Pith. sign in

REVIEW 3 major objections 4 minor 7 references

Significance of Residual Learning and Boundary Weighted Loss in Ischaemic Stroke Lesion Segmentation

T0 review · 3 major / 4 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 Adding residual connections and a boundary-weighted loss to adversarial training lifts ischaemic stroke lesion Dice to 0.881 penumbra and 0.877 core.

desk verdict Residual connections get a clean test; the boundary-loss claim is confounded by Lovász-Softmax in every BD condition. read the letter →

arxiv 1908.04840 v1 pith:NQ4CIRKC submitted 2019-08-13 eess.IV

Ronnie Rajan , Rachana Sathish , Debdoot Sheet This is my paper · ORCID
classification eess.IV
keywords ischaemicstrokelesionsegmentationresiduallearningboundaryweightedlossadversarialtrainingDicecoefficientSPES-ISLES2015multi-sequenceMRI
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

Accurate delineation of the ischaemic stroke lesion's core and salvageable penumbra matters because thrombolytic treatment decisions depend on their extent. This paper argues that two targeted changes to a deep segmentation network—adding residual connections to the encoder and weighting a 3-pixel boundary band by a factor of 10 in the loss—substantially improve that delineation. On the SPES-ISLES 2015 dataset, the resulting adversarially trained Residual-SUMNet reaches average Dice coefficients of 0.881 for penumbra and 0.877 for core, beating all seven baseline configurations. The largest single gain comes from residual connections for the smaller core region, which suggests the mechanism is preservation of small-structure information through the depth of the network.

What carries the argument

The central machinery is Residual-SUMNet: a fully convolutional segmentation network with a VGG11-like encoder modified so that a residual connection is added after each convolutional block, plus three relativistic discriminators operating on the predicted core, the predicted penumbra, and the pair. Training combines four losses: cross-entropy, Lovasz-Softmax, the adversarial loss from the discriminators, and a boundary-weighted negative log-likelihood in which a 3-pixel-thick boundary band extracted from the ground truth by subtracting the eroded label from the dilated label is weighted by factor 10. The residual connections carry multi-sequence information forward through the network depth, while the boundary term concentrates learning on the lesion edge; together they are what the paper credits for the measured Dice gains.

What would settle it

Run Residual-SUMNet with the proposed losses on the full ISLES 2015 training set or a held-out test set with the same boundary parameters; if the Dice advantage over BL7 shrinks to within fold-to-fold variability, or if reducing the boundary thickness erases the gain, the central claim that boundary weighting drives the improvement would fail.

Watch

Extended reading notes

Core claim

The paper claims that residual connections and a boundary-weighted loss are not minor tweaks but decisive components of ischaemic stroke lesion segmentation. Residual-SUMNet, trained from scratch without ImageNet weights on three MRI sequences (TMax, TTP, DWI), with three relativistic discriminators and a combined loss of cross-entropy, Lovasz-Softmax, adversarial, and boundary-weighted negative log-likelihood, achieves average Dice of 0.881 for penumbra and 0.877 for core on SPES-ISLES 2015. The reported comparisons show residual connections raising core Dice from 0.792 to 0.867 between the non-residual and residual CE-only baselines, and the full proposed configuration outperforming its closest adversarial baseline, BL7, from 0.852/0.865 to 0.881/0.877. The intended conclusion is that these additions systematically improve boundary delineation and small-region segmentation.

Load-bearing premise

The load-bearing premise is that improvements measured in a single three-fold cross-validation on the SPES subset, with boundary thickness and weight fixed at 3 pixels and 10, will generalize to other stroke scans and settings.

Editorial extensions

If this is right

  • If the reported comparisons hold, adding residual connections after each encoder block is a cheap and general way to improve segmentation of small anatomical targets, since the largest gain appears in the small core region.
  • The boundary-weighted loss makes predicted lesion edges track ground-truth boundaries more closely, which is clinically relevant because thrombolytic decisions depend on penumbra extent.
  • Training from scratch with batch normalization on three concatenated MRI sequences can outperform the ImageNet-pretrained baseline, removing a common practical dependency.
  • The best result comes from combining adversarial training with both Lovasz-Softmax and boundary weighting, implying the components are complementary rather than redundant.

Reading between the lines

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

  • The paper leaves implicit that a 3-pixel boundary thickness and weight 10 are optimal; varying these parameters could either enlarge or erase the reported gains, since no ablation is provided.
  • A single three-fold cross-validation on the SPES subset leaves open whether the method transfers to the full ISLES 2015 challenge data or to later editions with different scanners and protocols.
  • The large core-Dice jump from residual connections suggests that depth-driven information loss, not just class imbalance, limits small-lesion segmentation; a testable consequence is that shallower encoders would show smaller residual gains.
  • The same residual-plus-boundary recipe may transfer to other small-structure segmentation tasks, such as tumour or haemorrhage delineation, but that transfer is the authors' implicit hope rather than an experimentally established claim.
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.

Referee Report

3 major / 4 minor

Summary. The paper extends SUMNet, an adversarially trained network for ischaemic stroke lesion segmentation, by adding residual connections in the encoder and a morphological boundary-weighted loss (BD) combined with cross-entropy (CE), Lovász-Softmax (LS), and adversarial losses. The authors report average Dice coefficients of 0.881 for penumbra and 0.877 for core on the SPES-ISLES 2015 dataset, and attribute the improvement to residual connections and the boundary-weighted loss. The evaluation uses three-fold cross-validation with eight baselines (BL1–BL7 and Proposed) that progressively add residual connections, adversarial training, and the LS+BD losses.

Significance. If the claimed improvements are reproducible and correctly attributed, the work would offer a practical recipe for improving stroke lesion delineation, which is clinically important for treatment decisions. The paper's ablation structure is a strength for the residual-connection component: the comparisons BL5 vs BL1 and BL7 vs BL3 isolate the effect of residual connections cleanly, and the reported core Dice gains (0.792 to 0.867 and 0.730 to 0.865) are substantial. However, the paper's central claim that the boundary-weighted loss 'significantly' improves segmentation is not supported by the experiments as reported, because BD is never added independently of LS. The absence of error bars or significance tests further weakens the strength of the conclusions. With additional targeted ablations and statistical reporting, the contribution could become solid; in its current form, the attribution of the gains to BD specifically is not established.

major comments (3)
  1. [Section 2 and Table 1] The boundary-weighted loss (BD) is confounded with the Lovász-Softmax (LS) loss. In every comparison that the authors cite as evidence for the BD contribution (BL2 vs BL1, BL4 vs BL3, BL6 vs BL5, and Proposed vs BL7), both LS and BD are added together. There is no condition that adds BD alone to CE, nor a condition that adds LS alone. Since LS is itself a loss designed to maximize overlap (Dice/Jaccard), the observed improvements in Table 1 could be entirely due to LS rather than the boundary weighting. The text in Section 3 ('the weighted boundary loss improves the segmentation along the boundary') is therefore an untested attribution. The authors should add, at minimum, a CE+BD-only condition and ideally a CE+LS-only condition for the same network settings to separate the two effects.
  2. [Section 3 and Table 1] The paper uses the word 'significantly' repeatedly (abstract, Section 3, and conclusion) to describe improvements, but reports only average Dice coefficients over three folds with no standard deviations, no confidence intervals, and no statistical test. Several of the differences are small (e.g., penumbra 0.835 to 0.838 in BL1 vs BL2 and core 0.792 to 0.802), and without variance estimates or a paired test, these differences cannot be claimed as significant. This is a load-bearing issue because the central claim rests on the magnitude and reliability of these gains. The authors should provide per-fold results, error bars, and an appropriate significance test (e.g., paired t-test or Wilcoxon signed-rank test across folds or patients).
  3. [Section 3 and Table 1] The evaluation is limited to the SPES subset of ISLES 2015 with internal three-fold cross-validation only. The paper does not report results on the full ISLES test set or on any external dataset, and it does not compare against previously published ISLES results. Because the reported numbers are averages over a small, single-site dataset, the generalizability of the claimed improvement to the broader population of ischaemic stroke scans remains unsubstantiated. A comparison with published state-of-the-art methods or at least an external validation split would strengthen the conclusion.
minor comments (4)
  1. [Abstract and Conclusion] There are several typographical and grammatical errors: 'the the accurate' in the abstract, and 'it's effectiveness' in the conclusion should be 'its effectiveness'. These should be corrected.
  2. [Section 2] The boundary extraction parameters (3-pixel thickness, 3x3 structuring element, weight 10) are presented without justification or ablation. Since the paper's central claim concerns the boundary-weighted loss, an ablation over boundary thickness and weight would clarify the sensitivity of the results to these hyperparameters.
  3. [Section 2 and Table 1] Training details are missing: the paper does not specify the optimizer, learning rate, number of epochs, batch size, or the weighting coefficients used to combine the CE, adversarial, LS, and BD losses. Without these, the experiments are not fully reproducible.
  4. [Figure 2] The qualitative figure shows only selected baseline results (BL3, BL4, BL7, Proposed). It would be helpful to show the full set of baselines, or at least BL1 and BL5, so that the claimed improvement from residual connections is visually evident.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: central claims rest on in-paper ablations; minor self-citations are not load-bearing.

full rationale

This is an empirical segmentation paper rather than a formal derivation. The claimed improvements are evaluated by three-fold cross-validation on the external SPES-ISLES 2015 benchmark, with baselines defined in the paper. The residual-connection claim is isolated (BL5 vs BL1; BL7 vs BL3), and the adversarial-training component is taken from the authors' own prior work but is not the load-bearing claim. The self-citations to SUMNet (Nandamuri et al., 2019) and to Sathish et al. (2019) supply the base architecture and the adversarial scheme; neither is used to prove the residual or boundary-loss benefit. No fitted parameter is renamed as a prediction, no quantity is defined in terms of the reported Dice, and no uniqueness theorem is imported. A separate experimental-design concern should be noted: the boundary-weighted (BD) loss is never added alone. Every BD condition also includes Lovász-Softmax (LS) loss, so the improvement attributed to 'the weighted boundary loss' in Section 3 is confounded with LS. This is a validity/attribution weakness, not circularity, because the conclusion does not reduce to an input by construction. The only circularity-adjacent feature is the presence of self-citations, but they are not load-bearing, hence the low score.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central empirical comparison depends on three hand-picked segmentation hyperparameters (boundary weight 10, boundary thickness 3, structuring element 3x3) and an unspecified loss-weighting scheme, plus the authors' own adversarial training recipe. No new entities or physics are introduced.

free parameters (4)
  • boundary pixel weight = 10
    Chosen by hand in Section 2 to upweight boundary pixels; no ablation or sensitivity analysis is reported.
  • boundary thickness = 3 pixels
    Boundary is defined by subtracting eroded from dilated ground truth using a 3x3 structuring element (Section 2); the thickness is not varied.
  • structuring element size = 3 x 3
    Used for morphological boundary extraction in Section 2; no alternative sizes tested.
  • loss weighting coefficients = not reported
    The combination of CE, adversarial, Lovasz-Softmax and boundary losses is not specified quantitatively, so the relative balance is a hidden free choice.
assumptions (3)
  • domain assumption Adversarial training with three relativistic discriminators (Sathish et al., 2019) improves segmentation.
    The proposed method adopts this strategy from the authors' own prior work and does not independently validate it.
  • domain assumption The SPES-ISLES 2015 ground-truth annotations for core and penumbra are correct and clinically meaningful.
    The evaluation treats the challenge annotations as truth without verification.
  • ad hoc to paper Morphological boundary extraction with a 3x3 structuring element approximates clinically relevant lesion boundaries.
    The boundary loss is defined on this hand-crafted boundary, which is not clinically validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Significance of Residual Learning and Boundary Weighted Loss in Ischaemic Stroke Lesion Segmentation." pith.science (2026). https://pith.science/paper/NQ4CIRKC

@misc{pith2026190804840,
  author       = {Pith},
  title        = {Pith review of: Significance of Residual Learning and Boundary Weighted Loss in Ischaemic Stroke Lesion Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NQ4CIRKC}},
  note         = {Machine review of arXiv:1908.04840}
}
abstract

Radiologists use various imaging modalities to aid in different tasks like diagnosis of disease, lesion visualization, surgical planning and prognostic evaluation. Most of these tasks rely on the the accurate delineation of the anatomical morphology of the organ, lesion or tumor. Deep learning frameworks can be designed to facilitate automated delineation of the region of interest in such cases with high accuracy. Performance of such automated frameworks for medical image segmentation can be improved with efficient integration of information from multiple modalities aided by suitable learning strategies. In this direction, we show the effectiveness of residual network trained adversarially in addition to a boundary weighted loss. The proposed methodology is experimentally verified on the SPES-ISLES 2015 dataset for ischaemic stroke segmentation with an average Dice coefficient of $0.881$ for penumbra and $0.877$ for core. It was observed that addition of residual connections and boundary weighted loss improved the performance significantly.

Figures

Figures reproduced from arXiv: 1908.04840 by the authors.

Figure 1
Figure 1. Architecture of the proposed network Performance of the proposed method is compared with the following baselines: BL1: SUMNet trained using only CE loss BL2: SUMNet with CE, LS and BD losses BL3: SUMNet trained adversarially along with CE loss BL4: SUMNet trained adversarially along with CE, LS and BD losses 1. http://www.isles-challenge.org/ISLES2015/ 2 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure shows sample inputs, ground truth annotation and the segmented results (e)-(h) for different baselines (BL) and the proposed method. References The ATLANTIS et al. Association of outcome with early stroke treatment: pooled analysis of atlantis, ecass, and ninds rt-pa stroke trials. The Lancet, 363(9411):768–774, 2004. Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lov´asz-softmax loss: A tractab… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

7 extracted references · 6 canonical work pages

  1. [1]

    Association of outcome with early stroke treatment: pooled analysis of atlantis, ecass, and ninds rt-pa stroke trials

    The ATLANTIS et al. Association of outcome with early stroke treatment: pooled analysis of atlantis, ecass, and ninds rt-pa stroke trials. The Lancet, 363 0 (9411): 0 768--774, 2004

    work page 2004

  2. [2]

    The lov \'a sz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks

    Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lov \'a sz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In Proc. IEEE Conf. Comp. Vis. Patt. Recog., pages 4413--4421, 2018

    work page 2018

  3. [3]

    Pathobiology of ischaemic stroke: an integrated view

    Ulrich Dirnagl, Costantino Iadecola, and Michael A Moskowitz. Pathobiology of ischaemic stroke: an integrated view. Trends in neurosciences, 22 0 (9): 0 391--397, 1999

    work page 1999

  4. [4]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. Comp. Vis. Patt. Recog., pages 770--778, 2016

    work page 2016

  5. [5]

    SUMNet: Fully Convolutional Model for Fast Segmentation of Anatomical Structures in Ultrasound Volumes

    Sumanth Nandamuri, Debarghya China, Pabitra Mitra, and Debdoot Sheet. Sumnet: Fully convolutional model for fast segmentation of anatomical structures in ultrasound volumes. arXiv preprint arXiv:1901.06920, 2019

    work page Pith review arXiv 1901

  6. [6]

    Adversarially Trained Convolutional Neural Networks for Semantic Segmentation of Ischaemic Stroke Lesion using Multisequence Magnetic Resonance Imaging

    Rachana Sathish, Ronnie Rajan, Anusha Vupputuri, Nirmalya Ghosh, and Debdoot Sheet. Adversarially trained convolutional neural networks for semantic segmentation of ischaemic stroke lesion using multisequence magnetic resonance imaging. arXiv preprint arXiv:1908.01176, 2019

    work page Pith review arXiv 1908

  7. [7]

    Very deep convolutional networks for large-scale image recognition

    Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014

    arXiv 2014

Pith tools

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