Deep Active Lesion Segmentation

Ali Hatamizadeh; Assaf Hoogi; Brian Wilcox; Daniel Rubin; Debleena Sengupta; Demetri Terzopoulos; Wuyue Lu

REVIEW 3 major objections 5 minor 1 cited by

Deep Active Lesion Segmentation

T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Deep Active Lesion Segmentation couples a CNN encoder-decoder with a level-set active contour model, using learned per-pixel parameter maps to initialize and tune the contour, and reports significantly more accurate lesion boundaries than…

desk verdict Solid CNN+level-set hybrid with a new multi-organ dataset, but the per-organ significance claims exceed what the test sizes can support. read the letter →

arxiv 1908.06933 v4 pith:LCHGXQTG submitted 2019-08-19 eess.IV cs.CVcs.LG

Ali Hatamizadeh , Assaf Hoogi , Debleena Sengupta , Wuyue Lu , Brian Wilcox , Daniel Rubin , Demetri Terzopoulos This is my paper · ORCID

classification eess.IVcs.CVcs.LG

keywords lesionsegmentationactivecontourmodelslevelsetsconvolutionalneuralnetworksmedicalimageper-pixelparametermapsmultiorgandatasetMRandCTimaging

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces a fully automatic framework that combines a convolutional neural network with a level-set active contour model for lesion segmentation in medical images. The network learns an initial probability map and per-pixel parameter maps that initialize and guide the contour evolution, replacing manually tuned scalar weights with spatially varying ones. The authors claim this yields significantly more accurate boundaries than conventional CNNs and can reliably segment lesions of various sizes. They report superior Dice, Hausdorff distance, and boundary F-measure results across brain, liver, and lung lesions in MR and CT, with the largest advantages appearing when training data are small. If the claim holds, it would remove two practical obstacles in medical image segmentation: manual contour initialization and parameter tuning.

What carries the argument

The load-bearing mechanism is a per-pixel-parameterized level-set energy functional, a generalization of the Chan-Vese active contour energy in which the scalar constants $\lambda_1$ and $\lambda_2$ become functions $\lambda_1(x,y)$ and $\lambda_2(x,y)$ over the image domain. An encoder-decoder CNN with dense blocks and a multiscale dilated-convolution block produces the initial probability map $Y_{\mathrm{prob}}$; the Transformer converts it into the signed distance map that initializes the contour, and the exponential formulas above convert it into the parameter maps. During evolution, these local weights let the contour bend precisely to irregular, low-contrast lesion boundaries, which is the mechanism the paper credits for its boundary accuracy.

What would settle it

Run DALS and its CNN backbone on repeated random splits or k-fold cross-validation of each MLS organ subset; if the Dice advantage shrinks or the Wilcoxon p-value rises above 0.05 in any organ, the claimed significant boundary improvement is not supported.

Watch

Extended reading notes

Core claim

The central claim is that lesion segmentation can be made both fully automatic and boundary-precise by letting a CNN supply the initial contour and the spatially varying weights of a level-set active contour model. Concretely, the CNN outputs a probability map $Y_{\mathrm{prob}}(x,y)$; a transformer converts it into a signed distance map $\varphi(x,y,0)$ that seeds the contour, and the same map defines per-pixel parameters $\lambda_1(x,y)=\exp\big((2-Y_{\mathrm{prob}})/(1+Y_{\mathrm{prob}})\big)$ and $\lambda_2(x,y)=\exp\big((1+Y_{\mathrm{prob}})/(2-Y_{\mathrm{prob}})\big)$ in the energy functional. The contour then evolves by the Euler-Lagrange PDE of a localized Chan-Vese energy, so each boundary point carries its own interior and exterior weighting. On the paper's Multiorgan Lesion Segmentation dataset, DALS reports higher Dice and boundary metrics than U-Net, a manually initialized level-set model, and its own CNN backbone, and the authors state that the improvements are statistically significant.

Load-bearing premise

The load-bearing premise is that the 10-percent held-out test subsets of the MLS dataset, as few as nine lung images, are representative enough to support the reported significance given that DALS's confidence intervals overlap with its own CNN backbone on several organs.

Editorial extensions

If this is right

Lesion segmentation becomes fully automatic: no manual contour initialization and no hand-picked active-contour weights, because the network supplies both.
Boundary quality should improve most where intensity contrast is low or the lesion edge is irregular, exactly the cases pure CNNs handle poorly.
The same architecture can be retrained for other organs and modalities without changing the level-set machinery.
The method should remain usable with small training sets, because the CNN only needs to produce a coarse localization and parameter maps rather than a pixel-perfect boundary.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural ablation not reported in the paper: replace the learned $\lambda_1(x,y)$ and $\lambda_2(x,y)$ maps with their spatial means while keeping the same initialization; if boundary metrics degrade but Dice does not, the parameter maps are doing the boundary work.
Because the exponential transforms amplify $Y_{\mathrm{prob}}$ values near 0 and 1, a poorly calibrated probability map could produce extreme local weights; measuring calibration per organ would predict where DALS is most likely to fail.
The same pipeline should transfer to non-radiology segmentation tasks with irregular boundaries, such as ultrasound lesions or cells, since the level-set evolution is agnostic to image modality.
The significance claim could be stress-tested by repeated cross-validation; the reported 10 percent test splits leave as few as nine lung images, and several confidence intervals overlap with the backbone CNN.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is a solid, unglamorous hybrid method—CNN gives an initial probability map and per-pixel weight maps for a level-set ACM—plus a new multi-organ lesion dataset. The paper is honest about what it's adding, but the statistical support for the headline claim is not there.

The genuinely new bit is the generalization of the scalar Chan-Vese lambdas to per-pixel parameter functions lambda1(x,y), lambda2(x,y), computed from the CNN probability map via an exponential transform. The multiscale dense encoder-decoder with dilation rates 2-16 is a reasonable architecture. The MLS dataset (brain MR, lung CT, liver MR, liver CT) is a useful contribution, and the source code is public.

The ablation consistently shows DALS beating its own backbone on mean Dice for every organ, and the boundary plots show the level-set refinement does something. That's real evidence.

Now the soft spots, in order of severity. The significance claims don't hold up. With an 85/10/5 split, the lung CT test set is about 9 images. A paired Wilcoxon test on n=9 cannot produce p<0.001; the smallest two-sided p is roughly 0.0039. Either the test was pooled across organs, in which case it doesn't support the 'all datasets' sentence, or it's just wrong. Either way, 'significantly more accurate' is an overstatement. The confidence intervals in Table 2 overlap between DALS and the backbone for lung CT and liver CT, so the mean gain is not clearly significant on those organs. The manual level-set baseline is unsupervised and initialized by hand, so it's not a fair comparison. Also, the lambda maps in Eq. 3 are not learned network outputs; they are deterministic functions of Y_prob. Calling them 'learned parameter maps' is misleading. Finally, no dataset release, no commit hash, and the ACM evolution details (iterations, convergence criteria) are missing.

The core idea is coherent and the code makes reimplementation feasible. But the statistical evidence as reported is too thin for the claims. I'd send it out for review with a request to fix the stats and re-analyze with proper per-organ tests and ideally multiple splits. The paper is useful for anyone working on combining CNNs with active contours; just don't trust the significance asterisks.

Referee Report

3 major / 5 minor

Summary. The paper introduces Deep Active Lesion Segmentation (DALS), a framework combining a fully convolutional encoder-decoder with a level-set active contour model (ACM). The CNN predicts a lesion probability map that initializes the signed distance function and also produces per-pixel parameter maps λ1(x,y) and λ2(x,y) used in a localized Chan-Vese style energy functional. The method is evaluated on a new Multiorgan Lesion Segmentation (MLS) dataset with brain MR, lung CT, liver CT, and liver MR images, comparing against U-Net, a manually initialized level-set ACM, and the authors' CNN backbone. The authors report improved Dice, Hausdorff distance, and Boundary F1 scores, and claim statistically significant superiority via Wilcoxon paired tests.

Significance. If the reported results hold, the work makes a useful contribution to medical image segmentation by demonstrating a practical method for learning per-pixel ACM energy parameters from a CNN, moving beyond scalar or hand-tuned λ values. The release of source code and the introduction of the MLS dataset are positive aspects that support reproducibility and further research. The central architectural idea is plausible and the paper includes a comparison against a strong U-Net baseline, which lends credibility to the approach.

major comments (3)

[Section 4, Table 2] The statistical support for the headline claim of 'significantly more accurate boundaries' is not established. The paper reports Wilcoxon paired test p-values of p<0.001 against U-Net and the level-set baseline, and p<0.005 against the CNN backbone, but with the stated split (85/10/5) the lung CT test set consists of roughly 9 images and the liver CT test set of roughly 11 images. For a paired Wilcoxon test with n=9, the smallest attainable two-sided p-value is 2/2^9 = 0.0039, so p<0.001 cannot be correct if the test was performed per organ. If the test was pooled across organs, then it does not independently support the per-dataset claim 'in all datasets' made in the same paragraph. In addition, the confidence intervals reported in Table 2 for DALS and the CNN backbone overlap on lung CT (Dice 0.869 ± 0.092 vs 0.822 ± 0.094) and on liver CT (0.846 ± 0.081 vs 0.801 ± 0.159), which further undermines the per-organ significance claim. The authors must specify the exact test procedure, provide per-organ p-values or confidence intervals of the differences, and correct the reported p-values if they are unattainable.
[Section 3, data split] The evaluation rests on a single 10% test split per organ, with no repeated resampling, cross-validation, or description of patient-level splitting. For the lung CT subset, the test set is only about 8–9 images, which is too small to support general claims about 'all datasets' without additional evidence. The paper should report the exact number of test images per organ, clarify whether images from the same patient appear across train/test splits, and ideally provide leave-one-out or repeated random sub-sampling results to demonstrate that the reported gains are not an artifact of a particular split.
[Section 4, Algorithm Comparison] The manually initialized level-set ACM baseline is unsupervised: it does not use the training labels, while DALS is fully supervised. Comparing a supervised method against an unsupervised baseline does not isolate the contribution of the per-pixel parameter functions, because the baseline does not receive any task-specific tuning. A fairer ablation would either train the same backbone with scalar λ constants (for example, by integrating the level-set evolution into the training loop, as in the cited deep level-set works) or at least tune a scalar λ on the validation set and then compare. Without such a control, the claim that the learned parameter maps are responsible for the improvement is not convincingly supported.

minor comments (5)

[Abstract] The sentence 'we introduce Deep Active Lesion Segmentation (DALS), a fully automated segmentation framework that leverages...' contains a grammatical issue ('that leverages' is split from its antecedent by 'framework'); consider revising for readability.
[Section 2.3, Eq. (3)] The exponential forms for λ1 and λ2 are presented without motivation for the specific formulas. A brief explanation of why the chosen exponential functions are appropriate, or a reference to the prior work that used them, would help readers understand the design choice.
[Table 2] In the Liver CT row, the Hausdorff CI for the CNN Backbone is listed as 1.6, which is an order of magnitude larger than all other CI values in the table and appears to be a typographical error. Please check whether this should be 0.16 or another value.
[Section 4, statistics] The Wilcoxon test description does not state whether the test is one-sided or two-sided, how ties are handled, or whether any multiple-comparison correction was applied across the four organs and three comparisons. Please provide these details.
[Figure 2] The tensor dimensions shown in Figure 2 are not fully consistent: for example, the input is [256,256,1] but the first feature block is [128,128,64], and the concatenated feature maps before the multiscale dilation block have different listed sizes. Clarify the notation to show how these dimensions are obtained.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is evaluated against held-out ground truth and an external U-Net baseline, and the CNN-to-ACM coupling is an architectural choice rather than a fitted-input-as-prediction design.

full rationale

The paper's central claim is that DALS, which combines a CNN backbone with a level-set ACM, yields more accurate lesion boundaries than U-Net, a manually initialized level-set ACM, and its own CNN backbone. The evaluation is performed on held-out test images against expert ground truth, and the comparisons include an external baseline (U-Net). The lambda parameter maps in Eq. (3) are computed from the CNN probability map Yprob, which was trained with a Dice loss against ground truth; this means the ACM refinement is coupled to the CNN prediction rather than being fully independent. However, this is a design coupling, not a definitional circularity: the paper does not define the segmentation accuracy in terms of the fitted parameters, nor does it rename a fitted quantity as a prediction. The extension of Hoogi et al. [3] is cited for the form of the lambda functions, but the formula is stated explicitly in Eq. (3), so the argument does not reduce to an unverified self-citation. The reported Wilcoxon p-values may be statistically questionable given the small per-organ test sizes, but that is a correctness or statistical-power concern, not evidence of circularity. No load-bearing step in the derivation chain reduces by construction to its own inputs, and no fitted parameter is relabeled as a prediction. Therefore, the paper receives a circularity score of 0.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard level-set machinery, an unverified convergence assumption, reliable manual ground truth, and a hand-chosen formula for converting CNN probabilities into active contour weights. No new physical entities or free constants beyond mu=0.1 and the lambda-map formula are introduced.

free parameters (2)

mu (length penalty) = 0.1
Set by hand in the energy functional (Eq. 1) to penalize contour length; the paper states mu=0.1 without tuning analysis.
lambda map exponential formula = exponents (2-Y)/(1+Y) and (1+Y)/(2-Y)
Eq. (3) is an ad hoc choice for converting CNN probability into per-pixel weights; no derivation or sensitivity analysis is given.

assumptions (5)

standard math Chan-Vese uniform internal energy model
Eq. (1) assumes intensity means m1 and m2 inside and outside the contour characterize the regions; this is a standard model from [1].
standard math Lankton localized region integration
The windowed energy Ws is taken from [7]; no proof or parameter sensitivity is provided.
domain assumption Numerical time integration of the level-set PDE converges to a useful minimum
The paper states the contour is evolved by finite-difference time integration but gives no iteration count, time step, or convergence criterion.
domain assumption Manual ground truth annotations are accurate and consistent
All metrics are computed against expert manual segmentations; no inter-observer variability is reported.
ad hoc to paper The CNN probability map Yprob is sufficiently accurate to initialize and weight the ACM
DALS's core design assumes the encoder-decoder output is good enough that ACM refinement improves rather than degrades segmentation; this is empirically validated only on the small test sets.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Deep Active Lesion Segmentation." pith.science (2026). https://pith.science/paper/LCHGXQTG

@misc{pith2026190806933,
  author       = {Pith},
  title        = {Pith review of: Deep Active Lesion Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LCHGXQTG}},
  note         = {Machine review of arXiv:1908.06933}
}

abstract

Lesion segmentation is an important problem in computer-assisted diagnosis that remains challenging due to the prevalence of low contrast, irregular boundaries that are unamenable to shape priors. We introduce Deep Active Lesion Segmentation (DALS), a fully automated segmentation framework for that leverages the powerful nonlinear feature extraction abilities of fully Convolutional Neural Networks (CNNs) and the precise boundary delineation abilities of Active Contour Models (ACMs). Our DALS framework benefits from an improved level-set ACM formulation with a per-pixel-parameterized energy functional and a novel multiscale encoder-decoder CNN that learns an initialization probability map along with parameter maps for the ACM. We evaluate our lesion segmentation model on a new Multiorgan Lesion Segmentation (MLS) dataset that contains images of various organs, including brain, liver, and lung, across different imaging modalities---MR and CT. Our results demonstrate favorable performance compared to competing methods, especially for small training datasets. Source code : $\text{https://github.com/ahatamiz/dals}$

Figures

Figures reproduced from arXiv: 1908.06933 by the authors.

**Figure 2.** The proposed DALS architecture. DALS is a fully automatic framework [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Box and whisker plots of: (a) Dice score; (b) Hausdorff distance. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Comparison of the output segmentation of our DALS (red) against the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: (a) Labeled image. (b) Level-set (analogous to scalar [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

End-to-End Boundary Aware Networks for Medical Image Segmentation
cs.CV 2019-08 conditional novelty 4.0 of 10

A boundary-aware CNN with a dedicated edge stream and edge losses improves tumor-core segmentation on 2D slices of BraTS 2018 compared to U-Net and V-Net, though the evaluation has internal inconsistencies.

Reference graph

Works this paper leans on

11 extracted references · 8 canonical work pages · cited by 1 Pith paper

[1]

IEEE Transactions on Image Processing 10(2), 266--277 (2001)

Chan, T.F., Vese, L.A.: Active contours without edges. IEEE Transactions on Image Processing 10(2), 266--277 (2001)

work page 2001
[2]

arXiv preprint arXiv:1905.12120 (2019)

Hatamizadeh, A., Hosseini, H., Liu, Z., Schwartz, S.D., Terzopoulos, D.: Deep dilated convolutional nets for the automatic segmentation of retinal vessels. arXiv preprint arXiv:1905.12120 (2019)

arXiv 2019
[3]

IEEE Transactions on Medical Imaging 36(3), 781--791 (2017)

Hoogi, A., Subramaniam, A., Veerapaneni, R., Rubin, D.L.: Adaptive estimation of active contour parameters using convolutional neural networks and texture analysis. IEEE Transactions on Medical Imaging 36(3), 781--791 (2017)

work page 2017
[4]

In: Proc

Hu, P., Shuai, B., Liu, J., Wang, G.: Deep level sets for salient object detection. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (2017)

work page 2017
[5]

In: Deep Learning in Medical Image Analysis, Lecture Notes in Computer Science, vol

Imran, A.A.Z., Hatamizadeh, A., Ananth, S.P., Ding, X., Terzopoulos, D., Tajbakhsh, N.: Automatic segmentation of pulmonary lobes using a progressive dense V -network. In: Deep Learning in Medical Image Analysis, Lecture Notes in Computer Science, vol. 11045, pp. 282--290. Springer (2018)

work page 2018
[6]

International Journal of Computer Vision 1(4), 321--331 (1988)

Kass, M., Witkin, A., Terzopoulos, D.: Snakes: Active contour models. International Journal of Computer Vision 1(4), 321--331 (1988)

work page 1988
[7]

IEEE Transactions on Image Processing 17(11), 2029--2039 (2008)

Lankton, S., Tannenbaum, A.: Localizing region-based active contours. IEEE Transactions on Image Processing 17(11), 2029--2039 (2008)

work page 2008
[8]

In: Proc

Marcos, D., Tuia, D., Kellenberger, B., Zhang, L., Bai, M., Liao, R., Urtasun, R.: Learning deep structured active contours end-to-end. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 8877--8885 (2018)

work page 2018

Show all 11 references

[9]

In: LNCS vol

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: LNCS vol. 9351 (Proc. MICCAI). pp. 234--241 (2015)

2015
[10]

, " * write output.state after.block = add.period write newline

ENTRY address author booktitle chapter edition editor eid howpublished institution journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.se...
[11]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

IEEE Transactions on Image Processing 10(2), 266--277 (2001)

Chan, T.F., Vese, L.A.: Active contours without edges. IEEE Transactions on Image Processing 10(2), 266--277 (2001)

work page 2001

[2] [2]

arXiv preprint arXiv:1905.12120 (2019)

Hatamizadeh, A., Hosseini, H., Liu, Z., Schwartz, S.D., Terzopoulos, D.: Deep dilated convolutional nets for the automatic segmentation of retinal vessels. arXiv preprint arXiv:1905.12120 (2019)

arXiv 2019

[3] [3]

IEEE Transactions on Medical Imaging 36(3), 781--791 (2017)

Hoogi, A., Subramaniam, A., Veerapaneni, R., Rubin, D.L.: Adaptive estimation of active contour parameters using convolutional neural networks and texture analysis. IEEE Transactions on Medical Imaging 36(3), 781--791 (2017)

work page 2017

[4] [4]

In: Proc

Hu, P., Shuai, B., Liu, J., Wang, G.: Deep level sets for salient object detection. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (2017)

work page 2017

[5] [5]

In: Deep Learning in Medical Image Analysis, Lecture Notes in Computer Science, vol

Imran, A.A.Z., Hatamizadeh, A., Ananth, S.P., Ding, X., Terzopoulos, D., Tajbakhsh, N.: Automatic segmentation of pulmonary lobes using a progressive dense V -network. In: Deep Learning in Medical Image Analysis, Lecture Notes in Computer Science, vol. 11045, pp. 282--290. Springer (2018)

work page 2018

[6] [6]

International Journal of Computer Vision 1(4), 321--331 (1988)

Kass, M., Witkin, A., Terzopoulos, D.: Snakes: Active contour models. International Journal of Computer Vision 1(4), 321--331 (1988)

work page 1988

[7] [7]

IEEE Transactions on Image Processing 17(11), 2029--2039 (2008)

Lankton, S., Tannenbaum, A.: Localizing region-based active contours. IEEE Transactions on Image Processing 17(11), 2029--2039 (2008)

work page 2008

[8] [8]

In: Proc

Marcos, D., Tuia, D., Kellenberger, B., Zhang, L., Bai, M., Liao, R., Urtasun, R.: Learning deep structured active contours end-to-end. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 8877--8885 (2018)

work page 2018