Pith. sign in

REVIEW 4 major objections 5 minor 49 references

Semi-supervised learning and integration of multi-sequence MR-images for carotid vessel wall and plaque segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that the fusion point in a U-Net—bottleneck rather than input—drives multi-sequence carotid MRI segmentation quality, with the best binary Dice of 0.8725 and IoU of 0.7930.

desk verdict Competent empirical study whose headline fusion-point claim is confounded by model capacity; worth reviewing with a parameter-matched ablation required. read the letter →

arxiv 2507.07496 v1 pith:Y52WYHZL submitted 2025-07-10 eess.IV cs.CVcs.LG

Marie-Christine Pali , Christina Schwaiger , Malik Galijasevic , Valentin K. Ladenhauf , Stephanie Mangesius , Elke R. Gizewski This is my paper
classification eess.IVcs.CVcs.LG
keywords carotidarterysegmentationmulti-sequenceMRIsemi-supervisedlearningconsistencyregularizationfusionstrategyU-Netvesselwallandplaquegeometricprior
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

Carotid plaque is a leading stroke risk factor, and multi-sequence MRI can characterize it, but automatic segmentation of vessel wall and plaque is hard because plaque shapes vary and expert labels are scarce. This paper proposes a two-stage pipeline: a localization network that finds the carotid arteries under a geometric prior, followed by a fine segmentation network, both built as multi-sequence U-Nets. Its central claim is that where the five MRI sequences are mixed inside the U-Net matters—fusing feature maps at the bottleneck outperforms concatenating them at the input—and the proposed architecture reaches a binary Dice of 0.8725 and IoU of 0.7930, versus 0.8265 and 0.7350 for a basic U-Net with input fusion. The accompanying semi-supervised, perturbation-consistent training with uncertainty filtering raises binary Dice to 0.8773 and improves behavior on difficult slices, suggesting that limited labeled data need not block usable carotid segmentations.

What carries the argument

The central mechanism is the fusion point of a multi-sequence U-Net: each of the five sequences (PDw, T1w, T1ce, T2w, TOF) is fed to its own encoder branch, and the per-sequence feature maps are concatenated at the bottleneck and at every decoder level, so high-level modality-specific features are preserved before merging. Around this sit two supporting mechanisms: a one-way consistency semi-supervised scheme, where a clean teacher (exponential moving average of student weights) supervises a perturbed student under geometric and photometric transformations with an uncertainty-weighted MSE loss; and a coarse localization model whose loss includes a prior that penalizes wrong numbers of connected components and excessive asymmetry of the two carotid artery centers.

What would settle it

Deliberately misalign the five sequences by known small translations and rotations and repeat the input-vs-bottleneck fusion comparison: if bottleneck fusion no longer beats input fusion as misalignment grows, the claim that fusion-point selection is the operative factor is refuted; alternatively, a multi-center replication with consensus ground truth that fails to reproduce the reported Dice gap (0.8725 vs 0.8265) would also settle it.

Watch

Extended reading notes

Core claim

The paper establishes that fusion-point selection is a first-order design choice in U-Net-based multi-sequence segmentation. Processing each MRI sequence through its own encoder and merging the streams at the bottleneck—with the merged maps also forwarded through the decoder's skip connections—yields binary Dice 0.8725 and IoU 0.7930 for carotid vessel wall and plaque, beating the same architecture with input fusion (0.8549/0.7706) and a basic U-Net with input fusion (0.8265/0.7350). The same ordering holds for a basic U-Net, where bottleneck fusion adds more than four Dice points over input fusion. The paper also develops a semi-supervised training scheme that enforces transformation consistency between a perturbed student and an EMA teacher, filters the consistency signal by Monte-Carlo-dropout uncertainty, and reports expert-rated predicted masks close in quality to ground truth (mean 4.56 vs 4.75 on a 1–5 scale for binary segmentation).

Load-bearing premise

The whole comparison rests on the assumption that the five MRI sequences are accurately registered to the labeled PDw sequence, so that the same anatomical locations line up across modalities; if the registration is imperfect, the measured fusion-point differences could be artifacts of misalignment rather than real architectural effects.

Editorial extensions

If this is right

  • A basic U-Net gains more than four Dice points by switching from input fusion to bottleneck fusion, so fusion-point choice should be an explicit design variable in any multi-sequence segmentation pipeline.
  • Because the proposed U-Net shrinks the input-vs-bottleneck gap relative to the basic U-Net, architectural upgrades (residual blocks, SE modules, deep supervision) can partly compensate for a poor fusion choice.
  • Uncertainty-gated one-way consistency training improves binary Dice from 0.8645 to 0.8773 and is reported to produce visually more stable predictions on slices with artifacts and unusual plaque morphology.
  • The geometric prior lifts carotid detection from 95.18% of slices to 99.87%, showing that simple anatomical constraints can replace additional data.
  • Because vessel wall and plaque are segmented jointly, downstream plaque-only analysis is possible without a separate network to isolate plaque from wall.

Reading between the lines

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

  • The bottleneck-over-input ordering should transfer to other multi-sequence U-Net applications, such as brain tumor or cardiac MRI, since the mechanism—preserving per-modality high-level features before merging—is not specific to carotid anatomy; this is a testable prediction the paper does not make.
  • The small semi-supervised gain may reflect the near-balanced labeled/unlabeled ratio; adding many more unlabeled patients while holding the architecture fixed should make consistency regularization contribute more.
  • Because the authors flag registration mismatches as a remaining limitation, an experiment that measures fusion gain under controlled misregistration could separate a true architectural benefit from tolerance to alignment error.
  • Re-annotating the test set with consensus ground truth would test the paper's explanation for the multiclass gap—that annotation errors, not model errors, dominate the vessel-wall/plaque confusion.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a two-stage semi-supervised segmentation pipeline for carotid artery vessel wall and plaque in five-sequence MRI data. A coarse localization network, guided by anatomical priors on carotid number and symmetry, identifies a region of interest; a fine segmentation network, built on a multi-level multi-sequence U-Net, then delineates vessel wall and plaque. The authors compare input-level fusion with bottleneck fusion in both a basic U-Net and their proposed architecture, evaluate binary and multiclass segmentation on a 52-patient dataset, and extend a perturbed-student/clean-teacher consistency framework with an uncertainty-weighted loss. Quantitative results are complemented by an expert rating study.

Significance. If the fusion-point claim holds, it is practically valuable: the paper suggests that a simple architectural choice, fusing multi-sequence feature maps at the bottleneck rather than at the input, substantially improves carotid segmentation, and that this effect can be larger than architectural upgrades. The prior-based localization result is also notable, reducing failed detections from 38/789 to 1/789 slices. The semi-supervised framework for multi-sequence carotid MRI, including uncertainty-aware consistency, is a sensible extension of existing self-ensembling methods. However, the quantitative evidence is weakened by a capacity confound in the central fusion comparison and by the absence of variance or significance reporting; the results are promising but not yet conclusive.

major comments (4)
  1. [Section 3.1, Section 5, Table 2] The central claim that 'bottleneck fusion outperforms early fusion in U-Net-based architectures' rests on Table 2, but that comparison is not capacity-matched. In the bottleneck-fusion condition, each of the five MRI sequences is processed by its own independent encoder path, whereas input fusion concatenates the five channels into a single shared encoder. The bottleneck-fusion models therefore have roughly five times the encoder parameters and compute of the input-fusion models. The observed gains, such as 0.8265 to 0.8702 Dice for basic U-Net, could be explained by increased capacity rather than by fusion point alone. Please provide parameter counts, FLOPs, and an ablation that matches capacity (for example, a wider shared encoder, or proportionally fewer channels in each of the five encoder paths) before attributing the improvement to fusion location. This is load-bearing because the Discussion in Section 5 explicitly frames the key finding as the choice of fusion point.
  2. [Section 4.2, Tables 2-5] All quantitative segmentation results are reported only as mean metrics across folds, with no standard deviations, confidence intervals, or significance tests. Many comparisons that drive the narrative are small: 0.8725 versus 0.8702 in Table 2, 0.8773 versus 0.8725 in Table 4, and 0.6884 versus 0.6787 in Table 5. With five folds and a limited patient cohort, such differences are plausibly within fold-to-fold or run-to-run variation. Please report per-fold results, standard deviations, and a paired significance test over patient-level predictions, so that the claimed benefits of the proposed architecture and of semi-supervised learning can be assessed.
  3. [Section 4.2, Table 3 and Figure 9] The multiclass results reverse the binary-case ranking of the two architectures: basic U-Net with bottleneck fusion achieves Dice 0.7010, while the proposed U-Net with bottleneck fusion achieves Dice 0.6787. The paper explains this by arguing that the proposed model's 'wrong' plaque regions reflect ground-truth labeling errors, but this explanation is post hoc and is not backed by a quantitative analysis, such as a re-evaluation on adjudicated or corrected labels, a label-noise model, or an inter-expert consistency measure for the specific test cases. Since one of the contributions is the proposed architecture, this contradiction needs direct supporting evidence rather than visual inspection alone.
  4. [Section 4.1, Section 5] The multi-sequence fusion comparison assumes that all five sequences are accurately aligned to the annotated PDw space. The paper states only that rigid and scaling transformations were used for registration and that some slices with distortions were manually removed; no registration error is reported. If residual misalignment is non-negligible, it could affect the fusion-point comparison, because input fusion and bottleneck fusion combine misregistered features differently. The manuscript itself acknowledges this by listing 'addressing registration mismatches' as future work. Please quantify registration accuracy (for example, target registration error on anatomical landmarks, or the number of slices rejected for registration failure) or otherwise bound the effect of residual misalignment on the central comparison.
minor comments (5)
  1. [Title page and Section 5] There are several typos and spacing issues, including 'V ASCage' on the title page and 'medial image analysis' in the Discussion; these should be corrected.
  2. [Table 2] The precision and recall entries in the second row appear as '0.83390.9200' without a separating space; please fix the formatting.
  3. [Section 4.1] The T2w sequence description lists 'fat saturated' twice; this duplication should be removed.
  4. [Section 3.2] The notation for the total loss, L_total_t(Phi_theta(X), Y), is slightly inconsistent because the supervised and consistency terms are written with different arguments in the following line; please unify the notation for clarity.
  5. [Section 5] The Discussion states that 'the wrong plaque' regions suggest label inaccuracy, but the term is not defined; clarifying the criterion would help readers interpret Figure 9.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: all reported gains come from held-out empirical comparisons, and no load-bearing claim reduces to a fitted parameter or self-citation.

full rationale

This paper is an empirical segmentation study. The central claims—that bottleneck fusion improves Dice/IoU over input fusion (Table 2) and that semi-supervised consistency and uncertainty weighting add small gains (Tables 4 and 5)—are evaluated on patient-wise held-out splits with fixed hyperparameters, and the reported Dice/IoU metrics are not used to define or fit any loss term. The consistency loss is a training regularizer, not the evaluation target, so its contribution is an ablation rather than a self-fulfilling prediction. The prior-knowledge localization loss is an independent anatomical constraint (connected components and approximate symmetry), and its effect is measured by a with/without ablation in Table 1 rather than asserted by construction. No equation in the paper defines a prediction in terms of the quantity it is claimed to predict, and no load-bearing argument relies on a self-citation: the cited comparison architecture [35] is an external prior work, not an author self-citation. The capacity-confound concern about Table 2—that bottleneck fusion uses separate encoder paths and therefore more parameters—is a threat to the internal-validity interpretation of the fusion-point comparison, but it is an experimental confound, not a circularity, because the comparison remains an empirical measurement rather than a definitional equivalence. The acknowledged limitations (small dataset, registration mismatches, imperfect ground truth) are validity concerns for generality, not circular steps. No circular step was found.

Assumptions & free parameters 12 free parameters · 4 assumptions · 0 invented entities

The model depends on many hand-set hyperparameters and strong anatomical and registration assumptions. No new entities are introduced.

free parameters (12)
  • lambda_loc = 0.5
    Weight combining modified Tversky and BCE losses in coarse localization (Section 4.1).
  • delta_loc = 0.7
    Tversky index parameter balancing false positives and false negatives in localization.
  • lambda_seg = 0.5
    Weight combining modified asymmetric Focal and Focal Tversky losses in fine segmentation.
  • delta_seg = 0.6
    Tversky index parameter in fine segmentation, suggested by [43].
  • theta = 0.25
    Asymmetry parameter suppressing background and enhancing rare classes in fine segmentation.
  • omega = 0.1
    Weight of the prior loss in coarse localization (Eq. 4).
  • alpha = 0.999
    EMA momentum for teacher model weights.
  • k_seg = 20 (k_loc = 10)
    Maximum weight of consistency loss; k_loc for localization.
  • R = 60 for localization, 40 for fine segmentation
    Ramp-up length for the consistency weight lambda(t).
  • T_uncertainty = 8
    Number of stochastic forward passes for Monte Carlo dropout uncertainty estimation.
  • symmetry_threshold = 20 pixels
    Max allowed y-coordinate difference between centers of left and right carotid arteries in the prior.
  • num_connected_components = 1 to 2 per side
    Prior constraint on the number of detected carotid artery components per side.
assumptions (4)
  • domain assumption Each side of the image contains one or two carotid artery vessels, and the two carotid artery centers lie approximately on the same horizontal line (max y-difference <= 20 pixels).
    Used in the prior loss L_prior (Section 3.3, Eq. 4) to constrain localization; this anatomical prior is hard-coded and may fail in atypical anatomies or severe pathology.
  • domain assumption Rigid and scaling registration of the five MRI sequences to the PDw sequence is sufficiently accurate that labels transfer between sequences.
    Section 4.1 Preprocessing: all sequences are resampled and registered to PDw; all multi-sequence fusion assumes pixel-wise alignment. The authors later discuss 'registration mismatches' and propose random translations and rotations to mitigate them, indicating this assumption is imperfect.
  • domain assumption Manual annotations by two radiologists are reliable enough for training and evaluation.
    The entire supervised loss and evaluation depend on these labels; the authors themselves document errors in ground-truth masks in Section 5, partially explaining multiclass performance, which makes the reliability of the annotations load-bearing.
  • ad hoc to paper The uncertainty threshold tau(t) = ln(2) * (3/4 + lambda(t)/4) is a reasonable ad hoc schedule.
    Section 3.2 defines this threshold without derivation; it controls which teacher predictions are used in the consistency loss and is tuned to the training schedule.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Semi-supervised learning and integration of multi-sequence MR-images for carotid vessel wall and plaque segmentation." pith.science (2026). https://pith.science/paper/Y52WYHZL

@misc{pith2026250707496,
  author       = {Pith},
  title        = {Pith review of: Semi-supervised learning and integration of multi-sequence MR-images for carotid vessel wall and plaque segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y52WYHZL}},
  note         = {Machine review of arXiv:2507.07496}
}
read the original abstract

The analysis of carotid arteries, particularly plaques, in multi-sequence Magnetic Resonance Imaging (MRI) data is crucial for assessing the risk of atherosclerosis and ischemic stroke. In order to evaluate metrics and radiomic features, quantifying the state of atherosclerosis, accurate segmentation is important. However, the complex morphology of plaques and the scarcity of labeled data poses significant challenges. In this work, we address these problems and propose a semi-supervised deep learning-based approach designed to effectively integrate multi-sequence MRI data for the segmentation of carotid artery vessel wall and plaque. The proposed algorithm consists of two networks: a coarse localization model identifies the region of interest guided by some prior knowledge on the position and number of carotid arteries, followed by a fine segmentation model for precise delineation of vessel walls and plaques. To effectively integrate complementary information across different MRI sequences, we investigate different fusion strategies and introduce a multi-level multi-sequence version of U-Net architecture. To address the challenges of limited labeled data and the complexity of carotid artery MRI, we propose a semi-supervised approach that enforces consistency under various input transformations. Our approach is evaluated on 52 patients with arteriosclerosis, each with five MRI sequences. Comprehensive experiments demonstrate the effectiveness of our approach and emphasize the role of fusion point selection in U-Net-based architectures. To validate the accuracy of our results, we also include an expert-based assessment of model performance. Our findings highlight the potential of fusion strategies and semi-supervised learning for improving carotid artery segmentation in data-limited MRI applications.

Figures

Figures reproduced from arXiv: 2507.07496 by the authors.

Figure 1
Figure 1. Slices {x j ik} 5 j=1 for a fixed patient i and a slice index k for the 5 MRI sequences. Segmenting vessel wall and plaques within the carotid artery in multi-sequence MRI data poses significant challenges due to the morphological variability of plaques, imaging artifacts, and the limited availability of labeled data. Plaques exhibit diverse spatial and structural characteristics, and with only a small amount of lab… view at source ↗
Figure 2
Figure 2. Segmentation workflow. leverage complementary information provided by different modalities, and introduce a novel multi-level multi-sequence version of U-Net architecture. Considering the vast amount of information contained in multi-sequence MRI data and the class imbalance present in each image slice, we formulate our problem as hierarchical segmentation problem. This means, our segmentation approach consists of t… view at source ↗
Figure 3
Figure 3. Model architecture of fine segmentation model. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: One-way consistency semi-supervised learning approach with perturbed student and clean teacher. where t denotes the current training step and R the ramp-up length, i.e. the training step where λ(t) = k. This function allows a warm start, ensuring that the contribution …
Figure 5
Figure 5. Figure 5: Left and right carotid arteries of two patients with highlighted ground-truth ves￾sel walls and plaques. The center of each artery is highlighted by a horizontal line. Both images 5a and 5b show that the carotid arteries can be highly non￾symmetric, however their cente…
Figure 6
Figure 6. Figure 6: Performance of our U-Net and basic U-Net: Train and validation loss versus number of epochs. and bottleneck fusion became less severe than in the U-Net case. U-Net with bottleneck fusion provided the best recall scores, indicating that our U-Net model provided fewer fa…
Figure 7
Figure 7. Figure 7: Results of binary carotid artery vessel wall and plaque segmentation using our U-Net and late fusion. From left to right: Ground truth mask, predicted mask, their overlay, PDw image slice. For multiclass segmentation, [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: Results of multiclass carotid artery vessel wall and plaque segmentation using our U-Net and late fusion. From left to right: Ground truth mask, predicted mask, their overlay, PDw image slice. the segmentation scores of our U-Net model and a basic U-Net model using bot…
Figure 9
Figure 9. Figure 9: Comparison of multiclass ground truth masks, predicted masks using basic U-Net with late fusion, and predicted masks using our U-Net with late fusion. 5 Discussion In this work, we presented a semi-supervised deep learning-based segmentation approach designed to effect…
Figure 10
Figure 10. Figure 10: Demonstration of accuracy and integration of underlying image structures of our prediction. From left to right: Ground truth mask, predicted mask, their overlay, PDw image slice. Dice IoU Precision Recall Our U-Net, no semi-supervision 0.8645 0.7829 0.8608 0.8688 Our …
Figure 11
Figure 11. Figure 11: Histogram of average scores for predicted ratings minus ground truth ratings [PITH_FULL_IMAGE:figures/full_fig_p026_11.png]
Figure 12
Figure 12. Figure 12: Example predictions from our U-Net model using bottleneck fusion and ground truth masks from expert radiologists for multiclass and binary segmentation. From left to right: Ground truth mask, predicted mask, their overlay, PDw image slice. 29 [PITH_FULL_IMAGE:figures…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 35 canonical work pages

  1. [1]

    Enhancing medical image segmentation: Ground truth optimization through evaluating uncertainty in expert annotations.Mathematics, 11(17):3771, 2023

    Georgios Athanasiou, Josep Lluis Arcos, and Jesus Cerquides. Enhancing medical image segmentation: Ground truth optimization through evaluating uncertainty in expert annotations.Mathematics, 11(17):3771, 2023

    work page 2023

  2. [2]

    There are many consistent explanations of unlabeled data: Why you should average, 2019

    Ben Athiwaratkun, Marc Finzi, Pavel Izmailov, and Andrew Gordon Wilson. There are many consistent explanations of unlabeled data: Why you should average, 2019. URLhttps://arxiv.org/abs/1806.05594

    arXiv 2019

  3. [3]

    Multi Modal Convolutional Neural Networks for Brain Tumor Segmentation

    Mehmet Ayg¨ un, Yusuf Huseyin Sahin, and G¨ ozde B.¨Unal. Multi modal convolutional neural networks for brain tumor segmentation.CoRR, abs/1809.06191, 2018. URL http://arxiv.org/abs/1809.06191

    work page Pith review arXiv 2018

  4. [4]

    Matthews, and Daniel Rueck- ert

    Wenjia Bai, Ozan Oktay, Matthew Sinclair, Hideaki Suzuki, Martin Rajchl, Gia- como Tarroni, Ben Glocker, Andrew King, Paul M. Matthews, and Daniel Rueck- ert. Semi-supervised learning for network-based cardiac mr image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI 2017: 20th International Conference, Quebec City, QC,...

    2017

  5. [5]

    Albumentations: fast and flexible image augmen- tations.Information, 11(2):125, 2020

    Alexander Buslaev, Vladimir I Iglovikov, Eugene Khvedchenya, Alex Parinov, Mikhail Druzhinin, and Alexandr A Kalinin. Albumentations: fast and flexible image augmen- tations.Information, 11(2):125, 2020

    work page 2020

  6. [6]

    Hippe, Xihai Zhao, Rui Li, Thomas S

    Li Chen, Jie Sun, Gador Canton, Niranjan Balu, Daniel S. Hippe, Xihai Zhao, Rui Li, Thomas S. Hatsukami, Jenq-Neng Hwang, and Chun Yuan. Automated artery localiza- tion and vessel wall segmentation using tracklet refinement and polar conversion.IEEE Access, 8:217603–217614, 2020. ISSN 2169-3536. doi: 10.1109/ACCESS.2020.3040616

    work page arXiv 2020

  7. [7]

    Recent advances and clinical applications of deep learning in medical image analysis.Medical image analysis, 79:102444, 2022

    Xuxin Chen, Ximin Wang, Ke Zhang, Kar-Ming Fung, Theresa C Thai, Kathleen Moore, Robert S Mannel, Hong Liu, Bin Zheng, and Yuchen Qiu. Recent advances and clinical applications of deep learning in medical image analysis.Medical image analysis, 79:102444, 2022

    work page 2022

  8. [8]

    Stroke risk study based on deep learning-based magnetic resonance imaging carotid plaque automatic segmentation algorithm.Frontiers in Cardiovascular Medicine, 10, 2023

    Ya-Fang Chen, Zhen-Jie Chen, You-Yu Lin, Zhi-Qiang Lin, Chun-Nuan Chen, Mei-Li Yang, Jin-Yin Zhang, Yuan-zhe Li, Yi Wang, and Yin-Hui Huang. Stroke risk study based on deep learning-based magnetic resonance imaging carotid plaque automatic segmentation algorithm.Frontiers in Cardiovascular Medicine, 10, 2023

    work page 2023

Show all 49 references
  1. [9]

    Semi-supervised brain lesion segmentation with an adapted mean teacher model.ArXiv, abs/1903.01248, 2019

    Wenhui Cui, Yanling Liu, Yuxing Li, Meng-Hao Guo, Yiming Li, Xiuli Li, Tianle Wang, Xiangzhu Zeng, and Chuyang Ye. Semi-supervised brain lesion segmentation with an adapted mean teacher model.ArXiv, abs/1903.01248, 2019. URLhttps: //api.semanticscholar.org/CorpusID:67855384

    1903 arXiv

  2. [10]

    Deep learning technology in vascular image segmentation and disease diagnosis.Journal of Intelligent Medicine, 2024

    Chengyang Du, Jie Zhuang, and Xinglu Huang. Deep learning technology in vascular image segmentation and disease diagnosis.Journal of Intelligent Medicine, 2024. 30

    2024

  3. [11]

    Semi-supervised learning for pelvic mr image segmentation based on multi-task residual fully convolutional net- works

    Zishun Feng, Dong Nie, Li Wang, and Dinggang Shen. Semi-supervised learning for pelvic mr image segmentation based on multi-task residual fully convolutional net- works. In2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018), pages 885–888, 2018. doi: 10.11...

    2018

  4. [13]

    Semi-supervised learning by entropy minimiza- tion

    Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimiza- tion. InProceedings of the 18th International Conference on Neural Information Pro- cessing Systems, NIPS’04, page 529–536, Cambridge, MA, USA, 2004. MIT Press

    2004

  5. [14]

    Revisiting consistency for semi- supervised semantic segmentation.Sensors, 23(2):940, 2023

    Ivan Grubiˇ si´ c, Marin Orˇ si´ c, and Siniˇ saˇSegvi´ c. Revisiting consistency for semi- supervised semantic segmentation.Sensors, 23(2):940, 2023

    2023

  6. [15]

    Sheng, Yuqing Song, Yi Liu, Chengjian Qiu, Siqi Ma, and Zhe Liu

    Kai Han, Victor S. Sheng, Yuqing Song, Yi Liu, Chengjian Qiu, Siqi Ma, and Zhe Liu. Deep semi-supervised learning for medical image segmentation: A review.Ex- pert Systems with Applications, 245:123052, 2024. ISSN 0957-4174. doi: https://doi. org/10.1016/j.eswa.2023.123052. UR...

    2024

  7. [16]

    Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam

    Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convo- lutional neural networks for mobile vision applications.CoRR, abs/1704.04861, 2017. URLhttp://arxiv.org/abs/1704.04861

    2017 arXiv

  8. [17]

    Squeeze-and-excitation networks.CoRR, abs/1709.01507, 2017

    Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks.CoRR, abs/1709.01507, 2017. URLhttp://arxiv.org/abs/1709.01507

    2017 arXiv

  9. [18]

    J¨ ager, Simon A

    Fabian Isensee, Paul F. J¨ ager, Simon A. A. Kohl, Jens Petersen, and Klaus Maier- Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation.Nature Methods, 18:203 – 211, 2021. doi: https://doi.org/10.1038/ s41592-020-01008-z

    2021

  10. [19]

    Averaging weights leads to wider optima and better generalization.arXiv preprint arXiv:1803.05407, 2018

    Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gor- don Wilson. Averaging weights leads to wider optima and better generalization.arXiv preprint arXiv:1803.05407, 2018

    2018 arXiv

  11. [20]

    Deep learning applications in medical image analysis.Ieee Access, 6:9375–9389, 2017

    Justin Ker, Lipo Wang, Jai Rao, and Tchoyoson Lim. Deep learning applications in medical image analysis.Ieee Access, 6:9375–9389, 2017

    2017

  12. [21]

    Comparative review on traditional and deep learning methods for medical image segmentation

    Shadi Mahmoodi Khaniabadi, Haidi Ibrahim, Ilyas Ahmad Huqqani, Farzad Mahmoodi Khaniabadi, Harsa Amylia Mat Sakim, and Soo Siang Teoh. Comparative review on traditional and deep learning methods for medical image segmentation. In2023 IEEE 14th control and system graduate resea...

    2023

  13. [22]

    Londhe, S

    Anita Khanna, Narendra D. Londhe, S. Gupta, and Ashish Semwal. A deep resid- ual u-net convolutional neural network for automated lung segmentation in com- puted tomography images.Biocybernetics and Biomedical Engineering, 40(3):1314– 1327, 2020. ISSN 0208-5216. doi: https://d...

    2020 doi

  14. [23]

    Ran Li, Jie Zheng, Mohamed A Zayed, Jeffrey E Saffitz, Pamela K Woodard, and Abhinav K Jha. Carotid atherosclerotic plaque segmentation in multi-weighted mri using a two-stage neural network: advantages of training with high-resolution imaging and histology.Frontiers in Cardio...

    2023

  15. [24]

    Transformation-consistent self-ensembling model for semisupervised medical image seg- mentation.IEEE transactions on neural networks and learning systems, 32(2):523–534, 2020

    Xiaomeng Li, Lequan Yu, Hao Chen, Chi-Wing Fu, Lei Xing, and Pheng-Ann Heng. Transformation-consistent self-ensembling model for semisupervised medical image seg- mentation.IEEE transactions on neural networks and learning systems, 32(2):523–534, 2020

    2020

  16. [25]

    A comprehensive review of deep neural networks for medical image processing: Recent developments and future opportunities.Healthcare Analytics, page 100216, 2023

    Pawan Kumar Mall, Pradeep Kumar Singh, Swapnita Srivastav, Vipul Narayan, Marcin Paprzycki, Tatiana Jaworska, and Maria Ganzha. A comprehensive review of deep neural networks for medical image processing: Recent developments and future opportunities.Healthcare Analytics, page ...

    2023

  17. [26]

    Ozan Oktay, Jo Schlemper, Lo ¨ ıc Le Folgoc, Matthew C. H. Lee, Mattias P. Heinrich, Kazunari Misawa, Kensaku Mori, Steven G. McDonagh, Nils Y. Hammerla, Bernhard Kainz, Ben Glocker, and Daniel Rueckert. Attention u-net: Learning where to look for the pancreas.CoRR, abs/1804.0...

    2018 arXiv

  18. [27]

    Synthetic ground truth for validation of brain tumor mri segmentation

    Marcel Prastawa, Elizabeth Bullitt, and Guido Gerig. Synthetic ground truth for validation of brain tumor mri segmentation. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 26–33. Springer, 2005

    2005

  19. [28]

    Semi-supervised segmentation of retinoblastoma tumors in fundus images

    Amir Rahdar, Mohamad Javad Ahmadi, Masood Naseripour, Abtin Akhtari, Ahad Sedaghat, Vahid Zare Hosseinabadi, Parsa Yarmohamadi, Samin Hajihasani, and Reza Mirshahi. Semi-supervised segmentation of retinoblastoma tumors in fundus images. Research Square, 2023. doi: 10.21203/rs....

    2023 doi

  20. [29]

    Girshick, and Ali Farhadi

    Joseph Redmon, Santosh Kumar Divvala, Ross B. Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection.CoRR, abs/1506.02640, 2015. URL http://arxiv.org/abs/1506.02640

    2015 arXiv

  21. [30]

    U-net: Convolutional networks for biomedical image segmentation.CoRR, abs/1505.04597, 2015

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation.CoRR, abs/1505.04597, 2015. URLhttp://arxiv. org/abs/1505.04597

    2015 arXiv

  22. [31]

    Automated medical image segmentation techniques.Journal of Medical Physics / Association of Medical Physicists of India, 35:3 – 14, 2010

    Neeraj Sharma and Lalit Mohan Aggarwal. Automated medical image segmentation techniques.Journal of Medical Physics / Association of Medical Physicists of India, 35:3 – 14, 2010. URLhttps://api.semanticscholar.org/CorpusID:30824724. 32

    2010

  23. [32]

    Mean teachers are better role models: Weight- averaged consistency targets improve semi-supervised deep learning results.Advances in neural information processing systems, 30, 2017

    Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight- averaged consistency targets improve semi-supervised deep learning results.Advances in neural information processing systems, 30, 2017

    2017

  24. [33]

    Tsakanikas, Panagiotis K

    Vassilis D. Tsakanikas, Panagiotis K. Siogkas, Michalis D. Mantzaris, Vassiliki T. Potsika, Vassiliki I. Kigka, Themis P. Exarchos, Igor B. Koncar, Marija Jovanovic, Aleksandra Vujcic, Stefan Ducic, Jaroslav Pelisek, and Dimitrios I. Fotiadis. A deep learning oriented method f...

  25. [34]

    Understanding interobserver agreement: The kappa statistic.Family medicine, 37:360–3, 06 2005

    Anthony Viera and Joanne Garrett. Understanding interobserver agreement: The kappa statistic.Family medicine, 37:360–3, 06 2005

    2005

  26. [35]

    Jian Wang, Fan Yu, Mengze Zhang, Jie Lu, and Zhen Qian. A 3d framework for segmentation of carotid artery vessel wall and identification of plaque compositions in multi-sequence mr images.Computerized Medical Imaging and Graphics, 116:102402, September 2024. ISSN 08956111. doi...

    2024

  27. [36]

    Cottrell

    Panqu Wang, Pengfei Chen, Ye Yuan, Ding Liu, Zehua Huang, Xiaodi Hou, and Garrison W. Cottrell. Understanding convolution for semantic segmentation.CoRR, abs/1702.08502, 2017. URLhttp://arxiv.org/abs/1702.08502

    2017 arXiv

  28. [37]

    Application of artificial intelligence methods in carotid artery segmentation: a review.IEEE Access, 2023

    Yu Wang and Yudong Yao. Application of artificial intelligence methods in carotid artery segmentation: a review.IEEE Access, 2023

    2023

  29. [38]

    Simultaneous truth and perfor- mance level estimation (staple): an algorithm for the validation of image segmentation

    Simon K Warfield, Kelly H Zou, and William M Wells. Simultaneous truth and perfor- mance level estimation (staple): an algorithm for the validation of image segmentation. IEEE transactions on medical imaging, 23(7):903–921, 2004

    2004

  30. [39]

    Jiayi Wu, Jingmin Xin, Xiaofeng Yang, Jie Sun, Dongxiang Xu, Nanning Zheng, and Chun Yuan. Deep morphology aided diagnosis network for segmentation of carotid artery vessel wall and diagnosis of carotid atherosclerosis on black-blood vessel wall mri.Medical Physics, 46(12):554...

    2019

  31. [40]

    A comprehensive review of deep learning for medical image segmentation

    Qingling Xia, Hong Zheng, Haonan Zou, Dinghao Luo, Hongan Tang, Lingxiao Li, and Bin Jiang. A comprehensive review of deep learning for medical image segmentation. Neurocomputing, page 128740, 2024

    2024

  32. [41]

    A semantic segmentation method with emphasis on the edges for automatic vessel wall analysis.Applied Sciences, 12(14), 2022

    Wenjing Xu and Qing Zhu. A semantic segmentation method with emphasis on the edges for automatic vessel wall analysis.Applied Sciences, 12(14), 2022. ISSN 2076-

    2022

  33. [42]

    Deep learning- based automated detection of arterial vessel wall and plaque on magnetic resonance vessel wall images.Frontiers in Neuroscience, 16, June 2022

    Wenjing Xu, Xiong Yang, Yikang Li, Guihua Jiang, Sen Jia, Zhenhuan Gong, Yufei Mao, Shuheng Zhang, Yanqun Teng, Jiayu Zhu, Qiang He, Liwen Wan, Dong 33 Liang, Ye Li, Zhanli Hu, Hairong Zheng, Xin Liu, and Na Zhang. Deep learning- based automated detection of arterial vessel wa...

    2022

  34. [43]

    Michael Yeung, Evis Sala, Carola-Bibiane Sch¨ onlieb, and Leonardo Rundo. Uni- fied focal loss: Generalising dice and cross entropy-based losses to handle class imbalanced medical image segmentation.Computerized Medical Imaging and Graphics, 95:102026, 2022. ISSN 0895-6111. do...

    2022

  35. [44]

    Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmenta- tion

    Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmenta- tion. InMedical image computing and computer assisted intervention–MICCAI 2019: 22nd international conference, Shenzhen, C...

    2019

  36. [45]

    Road extraction by deep residual u-net.CoRR, abs/1711.10684, 2017

    Zhengxin Zhang, Qingjie Liu, and Yunhong Wang. Road extraction by deep residual u-net.CoRR, abs/1711.10684, 2017. URLhttp://arxiv.org/abs/1711.10684

    2017 arXiv

  37. [46]

    Chenglu Zhu, Xiaoyan Wang, Zhongzhao Teng, Shengyong Chen, Xiaojie Huang, Ming Xia, Lizhao Mao, and Cong Bai. Cascaded residual u-net for fully automatic segmen- tation of 3d carotid artery in high-resolution multi-contrast mr images.Physics in Medicine and Biology, 66(4):0450...

    2021 doi

  38. [47]

    Chenglu Zhu, Xiaoyan Wang, Shengyong Chen, Zhongzhao Teng, Cong Bai, Xiaojie Huang, Ming Xia, Zhanpeng Shao, Zheng Gu, and Peiliang Sun. Complex carotid artery segmentation in multi-contrast mr sequences by improved optimal surface graph cuts based on flow line learning.Medica...

    2022

  39. [2013]

    URLhttp://arxiv.org/abs/1311.2524

    arXiv

  40. [2020]

    doi: 10.1109/EMBC44109.2020.9176532

    ISBN 978-1-7281-1990-8. doi: 10.1109/EMBC44109.2020.9176532. URL https://ieeexplore.ieee.org/document/9176532/

    1990

  41. [3417]

    URLhttps://www.mdpi.com/2076-3417/12/14/ 7012

    doi: 10.3390/app12147012. URLhttps://www.mdpi.com/2076-3417/12/14/ 7012

    doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.