acceptodds
Under review as a conference paper at ICLR 2027

Learning Polyline Geometry for Length Measurement in Medical & Biological Images

11 views

Abstract

Object length measurement is a common objective in many biological and medical imaging applications, such as in identifying a tumor's longest axis, tracking larval growth, or measuring duodenal villi height. Conventional deep learning pipelines typically compute length as a byproduct of dense segmentation mask outputs. In this work, we question whether segmentation is even required if length measurement is the primary objective and propose polyline annotations for length measurement as an alternative. A polyline is defined as a sequence of points along an object's principal axis; we posit that it is inexpensive to annotate yet sufficient for measurement. We introduce MeasureNet, a framework that directly predicts polylines using specialized, measurement aware loss functions. We evaluate MeasureNet across five datasets spanning brightfield and fluorescence microscopy, optical coherence microscopy, computed tomography, and H&E histopathology, including on CeDeM, a newly curated duodenal biopsy dataset with application to Celiac disease detection. Across these measurement datasets, MeasureNet trained with polyline annotations outperforms SoTA segmentation models trained on dense segmentation masks, when controlled for the annotation budget, reducing length measurement Mean Absolute Error (MAE) by up to 30 points. Furthermore, MeasureNet outperforms the closest length measurement baselines across all datasets in measurement as well as localization, yielding up to 12.87 point MAE and 11.56 point mAP gains.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Medical Images, Histopatholgy Images, Polyline Detection, Measurement, Celiac Disease

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_3APkOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

In this paper, the authors question whether dense segmentation is necessary for object length measurement and propose MEASURENET, a framework that directly predicts object polylines using measurement-aware losses. The authors also introduce CeDeM, a new duodenal biopsy dataset with polyline annotations for villus and crypt measurement. Across four public datasets and the proposed CeDeM dataset, MEASURENET consistently achieves strong performance in both polyline localization and length measurement. The experiments further suggest that direct polyline-based measurement can be more annotation-efficient and achieve lower measurement error than conventional pipelines that rely on dense segmentation masks as an intermediate representation.

Critical strengths
  1. The research question is meaningful. Length measurement is important across biological and clinical domains. Conventional methods usually obtain polylines through intermediate segmentation rather than directly predicting them, while segmentation mask annotation is expensive.
  2. The CeDeM dataset is a valuable contribution. It contains expert pathologist annotations and addresses a clinically meaningful measurement task.
  3. The empirical coverage is strong. The evaluation includes both the proposed dataset and multiple public datasets, spanning CT, histopathology, brightfield/fluorescence microscopy, and optical coherence microscopy.
Critical weaknesses
  1. My biggest concern is the novelty. First, prior work has explored direct geometric prediction for object measurement, so this paper is not the first to perform direct measurement. The paper should clarify its differences from existing direct measurement methods. Second, the proposed measurement-aware losses are mainly combinations of existing losses, and the main architectural modification is just replacing the bounding-box head with a polyline head.
  2. The evaluation metric is problematic because it does not enforce one-to-one instance matching. For example, if the ground truth contains two polylines of length 10 but the model predicts only one polyline of length 10, the current metric can still report zero error, which is not appropriate.
  3. There is a mismatch between the annotation-efficiency claim and the training supervision on the public datasets. Across the four public datasets, the polyline labels are derived from segmentation masks rather than manually annotated. Therefore, the experiments do not fully validate the claimed annotation efficiency.
  4. Architectural validation is limited. The proposed framework is evaluated only on DINO-DETR with a Swin-L backbone. It remains unclear whether the proposed polyline representation and measurement-aware losses are architecture-agnostic.
Questions for authors
  1. Were the values of K selected exclusively using the training and validation sets, without using any information from the test set?
  2. Did the authors consider evaluating MEASURENET on existing public datasets with manually traced centerline annotations?
7 comments

We thank the reviewer for finding the research question meaningful, CeDeM a valuable contribution, and the empirical coverage across five datasets and imaging modalities strong. Changes in the revised manuscript are marked in red.

W1: Novelty relative to direct measurement methods, losses and architecture.

We agree that direct geometric prediction and the individual loss functions are not new. Our key novelty lies in the core hypothesis of formulating length measurement as an instance-wise polyline prediction task, rather than in introducing a new backbone or loss function.

Unlike direct scalar regression, our approach predicts the underlying geometry, allowing clinicians to visually verify and correct individual points when necessary. This is critical in applications such as healthcare, where the cost of errors can be high. Unlike existing line detectors (LETR, RANK-LETR) and general polyline detectors (YOLinO, MapTR), MeasureNet explicitly optimizes measurement accuracy through joint supervision of point localization, total length, and individual segment lengths. Although the architectural modification is subtle, the measurement-aware training objective provides consistent improvements (Table 3). We further demonstrate that sparse polyline annotations achieve competitive measurement accuracy with substantially lower annotation effort across five datasets.

W2: M-MAE does not match instances one-to-one: if the ground truth has two polylines of length 10 and the model predicts only one of length 10, the error is zero.

We agree that M-MAE alone does not penalize missing instances when their lengths match. However, M-MAE evaluates measurement accuracy, while precision, recall and mAP evaluate detection performance. In the given example, M-MAE can be zero, but recall is only 50%. We therefore report both measurement and detection metrics (Tables 1 and 2, and Appendix E). MeasureNet consistently achieves strong performance on both.

W3: On the four public datasets the polyline labels are derived from segmentation masks rather than annotated manually, so the annotation-efficiency claim is not fully validated.

We agree that for the existing datasets, polylines are derived by post-processing the existing segmentation masks, because no polyline-annotated datasets exist in the literature. To support our hypothesis about their effectiveness, we took several steps.

First, annotation efficiency was evaluated independently: two medical experts manually annotated polylines and dense masks for a subset of images across all five datasets, obtaining a 2.54–5.83× reduction in annotation time (Fig. 3(a), Table 5). Second, we visually inspected the polyline annotations and found no apparent errors in the inspected cases; Figures 4 and 5 further illustrate that the polylines adequately capture the underlying structures. Finally, and importantly, we release CeDeM, a dataset natively annotated with polylines by expert pathologists, which provides further validation of their effectiveness.

W4: Architecture-agnostic validation.

Our approach can be adapted to other object detectors by replacing the bounding-box head with a KK-point polyline head and adding the measurement-aware losses. To investigate this, we extended RT-DETR [1] with a ResNet-50 backbone and evaluated it with and without the proposed measurement-aware losses.

RT-DETRKiTS19 M-MAE ↓KiTS19 mAP ↑Drosophila M-MAE ↓Drosophila mAP ↑BBBC010 M-MAE ↓BBBC010 mAP ↑CeDeM MAE-V ↓CeDeM MAE-C ↓CeDeM mAP ↑
Base15.0268.891.7895.7610.0990.5018.9011.9139.40
+ Meas. losses9.7470.301.2595.848.7191.1017.259.8644.30

The measurement-aware losses consistently reduce measurement error on all four datasets while also improving detection performance. We have added this to Appendix F.2.

[1] DETRs Beat YOLOs on Real-time Object Detection, CVPR 2024.

Q1: Were the values of KK selected using only the training and validation sets, without any information from the test set?

Yes. All values of KK were selected using only the training and validation sets. For CeDeM, we evaluated K∈{2,3,4}K \in \{2,3,4\} on the validation set and selected K=3K=3 based on measurement accuracy. For the public datasets, KK was selected based on the polyline approximation error relative to the full skeleton, using validation data. The test sets were used only for the final evaluation after fixing KK.

Q2: Evaluation on public datasets with manually traced centerlines.

The closest public dataset we could identify is DeepTangle, which provides manually traced centerlines of C. elegans for evaluation, but no manually annotated training split. We evaluated MeasureNet and LETR (our closest baseline), both trained on BBBC010, directly on DeepTangle without fine-tuning. MeasureNet achieves an mAP of 57.3 versus 18.9 for LETR, and an M-MAE of 9.5 versus 13.3. If the reviewer has another relevant dataset in mind, we would be happy to evaluate it and provide additional comparisons.

Thanks for the authors' responses, which have addressed most of my concerns. I appreciate the clarification that the key novelty of this work lies in the core hypothesis of formulating length measurement as an instance-wise polyline prediction task, rather than in introducing a new backbone or loss function. I suggest that the authors further highlight this core hypothesis in the revised manuscript.

However, I still have one remaining concern regarding the methodological novelty. Specifically, I am curious about the distinction between the proposed MEASURENET and existing geometric detection methods adapted for length measurement by incorporating additional measurement-aware loss terms. Could the authors further clarify what unique advantages arise from the combination of the instance-wise polyline prediction formulation and the proposed measurement-aware losses, which may not be readily achieved by adapting existing geometric detection approaches?

Reviewer_SfqfOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
3(+1)
Presentation
2
Soundness
1(−1)
Read review
Paper summary

The authors propose an object detection and annotation approach for objects that are annotated via piecewise-linear functions. The proposed approach extends DINO-DETR to predict a set of class-label and point-sets representing the polylines with the goal to predict accurate lengths of the objects as measured by those lines. Training then uses a combination of the object detection, classification (focal loss) and task-specific losses (l1, earth-mover distance, total line-length)

In their evaluation, Eart Mover distance performed best under the matched mean absolute error measurements and comparison to other approaches shows that the presented work improves over the state of the art

Critical strengths

The results show that the new approach works well and it makes sense that directly predicting the line and training it towards accurate length predictions will improve over the clear segmentation baseline.

Critical weaknesses

The main weakness i can identify is that the paper uses conflicting definitions of a polyline and the evaluation is not consistent.

A polyline consists of line segments (p_i,p_{i+1}) and the lengths are therefore the sum of lengths of those sequences. The paper uses these definitions for parts of the losses, (1), (4) and (5) but then evaluates using matched point-sets and the chamfer distance (Table 4). The chamfer distance treats the lines as unordered point clouds, so the two-segment lines {(p_1,p_2),(p_2,p_3)} and {(p_1,p_3),(p_3,p_2)} are the same to them. As a result, predicting unordered point sets is significantly easier than predicting ordered polylines. The earth mover distance (3) is not a sensible metric on polylines, since it is invariant to the geometric content of the ordering, directly conflicting the authors statement: "The localization loss supervises each polyline point independently but does not enforce the overall polyline geometry".

Indeed, (1) is much more sensitive to permutations of the predicted points than (3), so the increased performance in Table 4 is likely a result of using the chamfer distance to match ground truth with predicted point sets, which is something that Earth Mover can help with. I wouldn't take this as indication that the line predictions of models trained without earth mover are worse, but the matching procedure favours Earth Mover.

Table 1 follows the common bad practice of marking best results in bold even if they are not statistically significantly better than the closest comparison (Drosophila mAP). This reviewer would prefer if all results with overlapping uncertainties would be marked in bold, equally. It is appreciated though, that at least the closest result is underlined.

Questions for authors

Why did the Authors choose chamfer distance to match ordered point sets?

4 comments

We thank the reviewer for the careful reading and the constructive comments. Changes in the revised manuscript are marked in red.

W1: Lengths and some losses treat a polyline as ordered segments, but Chamfer matching and EMD treat it as an unordered point set. The EMD gain in Table 4 may therefore come from Chamfer-based matching rather than from better ordered polylines.

We agree that Chamfer distance and EMD are insensitive to point ordering. To address the concern that Chamfer-based matching may favor EMD, we additionally evaluate order-sensitivity using Sequence Optimal Sub-Pattern Assignment (SOSPA) [1]. SOSPA is an order-aware polyline distance metric that, unlike Chamfer distance, distinguishes polylines with identical point sets but different point orderings. We replace Chamfer with SOSPA for evaluation time for instance matching, while keeping training procedure unchanged. The results are shown below.

MatchingLossBBBC010 M-MAE ↓CeDeM MAE-V ↓CeDeM MAE-C ↓CeDeM MAE-R ↓
Chamferℓ1\ell_112.7913.908.830.578
ChamferEMD10.5912.928.800.572
SOSPAℓ1\ell_112.7913.068.760.559
SOSPAEMD10.6012.988.500.550

EMD maintains lower measurement error than ℓ1\ell_1 under SOSPA matching across both datasets and all reported metrics. In particular, BBBC010 M-MAE improves from 12.79 to 10.60. These results indicate that the observed benefits of EMD are not solely attributable to Chamfer-based matching. We further inspected predictions on CeDeM (K=3K=3) and BBBC010 (K=7K=7) and found no arbitrary point permutations. We identified 45 reversed predictions in CeDeM and 3 in BBBC010, these trace the same geometric polylines as the ground truth, differing only in traversal direction.

We clarify that Chamfer/SOSPA is used for instance matching during evaluation, whereas M-MAE compares lengths computed by summing consecutive polyline segments. Importantly, M-MAE is also sensitive to incorrect point ordering, since arbitrary permutations generally change the sum of consecutive segment lengths and therefore can increase measurement error. Complete reversals, however, preserve the polyline geometry and length. EMD provides permutation-invariant spatial correspondence and discourages point collapse, while the ordered localization and part-length losses supervise polyline geometry. We have added the discussion to Appendix F.3.

[1]: Beyond Chamfer Distance: Granular Order-aware Evaluation Metric For Online Mapping

W2: Table 1 bolds the best mean even when it is not significantly better than the runner-up (e.g. Drosophila mAP); all results with overlapping uncertainties should be bolded equally.

Thank you for bringing this to our notice. We have updated the manuscript so that statistically significant cases are additionally marked with a *.

Q1: Why Chamfer distance to match ordered point sets?

We use Chamfer distance for instance matching, not for evaluating polyline geometry, because the annotation order can be reversed in some cases. After matching, M-MAE compares lengths computed from consecutive polyline segments; thus, incorrect point ordering that changes the length is penalized.

Thanks for the answer.

I think the argument that polylines are identical under reversal but l1 is not, is a good argument and should not be burried in the appendix. I think it is still a bad argument for EMD and more an argument against l1: EMD is invariant (bad), but l1 requires guessing the line order (worse). But even though EMD is invariant, order is still enforced by (4),(5) and so your total loss is still sensitive to permutations other than reversal. This would be a much better statement than "[l1] does not enforce the overall polyline geometry".

I think using chamfer distance is still a poor choice for phase 1, but probably fine for most of your datasets (overlapping organisms could look almost the same under chamfer but ordering would make it clear what matches). I would name it as a limitation, but it is a minor point.

We thank the reviewer for the constructive suggestions. We will emphasize reversal invariance in the main text and replace the statement that ℓ1\ell_1 "does not enforce the overall polyline geometry" with a clarification that EMD provides order-invariant alignment, while ordered localization and measurement-aware losses jointly constrain polyline structure. We will also acknowledge the limitation of Chamfer-based matching due to its insensitivity to point ordering. Do let us know if there are any remaining concerns we should address. If not, we would be grateful if you could reconsider your score.

Reviewer_SfqfOpenReview ↗

raised score

Thank you for your efforts.

In my book, this resolves my issues with the article and your reply to reviewer 3APk also resolved my latent novelty concern. It solves a real problem and shows it well. Also, it is a nice little idea and i like nice little ideas.

Compared to other papers in my batch its text does not seem to be dominantly AI generated and humans ensured that it reads decently well to a human. It is sad to point this out as specific strength, but in retrospect I have to distinguish it by what distinguishes it from other works.

It is the best paper in my batch and on a 10 point scale i would score it an 8. This is high enough, so I raised the score.

Reviewer_ok5wOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies direct length measurement in medical and biological images using sparse polylines instead of dense object masks. MEASURENET adapts DINO-DETR with a Swin backbone to predict a set of polylines, trained with localization and classification losses together with Earth Mover, total-length, and segment-length losses. Experiments cover five datasets, including the new CeDeM duodenal biopsy dataset. The evaluation compares measurement and localization accuracy, and considers both matched annotation time and dense-mask training on all available instances. The results support the practical value of direct polyline measurement, while leaving some questions about representation choices and component attribution.

Critical strengths
  1. A useful measurement objective with explicit attention to annotation effort. Predicting a sparse principal-axis polyline directly addresses the required measurement and reduces the need for dense contour annotation. The annotation-time study and the additional full-mask comparison make the practical tradeoff concrete. The paper also reports the KiTS19 case where dense masks retain an advantage, appropriately qualifying the scope of the benefit.
  2. Comparative evidence across diverse datasets. The five datasets include straight, curved, crowded, and overlapping structures. Comparisons include segmentation and polyline detection methods, with shared backbones and point counts for detection baselines. Measurement and localization metrics, five repeated training runs, and qualitative failure examples provide useful evidence beyond a single benchmark. CeDeM also contributes expert-annotated villus and crypt measurements.
Critical weaknesses
  1. The contributions of individual losses need further validation. Table 3 lacks full-model ablations removing EMD or total-length loss. Please provide these controls and variability across the reported five runs, particularly given the 0.08-pixel EMD–Chamfer difference in villus MAE. A targeted BBBC010 ablation would also help verify the benefits for curved structures.
  2. Boundary supervision and annotation costs need clarification. Please explain how VS/CB predictions contribute to final measurements, whether baselines and ablations receive the same boundary supervision, and whether annotation costs include VS/CB labeling. These details are needed to interpret both the accuracy gains and labeling savings.
Questions for authors
  1. Appendix A describes the point-count selection process. Please further explain why a fixed number of points is used for instances with different curvatures within the same class, and how the selected count balances measurement accuracy and annotation cost. Please also clarify which evaluation metric supports the statement that four points offer no improvement over three on CeDeM.

  2. Existing ablations demonstrate the contribution of the segment-length loss, but do not include full-model variants that separately remove EMD and total-length loss. Please provide these comparisons and conduct targeted ablations on BBBC010, which contains more highly curved structures, to assess their contributions to length measurement and shape preservation. In addition, the villus length errors of EMD and Chamfer in Table 4 differ by only 0.08 pixels. Please report variability across the repeated runs to assess whether this advantage is stable or attributable to training randomness.

  3. Besides the number of points, their placement can also affect length measurement accuracy. Please explain the rationale behind the current point-placement strategy and how it ensures adequate representation of highly curved regions.

  4. CeDeM includes VS/CB prediction heads. Please clarify whether their outputs contribute to the final length calculation and whether the ablations and baselines receive the same boundary supervision. The annotation-budget comparison should also specify whether the cost of annotating VS/CB is included.

3 comments

We thank the reviewer for recognising the practical value of direct polyline measurement and proving constructive comments. Changes in the revised manuscript are marked in red.

W1 / Q2: Ablate EMD and the total-length loss separately (also on the more curved BBBC010) to show their effect on length and shape. Report variability across repeated runs for Table 4

We conducted the ablation removing the measurement losses, Earth Mover loss (LEMDL_{\text{EMD}}), Length loss (LLL_{\text{L}}) and Part Length loss (LPLL_{\text{PL}}), one at a time. The table below provides the details; the same has been added to Appendix F.1. Removing the measurement-specific losses increases M-MAE by up to 3.09 points, with the largest drop coming from the Part Length loss, which ensures good curvature in curved cases.

BaseLEMD\mathcal{L}_{\text{EMD}}LL\mathcal{L}_{\text{L}}LPL\mathcal{L}_{\text{PL}}M-MAE ↓mAP ↑
✓✓✓✓10.5982.23
✓✓✓13.6879.14
✓✓✓12.9280.70
✓✓✓12.8780.93
✓20.3167.37

Regarding Table 4, we have updated Table 4 in the main paper (marked in red) with the standard deviation along with the mean; it is reproduced below for reference. We agree that Chamfer and EMD enforce a similar property, while EMD is slightly better as it uses a bijective mapping that prevents point collapse.

Point-set lossMAE-V ↓MAE-C ↓MAE-R ↓
Chamfer13.10 ± 0.79.0 ± 0.263.6 ± 0.04
Manhattan (ℓ1\ell_1)13.90 ± 0.48.83 ± 0.457.8 ± 0.01
Earth Mover12.92 ± 0.68.80 ± 0.257.2 ± 0.02

W2 / Q4: How do the villus-shoulder (VS) and crypt-border (CB) predictions contribute to the final measurements? Do baselines and ablations receive the same boundary supervision? Does the reported annotation cost include VS/CB labelling?

Contribution of VS/CB: Villus shoulder (VS) and crypt border (CB) predictions are not used to compute the final villus or crypt length measurements. Instead, these annotations provide auxiliary boundary supervision, motivated by pathological practice: when crypt boundaries are not clearly visible, pathologists approximate crypt length using the distance between the VS and CB. In MeasureNet, auxiliary VS and CB prediction heads are jointly trained with the villus and crypt prediction heads. These heads share the decoder and object queries, allowing boundary annotations to provide additional supervision during training without directly affecting the final measurements. We have clarified this in line 275 of the main text.

Dataset and annotation cost: All MeasureNet ablations and baseline models are trained using the same CeDeM annotations, including villi, crypts, VS, and CB. The annotation time reported for CeDeM also includes the annotation of VS and CB, in addition to villi and crypts.

Q1: Why is a fixed number of points used for instances of different curvature within a class, how does the chosen KK balance accuracy and annotation cost, and which metric shows that K=4K=4 is no better than K=3K=3 on CeDeM?

Fixed number of points: The model predicts a fixed number of points KK per instance, which is selected to capture the maximum curvature required to represent each class. For CeDeM, pathologists initially annotate each structure using 2–4 control points, depending on its curvature. These annotations are automatically standardized to K=3K=3 by interpolating a midpoint for two-point polylines or subsampling four-point polylines (see the response to Q3 for details). We select KK based on validation-set performance, evaluating MeasureNet with K∈{2,3,4}K \in \{2,3,4\}; the results are reported below. K=3K=3 achieves the lowest measurement error on all three metrics (MAE-V, MAE-C and MAE-R), while requiring less annotation time than K=4K=4. We have added this table to Appendix A, with the discussion in Appendix F.4 and Table 10.

KKMAE-V ↓MAE-C ↓MAE-R ↓Annot. time (s) ↓
218.109.950.6152.67
312.928.800.5724.05
420.079.480.7446.23

Q3: Besides the number of points, how are the points placed, and how does this strategy represent highly curved regions?

For CeDeM, pathologists place 2–4 control points along each structure's principal axis, with additional points at bends. These are resampled to K=3K=3 equidistant points. For the public datasets, points are sampled along the medial axis at equal arc-length intervals (Drosophila and BBBC010), or placed at the endpoints of the longest chord (KiTS19) or principal axis (DeepBacs).

We select KK per dataset based on curvature, using the validation set (Appendix A). For the highly curved BBBC010 structures, K=7K=7 yields approximately 5% average relative length error against the full skeleton. For CeDeM, K=4K=4 offers no improvement over K=3K=3. Although equal-arc-length sampling may under-represent sharp bends, it provides sufficient representation for our use case, as illustrated in Fig. 5 (column 5).

Thank you for the additional experiments and clarifications. The leave-one-out ablations on BBBC010 address my main concern about the contributions of the individual losses. The comparison of point counts on CeDeM also gives a clear empirical reason for choosing three points. I appreciate the clarification that VS and CB provide auxiliary supervision and that their labeling time is included in the annotation budget. I maintain my recommendation of weak acceptance. For the final version, please qualify the small difference between EMD and Chamfer on villus length and check the standard deviation reported for the EMD ratio error, which differs between Tables 2 and 4.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.