correction, a node type, and outgoing directions. This gives every tracing decision a training target along an annotated path.

How does TreeTracer compare with the closest alternatives? We trained and evaluated Trexplorer and Trexplorer Super on the same neuron tasks. Vanilla TreeTracer achieved nearly identical Fiber F1 on non-branching subtrees, while gained 6.0 points on CWMBS and 14.4 on CORAL on subtrees containing branches. Against the strongest two-stage baseline, Kimimaro (GBP-Net), the sandwich pipeline achieved the highest overall Fiber F1 on both datasets. Its gains over Kimimaro were +6.63 points on CWMBS (paired 95% CI [2.40, 10.54]) and +5.40 on CORAL ([1.38, 9.53]).

Does trajectory history improve the final reconstruction? Reviewer c9bY asked whether the benefit persists when the single-prompt tracer follows its own predictions. Reviewer 1sag asked whether history still contributes to full TreeTracer(sandwich) system. We tested both settings by increasing the inference history from one node to eight. The one-node condition retains the current heading.

Gain in Fiber F1 from one to eight history nodesCWMBSCORAL
Single-prompt TreeTracer (vanilla)+18.71 points+17.32 points
Full multi-seed system (sandwich)+3.65 points+9.87 points

Bifurcation F1 and leaf F1 also improved in all four settings Together, these experiments connect earlier trajectory information to better topology decisions and show that the benefit carries through to reconstruction when the tracer makes its own decisions.

What remains to be revised? Reviewer gcdR’s remaining concern was whether a first-time reader could follow Section 3. We provided an explicit definition of the prompted subtree in the discussion. In the manuscript, we will introduce it through a worked example, show its nodes, edges, and coordinate map in Eq. (2), and use one tracing step to connect each network output in Section 3.3 to the action it controls in Section 3.2. We will also state the scope of the evaluation: the experiments cover mouse whole-brain microscopy, including tracing across windows in brain-wide neurons. Generalization to the species and imaging modalities in BigNeuron remains open.

Reviewer c9bY indicated that our rebuttal addressed most of their concerns, and Reviewer 1sag acknowledged that the new experiments strengthened the evidence for both the full system and the value of trajectory history. Their follow-up questions focused on whether longer history improves reconstruction when the tracer follows its own predictions, both from a single prompt and within the full multi-seed pipeline. We tested both settings directly and observed gains in Fiber F1 on both datasets. Reviewer gcdR’s remaining concern was whether readers could follow the formulation in Section 3. In the discussion, we clarified how the prompt determines the subtree’s nodes, edges, and coordinates. In the revised manuscript, we will introduce these definitions through a worked example and add pseudocode showing how each network prediction determines the next tracing step.

We believe the strongest case for accepting TreeTracer is that its representation of the path traced so far improves decisions that matter for reconstructing a neuronal arbor. That effect is measurable in individual node decisions, in single-prompt reconstruction, and in the complete system on both datasets. We respectfully request your consideration for acceptance.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.