Pith. sign in

REVIEW 3 major objections 6 minor 62 references

A task-tuned U-Net cuts missed stroke lesions on DWI by up to 29% versus strong baselines.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 01:29 UTC pith:AAVTGPWB

load-bearing objection Solid single-center engineering paper with thorough stats and a real efficiency/recall trade-off; Abstract overstates overall superiority once nnU-Net is in the picture. the 3 major comments →

arxiv 2607.03568 v1 pith:AAVTGPWB submitted 2026-07-03 cs.CV cs.AI

EPRA U-Net: An Efficient Pyramid Residual Attention Framework for Accurate Infarct Segmentation in Diffusion-Weighted MRI

Hasan Ulutas , Muhammet Emin Sahin , Mustafa Fatih Erkoc , Esra Yuce , Turker Tuncer , Sengul Dogan , Serkan Kiranyaz This is my paper
classification cs.CV cs.AI
keywords infarct segmentationdiffusion-weighted MRIEPRA U-NetTversky lossEfficientNetASPPdual attentionstroke imaging
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Acute stroke care needs fast, accurate maps of ischemic infarcts on diffusion-weighted MRI, yet small lesions, fuzzy boundaries, and heavy class imbalance make ordinary segmenters miss too many. This paper builds EPRA U-Net, a U-Net variant that packs an EfficientNet encoder, residual-recurrent blocks, multi-scale ASPP context, and dual attention, then trains it with a Tversky loss that deliberately weights false negatives more heavily than false positives. On a new single-center set of 167 patients and 4,895 slices, the model reaches high Dice and lesion-level F1 while cutting missed lesions by 16–29% relative to UNet++, DeepLabV3+, and TransUNet. The practical claim is that the same architecture, optimized for clinical sensitivity rather than pure pixel overlap, can give more reliable infarct volume estimates for treatment decisions.

Core claim

EPRA U-Net, by combining an EfficientNet-B0 encoder, residual-recurrent (R2) blocks, ASPP multi-scale context, dual attention, and Tversky loss (α=0.4, β=0.6), attains superior infarct segmentation on DWI—pixel Dice 0.8984, per-sample Dice 0.9469, IoU 0.8155, recall 0.8887, lesion F1 0.9378, HD95 11.62 px—and reduces missed lesions by 16%, 25%, and 29% versus UNet++, DeepLabV3+, and TransUNet on a 167-patient held-out test set.

What carries the argument

EPRA U-Net: a hybrid encoder–decoder that uses EfficientNet-B0 for compact hierarchical features, R2 residual-recurrent blocks for spatial continuity, ASPP dilated multi-scale context, dual (position + channel) attention for lesion focus, and Tversky loss with β>α to penalize missed infarcts more than false alarms.

Load-bearing premise

That scores measured on one hospital’s retrospectively collected DWI scans of 167 patients will hold for other scanners, protocols, and patient populations.

What would settle it

An external multi-center DWI test set of comparable size on which EPRA U-Net no longer reduces false-negative lesion counts relative to the same three baselines under identical patient-level splitting and Tversky training.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The manuscript proposes EPRA U-Net, a hybrid encoder–decoder for acute ischemic infarct segmentation on DWI that combines an EfficientNet-B0 backbone, Residual-Recurrent (R2) blocks (t=2), ASPP multi-scale context, dual (position+channel) attention on skip pathways, and Tversky loss (α=0.4, β=0.6) to favor recall. On a single-center, patient-level split of 167 patients / 4,895 slices (14.7% positive), the model is compared with UNet++, DeepLabV3+, TransUNet under a matched training protocol and with 2D nnU-Net under its default self-configuring protocol. Headline results are pixel-aggregated Dice 0.8984, per-sample Dice 0.9469, IoU 0.8155, Recall 0.8887, Lesion F1 0.9378, HD95 11.62 px (median 2.00), with 16–29% fewer missed lesions versus the three matched baselines; ablation (Table 10) and multi-layered statistics (Wilcoxon, t-test, McNemar, bootstrap CIs, Cohen’s d) support the internal comparisons. The authors emphasize clinical prioritization of false-negative reduction and report a substantial training/inference cost advantage over nnU-Net.

Significance. If the reported gains hold under broader validation, the work is a solid, practically useful contribution to DWI infarct segmentation: it packages known components (EfficientNet, R2, ASPP, dual attention, Tversky) into a parameter-efficient pipeline (7.44M params, ~132 min training, ~34 ms/slice) that demonstrably reduces missed lesions relative to common matched-protocol baselines, with thorough lesion-level error analysis, ablations, and statistical testing. Strengths include patient-level splitting, clinically motivated loss design, multi-metric reporting (including FN lesion counts and HD95 median/mean), and explicit efficiency comparison to nnU-Net. The main scientific value is therefore a carefully engineered, efficiency-aware clinical tool rather than a fundamental architectural breakthrough; that is still of interest to medical imaging venues if claims are scoped accurately and external validity is strengthened.

major comments (3)
  1. Abstract and §3 / Table 3 frame EPRA as attaining “superior performance” (Dice 0.8984, HD95 11.62, fewer missed lesions vs UNet++/DeepLabV3+/TransUNet). Under its intended default protocol, 2D nnU-Net reaches Dice 0.8997, IoU 0.8177, mean HD95 7.98 (median 2.05 vs EPRA 2.00), Lesion F1 0.9376—essentially tied or better on overlap and boundary metrics—while EPRA leads mainly on Recall (0.8887 vs 0.8798) and lesion-level FP count (32 vs 69). The paper correctly notes the protocol/cost difference and excludes nnU-Net from paired tests, but the Abstract’s unqualified superiority language is not supported once the strongest practical baseline is included. Please reframe the claim as a favorable recall/efficiency trade-off against matched-protocol baselines (and competitive with nnU-Net at far lower cost), and align Abstract/Conclusion wording with Tables 3–4.
  2. Training protocol is not fully matched: EPRA is allowed 60 epochs while baselines use a 50-epoch ceiling (early stopping patience=15 applied uniformly; §2.4.2, Fig. 5). EPRA’s best validation Dice is reported at epoch 52. Although early stopping is cited, the extra budget and the fact that the selected checkpoint occurs after the baseline ceiling leave open a systematic training advantage. Either retrain all models under an identical epoch budget / identical early-stopping rule with the same patience clock, or provide a controlled ablation showing that the extra 10 epochs do not drive the reported gains.
  3. All quantitative claims rest on a single-institution retrospective cohort of 167 patients with no external test set (§2.1, §4.2). Patient-level splitting is correctly used, but scanner/protocol/population shift is untested. Given that the central clinical claim is reduced missed lesions and more reliable treatment-eligibility estimation, at least one external multi-center or multi-scanner hold-out (or a clear, quantified domain-shift experiment) is needed before the superiority/robustness language is appropriate for a journal audience. If external data cannot be added in revision, the claim scope must be narrowed throughout Abstract, Results, and Conclusion.
minor comments (6)
  1. Table 1 lists “Number of Epochs: Varies by model” without stating the actual ceilings; make the 50 vs 60 distinction explicit in the table for reproducibility.
  2. Data-availability URL contains a typo (“htps://” missing a “t”); correct to a working link and confirm that code/weights sufficient to reproduce Table 3 are released.
  3. Inter-annotator agreement is not quantified (§4.2 acknowledges this). Even a small double-read kappa or Dice on a subset would strengthen confidence in the ground truth used for all metrics.
  4. Fig. 5 / training curves: clarify whether the CosineAnnealing T_max equals the per-model epoch ceiling and whether any model was allowed to continue past early-stopping for plotting only.
  5. Notation: ASPP dilation rates (1, 6, 12, 18) and R2 unrolling t=2 are stated but not justified by a short sensitivity check; a one-row ablation or reference to the original formulations would help.
  6. Minor language/typos: “Atrous Spatial” double spaces; “Trans UNet” vs “TransUNet” inconsistency; “Yozgat Bozok” double spaces; Abstract omits nnU-Net while body includes it—align for consistency.

Circularity Check

0 steps flagged

No circularity: empirical supervised segmentation with held-out evaluation; architecture and loss choices are design decisions, not self-defining predictions.

full rationale

EPRA U-Net is a composite CNN (EfficientNet-B0 encoder + R2 blocks + ASPP + dual attention) trained with a fixed Tversky loss (α=0.4, β=0.6) on a patient-level split of an in-house DWI dataset. Performance claims (Dice, IoU, Recall, Lesion F1, HD95, missed-lesion reductions) are measured on held-out patients against baselines under a matched protocol; nnU-Net is reported separately under its own protocol. No quantity is fitted from the test set and then re-presented as a prediction; no uniqueness theorem or load-bearing premise is imported solely via overlapping-author citation; R2 unrolling t=2 and ASPP rates follow cited original formulations without redefining the evaluation metrics. Self-citations (e.g., authors’ prior COVID/R-CNN work) are ordinary background and do not close any definitional loop. The paper is therefore self-contained against its external benchmarks; circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

The work is an empirical deep-learning engineering paper. Load-bearing choices are standard supervised-learning assumptions plus a handful of hand-set hyper-parameters and the untested claim that single-center performance transfers. No new physical entities or free constants that redefine the target metric.

free parameters (4)
  • Tversky α, β = α=0.4, β=0.6
    Fixed at α=0.4, β=0.6 to emphasize recall; chosen by clinical preference rather than cross-validated search on the test set, yet directly shapes the reported FN reduction.
  • R2 recurrent unrolling steps t = t=2
    Set to t=2 following the original R2U-Net formulation; controls residual-recurrent depth.
  • ASPP dilation rates = 1,6,12,18
    Rates 1,6,12,18 taken from DeepLab literature; multi-scale context depends on them.
  • Learning-rate / epoch budget = 1e-4 / 60 vs 50 epochs
    1e-4 AdamW + CosineAnnealing; EPRA allowed 60 epochs while baselines used 50; early-stopping patience 15. Affects convergence comparison fairness.
axioms (5)
  • domain assumption Patient-level stratified split prevents data leakage and yields unbiased estimates of clinical performance.
    Stated in §2.1; standard but untested against multi-center shift.
  • domain assumption 2.5D three-slice stacks supply sufficient inter-slice context for acute infarct segmentation without full 3D modeling.
    §2.2 and limitations §4.2; computational convenience assumed adequate.
  • domain assumption Tversky loss with β>α correctly encodes the clinical cost asymmetry of false negatives over false positives for stroke.
    §2.4.2; drives the claimed FN reduction.
  • domain assumption Single-radiologist two-session polygon annotations are sufficiently reliable ground truth.
    §2.1; no kappa or multi-reader study provided.
  • standard math Standard U-Net skip-connection and encoder-decoder inductive biases remain appropriate for DWI infarcts.
    Background architecture assumption throughout §2.4.
invented entities (1)
  • EPRA U-Net architecture no independent evidence
    purpose: Task-specific hybrid that packages EfficientNet-B0 + R2 + ASPP + dual attention + Tversky for DWI infarct segmentation.
    The named model is the paper’s central artifact; it is a composition of prior modules rather than a new mathematical object with independent physical existence.

pith-pipeline@v1.1.0-grok45 · 28925 in / 3210 out tokens · 28477 ms · 2026-07-12T01:29:46.716981+00:00 · methodology

0 comments
read the original abstract

Objective: Accurate identification of acute ischemic infarcts on diffusion-weighted magnetic resonance imaging (DWI) is a critical prerequisite for reliable lesion quantification and effective clinical decision support in the management of cerebrovascular events. Methods: This study presents EPRA U-Net (Efficient Pyramid Residual Attention U-Net), a task-specific integrated architecture for efficient and accurate infarct segmentation of DWI images. In the proposed architecture, an EfficientNet-based encoder was used as a hierarchical feature extractor with a minimized parameterization. In addition, a Residual-Recurrent (R2) block (recurrent unrolling step t = 2, following the original formulation) and Atrous Spatial Pyramid Pooling (ASPP) were integrated to enhance the performance of spatial dependency modeling. Additionally, a dual attention mechanism was incorporated to highlight lesion-related activations while concurrently enabling the suppression of extraneous background responses. To prioritize lesion detection consistent with clinical imperative, a Tversky loss function was adopted, emphasizing the sensitivity of detection over its specificity during the optimization process. Results: Experimental evaluations were conducted utilizing an in-house dataset comprising 167 patients with 4,895 DWI slices; subsequently, the performance of the proposed EPRA U-Net was assessed in comparison with state-of-the-art models, specifically UNet++, DeepLabV3+, and TransUNet. The experimental results suggest that EPRA U-Net attained superior performance, evidenced by a pixel-aggregated Dice of 0.8984, a per-sample Dice of 0.9469, an IoU of 0.8155, a Recall of 0.8887, a Lesion F1 of 0.9378, and an HD95 of 11.62 px. Furthermore, a clear reduction in the rate of missed lesions, specifically by 16%, 25%, and 29%, was observed when compared with UNet++, DeepLabV3+, and TransUNet, respectively.

Figures

Figures reproduced from arXiv: 2607.03568 by Esra Yuce, Hasan Ulutas, Muhammet Emin Sahin, Mustafa Fatih Erkoc, Sengul Dogan, Serkan Kiranyaz, Turker Tuncer.

Figure 1
Figure 1. Figure 1: The patient-level dataset partitioning strategy is clarified; furthermore, representative Diffusion-Weighted Imaging (DWI) slice examples are presented, each accompanied by its corresponding ground-truth data. The prevention of potential data leakage and the assurance of unbiased model evaluation required splitting data at the patient level rather than the slice level. This distinction is critical in biome… view at source ↗
Figure 2
Figure 2. Figure 2: Data pre-processing and augmentation pipeline. Image Format Conversion and Windowing: The original Digital Imaging and Communications in Medicine (DICOM) images underwent conversion into 224×224×3 Portable Network Graphics (PNG) images. Intensity windowing simultaneously applied during this phase. This procedure augmented the contrast between lesion and background regions through the exclusion of superfluo… view at source ↗
Figure 3
Figure 3. Figure 3: Detailed functional components of EPRA U [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the deep learning pipeline for brain infarct segmentation using diffusion [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Curves depicting training and validation losses, as well as Dice scores. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Error analysis visualization. The qualitative comparison of error distributions on the pixel level for four segmentation networks is shown in [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: DSC and HD95 distributions (box plots) [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: presents violin plots that clarify DSC score distributions in greater detail; EPRA U-Net exhibits greater concentration in the high-performance region (DSC > 0.9); this configuration indicates that a substantial proportion of test images attained superior segmentation accuracy. On the other hand, the distribution of the comparison methods is less concentrated and shows longer lower tails, suggesting cases … view at source ↗
Figure 9
Figure 9. Figure 9: shows the Receiver Operating Characteristic (ROC) curves of the pixel-level classification results. EPRA U-Net obtains the largest Area Under the Curve (AUC = 0.9921), suggesting an outstanding discriminatory ability regardless of decision boundaries. This score outperforms other models such as TransUNet (0.9650), DeepLabV3+ (0.9566), and UNet++ (0.9415). Consequently, the EPRA U-Net model can achieve bett… view at source ↗
Figure 10
Figure 10. Figure 10: Precision–Recall curves and Average Precision values [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Radar chart comparison of all models across five core evaluation metrics (Dice, Precision, Recall, IoU, [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: presents the confusion matrices for each model, assessed at the pixel level, and these matrices serve to clarify the patterns of prediction and misclassification across distinct classes [41]. EPRA U-Net outperformed in the infarct (positive) class regarding its true positive rate, which was 0.889, followed by UNet++, TransUNet, and DeepLabV3+, which scored 0.853, 0.837, and 0.828, respectively. A high tru… view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative comparison on five representative test cases [PITH_FULL_IMAGE:figures/full_fig_p016_13.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

62 extracted references · 25 canonical work pages · 4 internal anchors

  1. [1]

    Current concepts on magnetic resonance imaging (MRI) perfusion- diffusion assessment in acute ischaemic stroke: a review & an update for the clinicians.,

    E. Roldán-Valadéz and M. López-Mejía, “Current concepts on magnetic resonance imaging (MRI) perfusion- diffusion assessment in acute ischaemic stroke: a review & an update for the clinicians.,” Indian J Med Res. , vol. 140, no. 6, pp. 717–28, Dec. 2014, PMID: 25758570; PMCID: PMC4365345

    2014

  2. [2]

    World Stroke Organization: Global Stroke Fact Sheet 2025,

    V. L. Feigin et al., “World Stroke Organization: Global Stroke Fact Sheet 2025,” International Journal of Stroke, vol. 20, no. 2, pp. 132–144, Dec. 2024, doi: 10.1177/17474930241308142

    work page doi:10.1177/17474930241308142 2025

  3. [3]

    TeleStroke: real -time stroke detection with federated learning and YOLOv8 on edge devices,

    A. Elhanashi, P. Dini, S. Saponara, and Q. Zheng, “TeleStroke: real -time stroke detection with federated learning and YOLOv8 on edge devices,” Journal of Real -Time Image Processing, vol. 21, no. 4, Jun. 2024, doi: 10.1007/s11554-024-01500-1

    work page doi:10.1007/s11554-024-01500-1 2024

  4. [4]

    Diffusion weighted imaging in acute ischemic stroke: A review of its interpretation pitfalls and advanced diffusion imaging application,

    N. Nagaraja, “Diffusion weighted imaging in acute ischemic stroke: A review of its interpretation pitfalls and advanced diffusion imaging application,” Journal of the Neurological Sciences , vol. 425, pp. 117435 –117435, Apr. 2021, doi: 10.1016/j.jns.2021.117435

    work page doi:10.1016/j.jns.2021.117435 2021

  5. [5]

    Artificial Intelligence in Stroke Care: A Narrative Review of Diagnostic, Predictive, and Workflow Applications,

    V. T. Heeralal, S. E. Chadee, B. Ilyaev, R. Ilyaev, and S. Ilyayeva, “Artificial Intelligence in Stroke Care: A Narrative Review of Diagnostic, Predictive, and Workflow Applications,” Cureus , vol. 17, no. 9, Sep. 2025, doi: 10.7759/cureus.93430

    work page doi:10.7759/cureus.93430 2025

  6. [6]

    U -Net: Convolutional Networks for Biomedical Image Segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U -Net: Convolutional Networks for Biomedical Image Segmentation,” in Lecture notes in computer science, Springer Science+Business Media, 2015, pp. 234–241. doi: 10.1007/978-3-319-24574-4_28

    doi:10.1007/978-3-319-24574-4_28 2015

  7. [7]

    UNet++: A Nested U-Net Architecture for Medical Image Segmentation,

    Z. Zhou, M. M. R. Siddiquee, N. Tajbakhsh, and J. Liang, “UNet++: A Nested U-Net Architecture for Medical Image Segmentation,” Lecture notes in computer science , vol. 11045, pp. 3 –11, Jan. 2018, doi: 10.1007/978 -3- 030-00889-5_1

    doi:10.1007/978 2018

  8. [8]

    Attention U-Net: Learning Where to Look for the Pancreas,

    O. Oktay et al., “Attention U-Net: Learning Where to Look for the Pancreas,” arXiv (Cornell University), Apr. 2018, doi: 10.48550/arxiv.1804.03999

    Pith/arXiv arXiv doi:10.48550/arxiv.1804.03999 2018

  9. [9]

    Recurrent Residual Convolutional Neural Network based on U -Net (R2U-Net) for Medical Image Segmentation,

    M. Z. Alom, Md. M. Hasan, C. Yakopcic, T. M. Taha, and V. K. Asari, “Recurrent Residual Convolutional Neural Network based on U -Net (R2U-Net) for Medical Image Segmentation,” arXiv (Cornell University), Feb. 2018, doi: 10.48550/arxiv.1802.06955

    Pith/arXiv arXiv doi:10.48550/arxiv.1802.06955 2018

  10. [10]

    TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation,

    J. Chen et al., “TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation,” arXiv (Cornell University), Feb. 2021, doi: 10.48550/arxiv.2102.04306

    Pith/arXiv arXiv doi:10.48550/arxiv.2102.04306 2021

  11. [11]

    Swin -Unet: Unet-like Pure Transformer for Medical Image Segmentation,

    H. Cao et al., “Swin -Unet: Unet-like Pure Transformer for Medical Image Segmentation,” arXiv (Cornell University), May 2021, doi: 10.48550/arxiv.2105.05537

    Pith/arXiv arXiv doi:10.48550/arxiv.2105.05537 2021

  12. [12]

    Medical Image Segmentation Review: The Success of U-Net,

    R. Azad et al., “Medical Image Segmentation Review: The Success of U-Net,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12. IEEE Computer Society, pp. 10076 –10095, Aug. 21, 2024. doi: 10.1109/tpami.2024.3435571

    doi:10.1109/tpami.2024.3435571 2024

  13. [13]

    3D U -Net: Learning Dense Volumetric Segmentation from Sparse Annotation,

    Ö. Çiçek, A. Abdulkadir, S. S. Lienkamp, T. Brox, and O. Ronneberger, “3D U -Net: Learning Dense Volumetric Segmentation from Sparse Annotation,” in Lecture notes in computer science, Springer Science+Business Media, 2016, pp. 424–432. doi: 10.1007/978-3-319-46723-8_49

    doi:10.1007/978-3-319-46723-8_49 2016

  14. [14]

    V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation,

    F. Milletarì, N. Navab, and S. Ahmadi, “V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation,” pp. 565–571, Oct. 2016, doi: 10.1109/3dv.2016.79

    doi:10.1109/3dv.2016.79 2016

  15. [15]

    Efficient multi -scale 3D CNN with fully connected CRF for accurate brain lesion segmentation,

    K. Kamnitsas et al., “Efficient multi -scale 3D CNN with fully connected CRF for accurate brain lesion segmentation,” Medical Image Analysis, vol. 36, pp. 61–78, Oct. 2016, doi: 10.1016/j.media.2016.10.004

    doi:10.1016/j.media.2016.10.004 2016

  16. [16]

    nnU-Net: a self -configuring method for deep learning -based biomedical image segmentation,

    F. Isensee, P. F. Jaeger, S. A. A. Kohl, J. Petersen, and K. H. Maier-Hein, “nnU-Net: a self -configuring method for deep learning -based biomedical image segmentation,” Nature Methods, vol. 18, no. 2, pp. 203 –211, Dec. 2020, doi: 10.1038/s41592-020-01008-z

    doi:10.1038/s41592-020-01008-z 2020

  17. [17]

    Encoder -Decoder with Atrous Separable Convolution for Semantic Image Segmentation,

    L. -C. Chen, Y. Zhu, G. Papandreou, F. Schroff, and H. Adam, “Encoder -Decoder with Atrous Separable Convolution for Semantic Image Segmentation,” in Lecture notes in computer science, Springer Science+Business Media, 2018, pp. 833–851. doi: 10.1007/978-3-030-01234-2_49

    doi:10.1007/978-3-030-01234-2_49 2018

  18. [18]

    Pyramid Scene Parsing Network,

    H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia, “Pyramid Scene Parsing Network,” pp. 6230 –6239, Jul. 2017, doi: 10.1109/cvpr.2017.660

    doi:10.1109/cvpr.2017.660 2017

  19. [19]

    Attention is All you Need,

    A. Vaswani et al., “Attention is All you Need,” in Adv. Neural Inf. Process. Syst., 2017

    2017

  20. [20]

    CBAM: Convolutional Block Attention Module,

    S. Woo, J. Park, J. Lee, and I. S. Kweon, “CBAM: Convolutional Block Attention Module,” in Lecture notes in computer science, Springer Science+Business Media, 2018, pp. 3 –19. doi: 10.1007/978-3-030-01234-2_1

    doi:10.1007/978-3-030-01234-2_1 2018

  21. [21]

    Dual Attention Network for Scene Segmentation,

    J. Fu et al., “Dual Attention Network for Scene Segmentation,” pp. 3141 –3149, Jun. 2019, doi : 10.1109/cvpr.2019.00326

    doi:10.1109/cvpr.2019.00326 2019

  22. [22]

    Generalised Dice Overlap as a Deep Learning Loss Function for Highly Unbalanced Segmentations,

    C. H. Sudre, W. Li, T. Vercauteren, S. Ourselin, and M. J. Cardoso, “Generalised Dice Overlap as a Deep Learning Loss Function for Highly Unbalanced Segmentations,” Lecture notes in computer science, vol. 2017, pp. 240–248, Jan. 2017, doi: 10.1007/978-3-319-67558-9_28

    work page doi:10.1007/978-3-319-67558-9_28 2017

  23. [23]

    Tackling the class imbalance problem of deep learning-based head and neck organ segmentation,

    E. Tappeiner, M. Welk, and R. Schubert, “Tackling the class imbalance problem of deep learning-based head and neck organ segmentation,” International Journal of Computer Assisted Radiology and Surgery, vol . 17, no. 11, pp. 2103–2111, May 2022, doi: 10.1007/s11548-022-02649-5

    work page doi:10.1007/s11548-022-02649-5 2022

  24. [24]

    Tversky Loss Function for Image Segmentation Using 3D Fully Convolutional Deep Networks,

    S. S. M. Salehi, D. Erdoğmuş, and A. Gholipour, “Tversky Loss Function for Image Segmentation Using 3D Fully Convolutional Deep Networks,” Lecture notes in computer science, pp. 379 –387, Jan. 2017, doi: 10.1007/978-3-319-67389-9_44

    doi:10.1007/978-3-319-67389-9_44 2017

  25. [25]

    A comprehensive survey of loss functions and metrics in deep learning,

    J. Terven, D. Córdova‐Esparza, J.-A. Romero-González, A. Ramírez-Pedraza, and E. A. Chávez‐Urbiola, “A comprehensive survey of loss functions and metrics in deep learning,” Artificial Intelligence Review, vol. 58, no. 7, Apr. 2025, doi: 10.1007/s10462-025-11198-7

    doi:10.1007/s10462-025-11198-7 2025

  26. [26]

    A Novel Focal Tversky Loss Function With Improved Attention U-Net for Lesion Segmentation,

    N. Abraham and N. Khan, “A Novel Focal Tversky Loss Function With Improved Attention U-Net for Lesion Segmentation,” pp. 683–687, Apr. 2019, doi: 10.1109/isbi.2019.8759329

    doi:10.1109/isbi.2019.8759329 2019

  27. [27]

    The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS),

    B. Menze et al., “The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS),” IEEE Transactions on Medical Imaging, vol. 34, no. 10, pp. 1993–2024, Dec. 2014, doi: 10.1109/tmi.2014.2377694

    doi:10.1109/tmi.2014.2377694 1993

  28. [28]

    Deep Learning-Driven Segmentation of Ischemic Stroke Lesions Using Multi-Channel MRI

    A. Rahman et al., “Deep Learning -Driven Segmentation of Ischemic Stroke Lesions Using Multi -Channel MRI,” arXiv (Cornell University), Jan. 2025, doi: 10.48550/arxiv.2501.02287

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2501.02287 2025

  29. [29]

    Random effects during training: Implications for deep learning -based medical image segmentation,

    J. Åkesson, J. Töger, and E. Heiberg, “Random effects during training: Implications for deep learning -based medical image segmentation,” Computers in Biology and Medicine, vol. 180, pp. 108944 –108944, Aug. 2024, doi: 10.1016/j.compbiomed.2024.108944

    work page doi:10.1016/j.compbiomed.2024.108944 2024

  30. [30]

    Robust chest CT image segmentation of COVID -19 lung infection based on limited data,

    D. Müller, I. Soto‐Rey, and F. Krämer, “Robust chest CT image segmentation of COVID -19 lung infection based on limited data,” Informatics in Medicine Unlocked, vol. 25, pp. 100681 –100681, Jan. 2021, doi: 10.1016/j.imu.2021.100681

    work page doi:10.1016/j.imu.2021.100681 2021

  31. [31]

    Detection, Diagnosis and Treatment of Acute Ischemic Stroke: Current and Future Perspectives,

    S. Patil, R. Rossi, D. Jabrah, and K. Doyle, “Detection, Diagnosis and Treatment of Acute Ischemic Stroke: Current and Future Perspectives,” Frontiers in Medical Technology, vol. 4. Frontiers Media, Jun. 24, 2022. doi: 10.3389/fmedt.2022.748949

    work page doi:10.3389/fmedt.2022.748949 2022

  32. [32]

    Make Sense: Free to use online annotation tool,

    P. Skalski, "Make Sense: Free to use online annotation tool," GitHub, 2019. [Online]. Available: https://github.com/SkalskiP/make-sense

    2019

  33. [33]

    Effect of data leakage in brain MRI classification using 2D convolutional neural networks,

    E. Yağış et al. , “Effect of data leakage in brain MRI classification using 2D convolutional neural networks,” Scientific Reports, vol. 11, no. 1, Nov. 2021, doi: 10.1038/s41598-021-01681-w

    doi:10.1038/s41598-021-01681-w 2021

  34. [34]

    A Flexible 2.5D Medical Image Segmentation Approach with In-Slice and Cross-Slice Attention

    A. Kumar et al., “A Flexible 2.5D Medical Image Segmentation Approach with In -Slice and Cross -Slice Attention,” arXiv (Cornell University), Apr. 2024, doi: 10.48550/arxiv.2405.00130

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2405.00130 2024

  35. [35]

    EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,

    M. Tan and Q. V. Le, “EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,” arXiv (Cornell University), May 2019, doi: 10.48550/arxiv.1905.11946

    Pith/arXiv arXiv doi:10.48550/arxiv.1905.11946 2019

  36. [36]

    Deep Residual Learning for Image Recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” pp. 770 –778, Jun. 2016, doi: 10.1109/cvpr.2016.90

    doi:10.1109/cvpr.2016.90 2016

  37. [37]

    DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs,

    L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. Yuille, “DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 40, no. 4, pp. 834–848, Apr. 2017, doi: 10.1109/tpami.2017.2699184

    doi:10.1109/tpami.2017.2699184 2017

  38. [38]

    Densely Connected Recurrent Residual (Dense R2UNet) Convolutional Neural Network for Segmentation of Lung CT Images

    K. Dutta, “Densely Connected Recurrent Residual (Dense R2UNet) Convolutional Neural Network for Segmentation of Lung CT Images,” arXiv (Cornell University) , Feb. 2021, doi: 10.48550/arxiv.2102.00663

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2102.00663 2021

  39. [39]

    A semi -automatic segmentation method for meningioma developed using a variational approach model,

    L. Burrows, J. Patel, A. I. Islim, M. D. Jenkinson, S. J. Mills, and K. Chen, “A semi -automatic segmentation method for meningioma developed using a variational approach model,” The Neuroradiology Journal, vol. 37, no. 2, pp. 199–205, Dec. 2023, doi: 10.1177/19714009231224442

    work page doi:10.1177/19714009231224442 2023

  40. [40]

    Image Segmentation Evaluation With the Dice Index: Methodological Issues,

    M. L. Seghier, “Image Segmentation Evaluation With the Dice Index: Methodological Issues,” International Journal of Imaging Systems and Technology, vol. 34, no. 6, Oct. 2024, doi: 10.1002/ima.23203

    doi:10.1002/ima.23203 2024

  41. [41]

    Towards a guideline for evaluation metrics in medical image segmentation,

    D. Müller, I. Soto‐Rey, and F. Krämer, “Towards a guideline for evaluation metrics in medical image segmentation,” BMC Research Notes, vol. 15, no. 1. BioMed Central, Jun. 20, 2022. doi: 10.1186/s13104 -022- 06096-y

    work page doi:10.1186/s13104 2022

  42. [42]

    Common Limitations of Image Processing Metrics: A Picture Story,

    A. Reinke et al, “Common Limitations of Image Processing Metrics: A Picture Story,” arXiv (Cornell University), vol. 2021, Apr. 2021, doi: 10.48550/arxiv.2104.05642

    Pith/arXiv arXiv doi:10.48550/arxiv.2104.05642 2021

  43. [43]

    Detection and classification of COVID-19 by using faster R-CNN and mask R-CNN on CT images,

    M. E. Şahin, H. Ulutaş, E. Yüce, and M. F. Erkoç, “Detection and classification of COVID-19 by using faster R-CNN and mask R-CNN on CT images,” Neural Computing and Applications, vol. 35, no. 18, pp. 13597–13611, Mar. 2023, doi: 10.1007/s00521-023-08450-y

    work page doi:10.1007/s00521-023-08450-y 2023

  44. [44]

    U -Net and Its Variants for Medical Image Segmentation: A Review of Theory and Applications,

    N. Siddique, S. Paheding, C. Elkin, and V. Devabhaktuni, “U -Net and Its Variants for Medical Image Segmentation: A Review of Theory and Applications,” IEEE Access, vol. 9, pp . 82031–82057, Jan. 2021, doi: 10.1109/access.2021.3086020

    doi:10.1109/access.2021.3086020 2021

  45. [45]

    Deep learning -based approach for detecting COVID -19 in chest X -rays,

    M. E. Şahin, “Deep learning -based approach for detecting COVID -19 in chest X -rays,” Biomedical Signal Processing and Control, vol. 78, pp. 103977–103977, Jul. 2022, doi: 10.1016/j.bspc.2022.103977

    work page doi:10.1016/j.bspc.2022.103977 2022

  46. [46]

    Calibrating the Dice Loss to Handle Neural Network Overconfidence for Biomedical Image Segmentation,

    M. Yeung, L. Rundo, N. Yang, E. Sala, C. Schönlieb, and G. Yang, “Calibrating the Dice Loss to Handle Neural Network Overconfidence for Biomedical Image Segmentation,” Journal of Digital Imaging, vol. 36, no. 2, pp. 739–752, Dec. 2022, doi: 10.1007/s10278-022-00735-3

    doi:10.1007/s10278-022-00735-3 2022

  47. [47]

    Retinoblastoma Detection via Image Processing and Interpretable Artificial Intelligence Techniques,

    S. Duraivenkatesh, A. Narayan, V. Srikanth, and A. F. Made, “Retinoblastoma Detection via Image Processing and Interpretable Artificial Intelligence Techniques,” medRxiv (Cold Spring Harbor Laboratory), May 2023, doi: 10.1101/2023.05.02.23289419

    work page doi:10.1101/2023.05.02.23289419 2023

  48. [48]

    From coarse to fine: a deep 3D probability volume contours framework for tumour segmentation and dose painting in PET images,

    W. Zhang and S. Ray, “From coarse to fine: a deep 3D probability volume contours framework for tumour segmentation and dose painting in PET images,” Frontiers in Radiology, vol. 3, Sep. 2023, doi: 10.3389/fradi.2023.1225215

    work page doi:10.3389/fradi.2023.1225215 2023

  49. [49]

    Individual Comparisons by Ranking Methods,

    F. Wilcoxon, “Individual Comparisons by Ranking Methods,” Biometrics Bulletin, vol. 1, no. 6, pp. 80 –80, Dec. 1945, doi: 10.2307/3001968

    doi:10.2307/3001968 1945

  50. [50]

    Parametric versus non -parametric statistics in the analysis of randomized trials with non - normally distributed data,

    A. J. Vickers, "Parametric versus non -parametric statistics in the analysis of randomized trials with non - normally distributed data," BMC Medical Research Methodology, vol. 5, no. 1, p. 35, 2005, doi: 10.1186/1471 - 2288-5-35

    doi:10.1186/1471 2005

  51. [51]

    Note on the Sampling Error of the Difference Between Correlated Proportions or Percentages,

    Q. McNemar, “Note on the Sampling Error of the Difference Between Correlated Proportions or Percentages,” Psychometrika, vol. 12, no. 2, pp. 153–157, Jun. 1947, doi: 10.1007/bf02295996

    doi:10.1007/bf02295996 1947

  52. [52]

    Efron and R

    B. Efron and R. Tibshirani, An Introduction to the Bootstrap. 1994. doi: 10.1201/9780429246593

    doi:10.1201/9780429246593 1994

  53. [53]

    Cohen, Statistical Power Analysis for the Behavioral Sciences, 2nd ed

    J. Cohen, Statistical Power Analysis for the Behavioral Sciences, 2nd ed. Hillsdale, NJ: Lawrence Erlbaum Associates, 1988

    1988

  54. [54]

    Resolving power: a general approach to compare the distinguishing ability of threshold -free evaluation metrics,

    C. A. Beam, “Resolving power: a general approach to compare the distinguishing ability of threshold -free evaluation metrics,” Machine Learning, vol. 114, no. 1, Jan. 2025, doi: 10.1007/s10994 -024-06723-8

    work page doi:10.1007/s10994 2025

  55. [55]

    Risk-based Evaluation of ML Classification Methods Used for Medical Devices,

    M. Haimerl and C. Reich, “Risk-based Evaluation of ML Classification Methods Used for Medical Devices,” Research Square (Research Square), Sep. 2023, doi: 10.21203/rs.3.rs -3317894/v1

    doi:10.21203/rs.3.rs 2023

  56. [56]

    Challenges in Deep Learning-Based Small Organ Segmentation: A Benchmarking Perspective for Medical Research with Limited Datasets

    P. M. Konrad, A. -A. Popa, Y. Sabzehmeidani, L. Zhong, E. A. Liehn, and S. Ayvaz, “Challenges in Deep Learning-Based Small Organ Segmentation: A Benchmarking Perspective for Medical Research with Limited Datasets,” arXiv (Cornell University), Sep. 2025, doi: 10.48550/arxiv.2509.05892

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2509.05892 2025

  57. [57]

    Accuracy of CT perfusion ischemic core volume and location estimation: A comparison between four ischemic core estimation approaches using syngo.via,

    J. W. Hoving et al., “Accuracy of CT perfusion ischemic core volume and location estimation: A comparison between four ischemic core estimation approaches using syngo.via,” PLoS ONE, vol. 17, no. 8, Aug. 2022, doi: 10.1371/journal.pone.0272276

    work page doi:10.1371/journal.pone.0272276 2022

  58. [58]

    Automated identification of thrombectomy amenable vessel occlusion on computed tomography angiography using deep learning,

    J. H. Han et al., “Automated identification of thrombectomy amenable vessel occlusion on computed tomography angiography using deep learning,” Frontiers in Neurology, vol. 15, Jul. 2024, doi: 10.3389/fneur.2024.1442025

    work page doi:10.3389/fneur.2024.1442025 2024

  59. [59]

    Early detection of white matter hyperintensities using SHIVA‐WMH detector,

    A. Tsuchida, P. Boutinaud, V. Verrecchia, C. Tzourio, S. Debette, and M. Joliot, “Early detection of white matter hyperintensities using SHIVA‐WMH detector,” Human Brain Mapping, vol. 45, no. 1, Dec. 2023, doi: 10.1002/hbm.26548

    work page doi:10.1002/hbm.26548 2023

  60. [60]

    Automatic Liver Segmentation from Multiphase CT Using Modified SegNet and ASPP Module,

    P. V. Nayantara, S. Kamath, R. Kadavigere, and K. Manjunath, “Automatic Liver Segmentation from Multiphase CT Using Modified SegNet and ASPP Module,” SN Computer Science, vol. 5, no. 4, Mar. 2024, doi: 10.1007/s42979-024-02719-2

    work page doi:10.1007/s42979-024-02719-2 2024

  61. [61]

    Rapid risk stratification of acute ischemic stroke patients in the emergency department: the incremental prognostic role of left atrial reservoir strain

    Sonaglioni, A., Di Cara, M., Nicolosi, G. L., Eusebio, A., Bordonali, M., Santalucia, P., & Lombardo, M. "Rapid risk stratification of acute ischemic stroke patients in the emergency department: the incremental prognostic role of left atrial reservoir strain." Journal of Stroke and Cerebrovascular Diseases 30.11 (2021): 106100

    2021

  62. [62]

    Weak Edge Target Segmentation Network Based on Dual Attention Mechanism,

    N. Wu, D. Jia, Z. Li, and Z. He, “Weak Edge Target Segmentation Network Based on Dual Attention Mechanism,” Applied Sciences, vol. 14, no. 19, pp. 8963–8963, Oct. 2024, doi: 10.3390/app14198963

    work page doi:10.3390/app14198963 2024