acceptodds
Under review as a conference paper at ICLR 2027

Sparse2Geo: Prediction-Derived Geometry for Sparsely Supervised Segmentation and Regional Quality Assessment in Volume Electron Microscopy

Abstract

Volume electron microscopy (vEM) enables nanometer-scale three-dimensional (3D) imaging of cellular ultrastructure, but semantic segmentation remains annotation-intensive. Sparse supervision reduces annotation cost, yet supervision confined to sparsely sampled locations provides limited constraints on object extent, boundaries, and 3D structure. Signed distance fields (SDFs) continuously encode object interior, exterior, and distance to the boundary, but are difficult to construct directly from sparse labels. Regional segmentation quality also remains difficult to assess without ground truth. To address both challenges, we introduce Sparse2Geo, a unified framework that exploits prediction-derived geometry for both sparsely supervised segmentation and regional quality assessment. In Stage I, Sparse2Geo constructs a teacher SDF from 3D consensus geometry aggregated from fixed full-volume predictions of an initial sparsely supervised model, providing geometric and boundary-aware supervision without additional dense semantic targets. In Stage II, local predictions are re-encoded as normalized local SDFs and combined with aligned raw volumetric patches to estimate regional segmentation quality without ground truth at inference. Our experiments use volumes from BetaSeg and UroCell, two public vEM datasets spanning mouse pancreatic cells and urinary bladder urothelial cells, comprising approximately 2.32 billion and 67 million voxels, respectively. Sparse2Geo achieves the highest Macro Dice among the compared segmentation methods (90.63% and 71.44%), improving over -Seg by 4.13 and 15.33 percentage points, respectively. For regional quality assessment, Sparse2Geo achieves the lowest Macro MAE on BetaSeg, reducing it by 12.3% relative to the best adapted external baseline, while incorporating local SDF geometry reduces Macro MAE by 35.1% relative to Raw-only on UroCell.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
sparse supervision, semantic segmentation, signed distance fields, segmentation quality assessment, volume electron microscopy, biomedical image analysis

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_nNfPOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a two-stage framework for sparsely supervised segmentation and regional quality estimation in volume electron microscopy. Stage I trains an initial segmenter, constructs a fixed three-dimensional consensus mask from its predictions, and converts that mask into a signed distance field (SDF). A separately initialized final segmenter receives sparse semantic supervision together with SDF regression and boundary-region consistency losses. Stage II predicts class-conditional local Dice from raw image regions and SDFs computed from the final segmentation. Experiments on BetaSeg and UroCell report improved segmentation and regional quality estimation.

Critical strengths
  1. The problem and representation are well motivated. Sparse labels constrain boundaries poorly, and a continuous distance representation is a reasonable way to expose spatial structure.

  2. The ablation study is well conducted. Table 3 separates the dense prediction head, validation protocol, and geometric objectives. Tables 10 and 13 examine teacher construction and compare SDFs with binary masks and class metadata.

  3. The final performance of the model is good on benchmarks.

Critical weaknesses
  1. The headline segmentation gain does not clearly reflect the contribution of the proposed geometric objectives. The reported 4.13 percentage-point improvement over ε-Seg on BetaSeg includes changes to both the output head and the checkpoint-selection protocol. Table 3 shows that the controlled baseline already achieves 90.07% Macro Dice, compared with 90.63% for the full method. Thus, most of the reported improvement comes from the dense head and representative-slice validation, whereas the methodological emphasis is on the two geometric losses.

  2. The evidence for robustness and generalization is limited. Each dataset contains only one held-out evaluation volume, and the reported configurations use a fixed random seed. This makes it difficult to assess variability across training runs and generalization across volumes.

  3. The benefit of SDF supervision over simpler alternatives remains unclear. What does SDF supervision add over directly training the final segmenter with the same consensus pseudo masks? A comparison using the same teacher predictions, consensus masks, backbonewould help isolate the value of the geometric representation.

  4. The choice of a separate quality estimator needs further justification. Why is regional quality estimation trained independently after the final segmenter is frozen, rather than incorporated into a jointly trained or shared-backbone model? I recognize that the current segmenter and quality estimator operate on different input dimensionalities and that the quality targets depend on the segmentation predictions. Nevertheless, discussing these trade-offs would help explain the design choice and clarify whether modularity, optimization stability, or empirical performance motivates the separate training procedure.

Questions for authors

The training workflow should be described more explicitly. Although the framework is organized into two conceptual stages, it involves three separate model-training procedures: training the initial segmenter, training a separately initialized final segmenter, and training the regional quality estimator. Between the first two procedures, the initial predictions are processed offline to construct the consensus masks and teacher SDFs. Now, it is rather harder to get the whole training pipeline picture.

6 comments
AuthorsOpenReview ↗

Response to Reviewer nNfP (Part 1/4) — W1/W2

Reviewer comment — W1

The headline segmentation gain does not clearly reflect the contribution of the proposed geometric objectives. The reported 4.13 percentage-point improvement over ε-Seg on BetaSeg includes changes to both the output head and the checkpoint-selection protocol. Table 3 shows that the controlled baseline already achieves 90.07% Macro Dice, compared with 90.63% for the full method. Thus, most of the reported improvement comes from the dense head and representative-slice validation, whereas the methodological emphasis is on the two geometric losses.

Reviewer comment — W2

The evidence for robustness and generalization is limited. Each dataset contains only one held-out evaluation volume, and the reported configurations use a fixed random seed. This makes it difficult to assess variability across training runs and generalization across volumes.

Response

We thank the reviewer for their comments on the contribution of geometric supervision and experimental stability. In earlier experiments with different random seeds, we observed small performance variations. The overall improvement over the official ϵ-Seg implementation reflects the combined effects of the dense semantic prediction head, validation protocol, and geometric supervision. To isolate the contribution of geometric supervision, we added matched comparisons under three configurations on BetaSeg and UroCell. Within each configuration, we keep the prediction head and validation protocol fixed and compare no additional teacher supervision with Full SDF.

DatasetDense headRep.-valNo additional teacher supervisionFull SDFGain (pp)
BetaSeg××86.5087.69+1.19
BetaSeg✓×88.8890.05+1.17
BetaSeg✓✓90.0790.63+0.56
UroCell××56.1160.47+4.36
UroCell✓×69.6970.87+1.18
UroCell✓✓70.7671.44+0.68

Results are Macro Dice (%) on the held-out test set, including the background class. × denotes the official center head or default validation protocol; ✓ denotes the dense head or representative-slice Dice validation (Rep.-val). Gains are in percentage points.

We also report progressive ablations on both datasets, showing the changes after successively adding geometric supervision, the dense semantic prediction head, and representative-slice validation.

DatasetFull SDFDense headRep.-valMacro Dice (%)Incremental gain (pp)
BetaSeg×××86.50—
BetaSeg✓××87.69+1.19
BetaSeg✓✓×90.05+2.36
BetaSeg✓✓✓90.63+0.58
UroCell×××56.11—
UroCell✓××60.47+4.36
UroCell✓✓×70.87+10.40
UroCell✓✓✓71.44+0.57

Full SDF improves performance in all six matched comparisons. With the official center head and default model selection, the final segmenter gains 1.19 and 4.36 percentage points on BetaSeg and UroCell, respectively. With the dense head and representative-slice validation, the gains are 0.56 and 0.68 percentage points. These results support an additional contribution from geometric supervision across the evaluated configurations. In the revised manuscript, Section 4.3 distinguishes the overall improvement from the incremental contribution of geometric supervision. Table 4 reports the matched comparisons, and Appendix B.2 reports the progressive ablation (Table 13).

BetaSeg follows the training/test split used in prior work, whereas UroCell uses our fixed split. Each dataset has one held-out test volume, and all compared methods use the same split to ensure a fair comparison.

AuthorsOpenReview ↗

Response to Reviewer nNfP (Part 2/4) — W3

Reviewer comment — W3

The benefit of SDF supervision over simpler alternatives remains unclear. What does SDF supervision add over directly training the final segmenter with the same consensus pseudo masks? A comparison using the same teacher predictions, consensus masks, backbone would help isolate the value of the geometric representation.

Response

We thank the reviewer for this suggestion. To compare SDF geometric supervision with direct supervision from consensus masks, we added comparisons under three configurations on both datasets. Within each configuration, the segmentation network and validation protocol are fixed. Mask and Full SDF use the same teacher predictions and consensus masks for direct Mask supervision and SDF geometric supervision, respectively.

DatasetDense headRep.-valNo additional teacher supervisionMaskFull SDF
BetaSeg××86.5087.3387.69
BetaSeg✓×88.8889.8290.05
BetaSeg✓✓90.0790.4690.63
UroCell××56.1154.9560.47
UroCell✓×69.6966.8970.87
UroCell✓✓70.7666.7571.44

Results are Macro Dice (%) on the held-out test set, including the background class. × denotes the official center head or default validation protocol; ✓ denotes the dense head or representative-slice Dice validation (Rep.-val). Bold indicates the best result within each configuration.

Full SDF outperforms direct Mask supervision and the baseline without additional teacher supervision in all six comparisons shown. Relative to Mask, the gains are 0.17–0.36 percentage points on BetaSeg and 3.98–5.52 percentage points on UroCell.

Whereas binary masks encode class membership, SDFs additionally encode continuous distances inside and outside the target and the location of its boundary. These results support the additional value of converting the same consensus prediction into geometric supervision. We report these comparisons in Table 4 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer nNfP (Part 3/4) — W4

Reviewer comment — W4

The choice of a separate quality estimator needs further justification. Why is regional quality estimation trained independently after the final segmenter is frozen, rather than incorporated into a jointly trained or shared-backbone model? I recognize that the current segmenter and quality estimator operate on different input dimensionalities and that the quality targets depend on the segmentation predictions. Nevertheless, discussing the trade-offs would help explain the design choice and clarify whether modularity, optimization stability, or empirical performance motivates the separate training procedure.

Response

We thank the reviewer for their question about independent training. Stage II is designed as a regional quality-assessment module applied after segmentation. It computes local SDFs from the final prediction masks and combines them with raw images as estimator inputs. The module does not depend on internal segmentation features or a specific backbone, allowing independent training and adjustment without modifying the segmentation network.

Independent training also keeps the quality-supervision task fixed. Regional Dice depends on the segmentation prediction, so freezing the final segmenter keeps both inputs and quality targets fixed during estimator training. The two-dimensional segmenter and three-dimensional regional quality estimator can also use architectures suited to their respective tasks.

To assess this module under different annotation conditions, we compare the original quality-supervision setting with extremely sparse supervision that reuses existing Stage I manually annotated points. In the extremely sparse quality-supervision setting, regional Dice is computed only at these points for estimator training and validation-based model selection. Both settings are evaluated against test-region Dice computed using the original evaluation protocol.

DatasetQuality supervisionInputMacro MAE ↓Macro ρ ↑AUPRC <0.90 ↑AUPRC <0.80 ↑
BetaSegOriginalRaw0.03080.59170.88940.3604
BetaSegOriginalRaw + local SDF0.02430.80000.96170.6814
BetaSegExtremely sparseRaw0.05510.36380.90970.1296
BetaSegExtremely sparseRaw + local SDF0.05670.67070.94580.2127
UroCellOriginalRaw0.22760.31250.89390.8342
UroCellOriginalRaw + local SDF0.14770.74210.97450.9487
UroCellExtremely sparseRaw0.32620.14750.88640.8223
UroCellExtremely sparseRaw + local SDF0.26390.40850.90110.8828

The original-setting results are from the paper; the extremely sparse setting is a new experiment. Both settings use the same 597 test regions on BetaSeg and 600 on UroCell. Bold indicates the better result within each dataset and supervision setting.

Local SDFs improve quality ranking and low-quality-region detection on both datasets under both the original and extremely sparse supervision settings. These results show that the independent quality estimator can reuse existing Stage I sparse annotations for training and use prediction-derived geometry for regional quality assessment.

The trade-off is an additional quality-estimation network and training step. We choose independent training primarily for modularity, fixed quality targets, and task-specific architectures. Section 3.2 explains the separate training design, and Appendix C.2 reports the supervision comparison (Table 19).

Reviewer_42YfOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces two methods for the analysis of cellular ultrastructure from volumetric electron microscopy imaging. The first method (stage 1) tackles the challenges in training a semantic segmentation model from sparse annotations. They use a teacher-student framework: 1) train and freeze a 2D segmentation model using only the sparse annotations; 2) create a 3D consensus mask using orthogonal re-slicing and u-segment3D; 3) compute a signed distance function from the 3D consensus mask; 4) train a second (final) segmentation model using the SDF as a regression target and for boundary consistency. While the individual parts of this approach are not novel, the combination is the contribution. The second method (stage 2) tackles the problem of predicting the quality of the model segmentation. While segmentation quality prediction is not a novel task, the contribution of the paper is its use of prediction-derived local SDF as the input for regional segmentation-quality prediction.

The experimental validation uses two public datasets, BetaSeg and UroCell. Stage 1 is evaluated using full-volume per-class and Macro Dice. Stage 2 is evaluated using MAE, Spearman correlation, and low-quality-region AUPRC. The paper reports higher Macro Dice than sparse, self-supervised, and fully supervised baselines for stage 1. Stage 2 outperforms statistical and learned baselines. Ablations separately examine the dense prediction head, validation protocol, SDF regression, boundary consistency, consensus construction, target class. For Stage 2 ablations study input representation. The largest Stage 1 improvement comes from the dense head and validation protocol.

Critical strengths

The paper addresses two practically important problems in large-scale volumetric electron microscopy from extremely sparse annotations and identifying unreliable segmentation regions—using a computationally simple representation that adds no Stage 1 inference overhead and requires no ground truth for Stage 2 inference.

The experimental results are consistently positive across two datasets. Ablations of the geometric losses, boundary accuracy, consensus construction, and SDF-versus-mask representations are informative.

Critical weaknesses

The technical motivation and formulation is weak for ICLR. The connection between the two stages is mostly conceptual and sequential, not algorithmically unified. Even though both stages leverage the signed distance function (SDF), it plays different roles in the two stages. Stage I uses a consensus-derived SDF as an auxiliary training target, whereas Stage II independently recomputes local SDFs from the final discrete prediction and treats them as inputs to a quality-regression network. The connection between the two stages is a simple step-wise progression with no joint parts/parameters, no feedback from stage 2 to 1 or a unified training objective. Furthermore, the novelty is mostly incremental over existing work. The individual components of stage 1 are not novel, but the particular combination of using a prediction-derived SDF as fixed dense geometric supervision for a second sparse-label model is likely novel.

The method used to build the consensus mask is heuristic and narrowly focused on the particular application. Component size and other filtering rules are not reported. The results of the filtering are likely to depend on the orientation of the dataset due to the orthogonal re-slicing/

Questions for authors

Clarification: In Eq. 3, what exact component-filtering thresholds and u-Segment3D parameters are used?

What experimental evidence establishes that Stage II specifically benefits from the proposed Stage I rather than operating as an independent quality estimator applicable to any segmentation output?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 42Yf (Part 1/3) — W1

Reviewer comment — W1

The technical motivation and formulation is weak for ICLR. The connection between the two stages is mostly conceptual and sequential, not algorithmically unified. Even though both stages leverage the signed distance function (SDF), it plays different roles in the two stages. Stage I uses a consensus-derived SDF as an auxiliary training target, whereas Stage II independently recomputes local SDFs from the final discrete prediction and treats them as inputs to a quality-regression network. The connection between the two stages is a simple step-wise progression with no joint parts/parameters, no feedback from stage 2 to 1 or a unified training objective. Furthermore, the novelty is mostly incremental over existing work. The individual components of stage 1 are not novel, but the particular combination of using a prediction-derived SDF as fixed dense geometric supervision for a second sparse-label model is likely novel.

Response

We thank the reviewer for discussing the technical motivation and contributions. Our central idea is to use spatial geometry already present in segmentation predictions for sparsely supervised segmentation and regional quality assessment. Stage I converts aggregated predictions into a fixed teacher SDF, supplementing sparse semantic labels with continuous distance and boundary supervision. Stage II converts final predictions into local SDFs, allowing the quality estimator to use prediction geometry rather than relying only on raw images.

In Stage I, the specific design is to use a prediction-derived SDF as geometric supervision for a separately initialized final segmenter. We added matched comparisons to assess its value relative to direct supervision with aggregated pseudo-labels. Within each prediction-head and validation configuration, Mask and Full SDF use the same teacher predictions and consensus masks.

DatasetDense headRep.-valNo additional teacher supervisionMaskFull SDF
BetaSeg××86.5087.3387.69
BetaSeg✓×88.8889.8290.05
BetaSeg✓✓90.0790.4690.63
UroCell××56.1154.9560.47
UroCell✓×69.6966.8970.87
UroCell✓✓70.7666.7571.44

Results are Macro Dice (%) on the held-out test set, including the background class. × denotes the official center head or default validation protocol; ✓ denotes the dense head or representative-slice Dice validation (Rep.-val).

Full SDF outperforms both no additional teacher supervision and direct Mask supervision in all six comparisons shown. Thus, converting the same aggregated predictions into geometric supervision provides additional gains when the prediction head and validation protocol are controlled.

In Stage II, the contribution is the use of prediction-derived local SDFs as an input representation for regional quality assessment. The table shows the effect of adding local SDFs under the same quality-supervision setting.

DatasetInputMacro MAE ↓Macro ρ ↑AUPRC <0.90 ↑AUPRC <0.80 ↑
BetaSegRaw0.03080.59170.88940.3604
BetaSegRaw + local SDF0.02430.80000.96170.6814
UroCellRaw0.22760.31250.89390.8342
UroCellRaw + local SDF0.14770.74210.97450.9487

Results use the original quality-supervision setting in the paper. Bold indicates the better result on each dataset.

Adding local SDFs reduces quality-estimation error and improves rank correlation and low-quality-region detection on both datasets. These results support the value of prediction geometry as input to the quality estimator.

The stages are connected through the final segmentation prediction. Stage I uses geometry to improve segmentation, and Stage II uses prediction geometry to assess regional quality. The modular design allows the quality estimator to be trained and used independently without modifying the segmentation network. Our contribution lies in these two uses of prediction-derived geometry and the resulting segmentation and quality-assessment workflow. Section 3 of the revised manuscript describes this workflow. Table 4 reports the Stage I comparisons, and Tables 6 and 18 report the Stage II input comparisons.

AuthorsOpenReview ↗

Response to Reviewer 42Yf (Part 2/3) — W2/Q1

Reviewer comment — W2

The method used to build the consensus mask is heuristic and narrowly focused on the particular application. Component size and other filtering rules are not reported. The results of the filtering are likely to depend on the orientation of the dataset due to the orthogonal re-slicing/

Reviewer question — Q1

Clarification: In Eq. 3, what exact component-filtering thresholds and u-Segment3D parameters are used?

Response

We thank the reviewer for pointing out the missing aggregation details. We construct teacher masks using indirect 2D-to-3D aggregation from u-Segment3D. Specifically, we extract the target-class binary mask from the initial segmentation prediction and re-slice it in the XY, XZ, and YZ orientations. Slice-wise connected-component labeling and small-component filtering are followed by 3D aggregation. Both datasets use the same main settings, listed below.

Step or parameterSetting
Orthogonal viewsXY, XZ, YZ
2D connected components8-connectivity; remove components smaller than 8 pixels
AggregationIndirect 2D-to-3D aggregation, dtform_method = cellpose_skel
Skeleton smoothingsmooth_skel_sigma = 1.0
Gradient-descent iterationsn_iter = 250
Gradient-descent step sizedelta = 1.0
Gradient decay and momentumgradient_decay = 0.25, momenta = 0.98
BinarizationThree-class Multi-Otsu thresholding; use the higher threshold, rounded down to one decimal place
3D binary-mask processingClosing and dilation with a spherical structuring element of radius 1 voxel, followed by hole filling
3D small-component filteringRemove components smaller than 64 voxels during binary-mask processing and label regions smaller than 64 voxels after aggregation

These rules remove small prediction fragments and construct continuous 3D teacher geometry. For C2, the class used for supervision, aggregation increases Dice from 85.30% to 86.53% on BetaSeg and from 68.32% to 69.50% on UroCell. Thus, this construction procedure improves target-class teacher-mask quality on both datasets.

Regarding orientation dependence, we combine three orthogonal views to integrate structural information from different cross-sections. These views are re-slicings of the same prediction volume, not outputs from three separately trained orientation-specific models. Two-dimensional component areas can vary with orientation, so this design does not guarantee invariance to arbitrary rotations. Appendix A.2 discusses orientation dependence and lists the aggregation settings and filtering rules in Table 9.

AuthorsOpenReview ↗

Response to Reviewer 42Yf (Part 3/3) — Q2

Reviewer question — Q2

What experimental evidence establishes that Stage II specifically benefits from the proposed Stage I rather than operating as an independent quality estimator applicable to any segmentation output?

Response

We thank the reviewer for their question about the relationship between the two stages. They are connected through the final segmentation prediction. Stage I uses a prediction-derived teacher SDF to provide geometric constraints for sparsely supervised segmentation. Stage II recomputes local SDFs from the final prediction masks to assess regional segmentation quality. The stages improve segmentation and identify unreliable regions, respectively, providing both semantic predictions and regional quality information.

Stage II is designed as an independently usable quality-assessment module. Its inputs depend only on raw images and prediction masks, without requiring access to internal segmentation features or changes to the segmentation network. Inputs can therefore also be constructed from other segmenters' outputs to train a corresponding quality estimator. This modular design allows the segmenter and quality estimator to be adjusted separately. Once the segmenter is frozen, the estimator is trained using fixed predictions and quality targets.

Our experiments support the estimator's effectiveness on outputs of the proposed segmentation workflow and the contribution of local SDFs to quality prediction. They do not, by themselves, establish a synergistic gain specific to these two stages. The design combines segmentation improvement and regional quality assessment in a sequential workflow whose modules can be configured independently. Sections 3 and 3.2 of the revised manuscript explain how the stages connect and why the quality estimator is trained independently.

Reviewer_cU3BOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Sparse2Geo is a two-stage framework that leverages prediction-derived geometry for (I) semantic segmentation and (II) regional quality assessment in volume electron microscopy (vEM).

In stage I, a 2D sparsely supervised segmentation model is trained and frozen (modified -Seg). This is used to generate full-volume predictions on a chosen target class, which is then converted to a clipped 3D teacher SDF. A newly initialized segmenter is subsequently trained with the original sparse losses plus an auxiliary SDF regression loss and a boundary-band binary cross-entropy consistency loss against the teacher SDF.

In Stage II, predictions from the Stage I model are cropped into 3D sub-volumes, converted into normalized local SDFs, and concatenated with corresponding raw EM patches to train a 3D CNN that regresses the local class-conditional Dice score. The method is evaluated on two public FIB-SEM datasets (BetaSeg and UroCell) against self-supervised, sparsely supervised, and fully supervised baselines. The authors report improvements in segmentation Macro Dice and regional quality-estimation error, with additional ablations examining the effects of the geometric objectives and the SDF representation.

Critical strengths
  1. The proposed approach uses predictions from an initial segmentation model to construct geometric supervision without needing additional dense semantic training data in Stage I. The results show improvements in segmentation performance and boundary quality, including on a second biological dataset.

  2. The Stage II experiments compare the proposed representation against several statistical and learned baselines. The authors also compare raw images, binary masks, and local SDFs under a controlled setting. The SDF-based approach achieves the lowest Macro MAE on BetaSeg and substantially improves quality estimation over the raw-only baseline on UroCell. These comparisons provide useful evidence that the geometric representation can help predict regional segmentation quality.

Critical weaknesses
  1. The paper reports a 4.13 percentage-point improvement in Macro Dice over -Seg on BetaSeg. However, Table 3 shows that most of this improvement comes from adding a dense semantic prediction head and changing the validation protocol. With these modifications already included, the proposed geometric objectives improve Macro Dice from 90.07% to 90.63%, a gain of only 0.56 percentage points.

  2. The much larger improvement on UroCell is promising, but a corresponding ablation separating the contributions of the architectural changes, validation procedure, and geometric objectives is missing. This makes it difficult to assess the actual contribution of the proposed geometric supervision.

  3. A central motivation of the paper is reducing annotation requirements through sparse supervision. However, this applies primarily to Stage I. Training the Stage II quality estimator requires ground-truth regional Dice scores. As specified in Appendix A.3, each accepted region must have valid ground-truth labels for at least 50% of its evaluation core. On BetaSeg, this corresponds to at least 32,768 labeled voxels within each 16 × 64 × 64 core, which is substantially different from the 0.05% sparse-label setting used for Stage I. The authors correctly state that Stage II does not require ground truth at inference. Nevertheless, the additional supervision required to train the quality estimator is important when considering the practical applicability of the complete framework in annotation-scarce settings. The paper does not sufficiently discuss this distinction or quantify the additional annotation cost. It is also unclear how effectively Stage II could be trained when only the sparse annotations used in Stage I are available.

  4. A related issue is that Stage I uses dense validation annotations for checkpoint selection on 45 representative slices (Appendix A.2). Since the validation protocol itself accounts for a substantial part of the improvement over -Seg, the overall annotation requirements and their implications for the comparisons deserve clearer discussion.

  5. All segmentation and quality assessment experiments are single-run point estimates using a fixed random seed (seed 42, Appendix A.2/A.3). Because the true geometric gain in Table 3 is quite small (+0.56% Macro Dice, and ~0.08% difference between Lsdf alone and Lsdf+Lcons), it is difficult to tell whether the synergy between the two geometric losses is statistically significant above training noise.

Questions for authors
  1. Table 3 shows that the proposed geometric objectives improve BetaSeg Macro Dice by 0.56 percentage points over the controlled baseline. Could the authors report variability across multiple random seeds and provide a similar progressive ablation on UroCell?

  2. Annotation requirements of Stage II. Given that Stage II requires ground-truth regional Dice targets computed from substantially annotated evaluation cores, how much additional annotation is needed to train the complete framework? Can the quality estimator be trained effectively using only the sparse annotations available in Stage I, or is a separate collection of densely annotated regions necessary?

  3. The effectiveness of geometric supervision appears sensitive to the choice of target class (k*). In particular, using Class 3 (mitochondria) reduces Macro Dice from 90.07% to 89.73%, whereas Class 2 (granules) improves it to 90.63%. Could the authors explain this variation and clarify how the target class should be selected in practice? Additionally, have the authors considered using multi-channel SDF supervision across all foreground classes instead of selecting a single target class?

6 comments
AuthorsOpenReview ↗

Response to Reviewer cU3B (Part 1/6) — W1/Q1

Reviewer comment — W1

The paper reports a 4.13 percentage-point improvement in Macro Dice over ε-Seg on BetaSeg. However, Table 3 shows that most of this improvement comes from adding a dense semantic prediction head and changing the validation protocol. With these modifications already included, the proposed geometric objectives improve Macro Dice from 90.07% to 90.63%, a gain of only 0.56 percentage points.

Reviewer question — Q1

Table 3 shows that the proposed geometric objectives improve BetaSeg Macro Dice by 0.56 percentage points over the controlled baseline. Could the authors report variability across multiple random seeds and provide a similar progressive ablation on UroCell?

Response

We thank the reviewer for their comments on the contribution of geometric supervision and experimental stability. In earlier experiments with different random seeds, we observed small performance variations. The overall improvement over the official ϵ-Seg implementation reflects the combined effects of the dense semantic prediction head, validation protocol, and geometric supervision. To isolate the contribution of geometric supervision, we added matched comparisons under three configurations on BetaSeg and UroCell. Within each configuration, we keep the prediction head and validation protocol fixed and compare no additional teacher supervision with Full SDF.

DatasetDense headRep.-valNo additional teacher supervisionFull SDFGain (pp)
BetaSeg××86.5087.69+1.19
BetaSeg✓×88.8890.05+1.17
BetaSeg✓✓90.0790.63+0.56
UroCell××56.1160.47+4.36
UroCell✓×69.6970.87+1.18
UroCell✓✓70.7671.44+0.68

Results are Macro Dice (%) on the held-out test set, including the background class. × denotes the official center head or default validation protocol; ✓ denotes the dense head or representative-slice Dice validation (Rep.-val). Gains are in percentage points.

Full SDF improves performance in all six matched comparisons. With the official center head and default model selection, the final segmenter gains 1.19 and 4.36 percentage points on BetaSeg and UroCell, respectively. With the dense head and representative-slice validation, the gains are 0.56 and 0.68 percentage points. These results support an additional contribution from geometric supervision across the evaluated configurations. In the revised manuscript, Section 4.3 distinguishes the overall improvement from the incremental contribution of geometric supervision, and Table 4 reports the matched comparisons.

AuthorsOpenReview ↗

Response to Reviewer cU3B (Part 2/6) — W2

Reviewer comment — W2

The much larger improvement on UroCell is promising, but a corresponding ablation separating the contributions of the architectural changes, validation procedure, and geometric objectives is missing. This makes it difficult to assess the actual contribution of the proposed geometric supervision.

Response

We thank the reviewer for this suggestion. We added a progressive ablation on UroCell and report it alongside the corresponding BetaSeg results. Starting with the final segmenter using the official center head and default model selection, we successively add Full SDF, the dense semantic prediction head, and representative-slice Dice validation.

DatasetFull SDFDense headRep.-valMacro Dice (%)Incremental gain (pp)
BetaSeg×××86.50—
BetaSeg✓××87.69+1.19
BetaSeg✓✓×90.05+2.36
BetaSeg✓✓✓90.63+0.58
UroCell×××56.11—
UroCell✓××60.47+4.36
UroCell✓✓×70.87+10.40
UroCell✓✓✓71.44+0.57

Results are Macro Dice (%) on the held-out test set, including the background class. Each incremental gain is the difference from the preceding row within the same dataset, in percentage points.

On UroCell, Full SDF increases Macro Dice from 56.11% to 60.47%. Adding the dense semantic prediction head and representative-slice validation further increases it to 70.87% and 71.44%, respectively. BetaSeg follows the same trend. This ablation shows the successive improvements in the complete framework. The dense head provides a large gain on UroCell, while geometric supervision also improves performance with the official center head and default model selection. We report the progressive ablation for both datasets in Appendix B.2 (Table 13).

AuthorsOpenReview ↗

Response to Reviewer cU3B (Part 3/6) — W3/Q2

Reviewer comment — W3

A central motivation of the paper is reducing annotation requirements through sparse supervision. However, this applies primarily to Stage I. Training the Stage II quality estimator requires ground-truth regional Dice scores. As specified in Appendix A.3, each accepted region must have valid ground-truth labels for at least 50% of its evaluation core. On BetaSeg, this corresponds to at least 32,768 labeled voxels within each 16 × 64 × 64 core, which is substantially different from the 0.05% sparse-label setting used for Stage I. The authors correctly state that Stage II does not require ground truth at inference. Nevertheless, the additional supervision required to train the quality estimator is important when considering the practical applicability of the complete framework in annotation-scarce settings. The paper does not sufficiently discuss this distinction or quantify the additional annotation cost. It is also unclear how effectively Stage II could be trained when only the sparse annotations used in Stage I are available.

Reviewer question — Q2

Annotation requirements of Stage II. Given that Stage II requires ground-truth regional Dice targets computed from substantially annotated evaluation cores, how much additional annotation is needed to train the complete framework? Can the quality estimator be trained effectively using only the sparse annotations available in Stage I, or is a separate collection of densely annotated regions necessary?

Response

We thank the reviewer for their question about annotation requirements. Stage I trains the segmenter using sparse semantic labels. In Stage II, local SDFs are derived from prediction masks, but the regional Dice targets used to train the quality estimator are computed from manual annotations. The quality supervision required by the original Stage II experiments should therefore be distinguished from the sparse supervision used in Stage I.

To examine whether the existing Stage I sparse annotations can be reused, we added experiments with extremely sparse quality supervision on BetaSeg and UroCell. These experiments reuse the candidate region pool screened using reference labels. Regional Dice targets for estimator training and validation-based model selection are computed only at the existing Stage I sparsely annotated points.

Within each dataset, both supervision settings are evaluated on the same test regions against regional Dice computed using the original evaluation protocol. The table compares Raw and Raw + local SDF under the original and extremely sparse supervision settings.

DatasetQuality supervisionInputMacro MAE ↓Macro ρ ↑AUPRC <0.90 ↑AUPRC <0.80 ↑
BetaSegOriginalRaw0.03080.59170.88940.3604
BetaSegOriginalRaw + local SDF0.02430.80000.96170.6814
BetaSegExtremely sparseRaw0.05510.36380.90970.1296
BetaSegExtremely sparseRaw + local SDF0.05670.67070.94580.2127
UroCellOriginalRaw0.22760.31250.89390.8342
UroCellOriginalRaw + local SDF0.14770.74210.97450.9487
UroCellExtremely sparseRaw0.32620.14750.88640.8223
UroCellExtremely sparseRaw + local SDF0.26390.40850.90110.8828

The original-setting results are from the paper; the extremely sparse setting is a new experiment. Both settings use the same 597 test regions on BetaSeg and 600 on UroCell. Bold indicates the better result within each dataset and supervision setting.

The extremely sparse setting reuses existing Stage I manually annotated points to construct quality targets. For validation regions, the median numbers of annotated points per evaluation core are 31 on BetaSeg and 9 on UroCell. The original setting requires at least 32,768 and 4,096 valid labeled voxels per core, respectively.

Local SDFs improve quality ranking and low-quality-region detection on both datasets under both supervision settings. Existing sparse annotations can therefore be used to train the quality estimator, although overall performance remains below that of the original supervision setting. Appendix A.3 details the original quality-supervision protocol, and Appendix C.2 describes the extremely sparse setting and reports the comparison (Table 19).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their time and constructive feedback.

Our point-by-point responses refer to the relevant sections and tables in the revised manuscript, where the changes are highlighted in blue. We hope that our clarifications and additional analyses will resolve the remaining concerns and strengthen the case for acceptance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.