Pith. sign in

REVIEW 5 major objections 6 minor 34 references

Automated MRI Tumor Segmentation using hybrid U-Net with Transformer and Efficient Attention

T0 review · 5 major / 6 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 reports that a hybrid U-Net with a transformer bottleneck and efficient attention reaches 0.7636 Dice and 0.7357 IoU on a local hospital MRI tumor dataset of 6,080 images, arguing that site-specific, small clinical datasets can…

desk verdict Interesting architecture integration and a useful local data pipeline, but the reported Dice/IoU are training metrics, so the generalization claim is unsupported. read the letter →

arxiv 2506.15562 v2 pith:JGVUTASI submitted 2025-06-18 eess.IV cs.CV

Syed Haider Ali , Asrar Ahmad , Muhammad Ali , Asifullah Khan , Nadeem Shaukat This is my paper
classification eess.IVcs.CV
keywords braintumorsegmentationMRIU-NettransformerbottleneckefficientattentionSqueeze-and-ExcitationCBAMclinicaldataset
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

This paper argues that a U-Net with a transformer bottleneck and several attention modules can segment brain tumors in MRI scans from a local hospital with a Dice score of 0.7636 and an IoU of 0.7357 while training on only 6,080 largely augmented images. The authors position the work against public-benchmark models trained on hundreds of thousands of slices, claiming that site-specific, privacy-preserving data better reflect local patient populations and that their hybrid architecture is accurate enough for clinical deployment despite limited resources. The central message is that a comparatively small, heterogeneous clinical dataset can support competitive segmentation when the encoder carries pretrained ImageNet weights and the network combines local convolutional features with global transformer context.

What carries the argument

The load-bearing object is the hybrid encoder-bottleneck-decoder stack. The encoder is a ResNet-50 pretrained on ImageNet, with Squeeze-and-Excitation and CBAM added to recalibrate channel and spatial features. The deepest 32 by 32 by 1024 feature map is flattened into 1024 tokens and fed through four transformer blocks with 8-head self-attention to capture global context. In the decoder, CBAM refines skip connections, an efficient attention module computes per-channel gates with linear complexity, and ResNeXt blocks with grouped convolutions fuse multi-scale patterns; the network is trained end-to-end with binary cross-entropy plus soft Dice loss.

What would settle it

Run the trained model on a held-out set of slices from patients not used in training, using the same preprocessing and augmentation; if the Dice score on that split falls substantially below 0.7636, the central claim of competitive generalizable segmentation is not supported.

Watch

Extended reading notes

Core claim

The paper reports that the complete hybrid model—a ResNet-50 encoder with Squeeze-and-Excitation and CBAM, a four-block transformer bottleneck, and a decoder with efficient attention, SE, CBAM, and ResNeXt blocks—reaches 0.7636 Dice and 0.7357 IoU on the local MRI dataset, compared with 0.7190 and 0.6818 for the plain U-Net baseline and 0.7286 and 0.7018 for a U-Net with only a transformer bottleneck. The authors take this as evidence that each added module contributes and that the combined design balances local feature extraction with long-range context. They interpret the result as competitive with published models such as HTTU-Net, H-DenseUNet, and MM-BiFPN even though those used far larger public datasets.

Load-bearing premise

The load-bearing premise is that the reported Dice and IoU numbers, which are shown as training curves, describe how well the model will segment MRI slices it has not seen before; the paper describes no held-out test set or patient-level split.

Editorial extensions

If this is right

  • If the reported figures hold on held-out scans, the same architecture could be transferred to other local hospital datasets with modest GPU budgets.
  • The module-wise comparison shows that adding the transformer bottleneck and attention modules raises Dice by roughly 0.04 over the baseline, suggesting measurable gains from combining local convolutional features with global context.
  • Pretrained ImageNet weights plus checkpointed training offer a practical recipe for resource-limited sites that cannot train large models from scratch.
  • The comparison against BraTS- and LiTS-trained models implies that dataset provenance matters as much as architecture when the goal is clinical deployment.

Reading between the lines

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

  • Extending the paper's argument, the reported metrics are shown as training curves, so a fairer test of the clinical-deployment claim would require a held-out patient-level split; this is my inference, not a claim the paper makes.
  • The augmentation pipeline expands roughly 1,000 images into 6,080, but augmented copies share label structure, so the model may be overfit to six patients; testing on truly unseen patients would clarify this.
  • A testable extension is to run the same hybrid architecture on a public benchmark under a matched small-data regime to isolate whether the gains come from the architecture itself or from characteristics of the local dataset.
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

5 major / 6 minor

Summary. The paper proposes a hybrid U-Net architecture for automated MRI tumor segmentation, combining a ResNet-50 encoder with a transformer bottleneck and several attention modules (efficient attention, SE, CBAM, ResNeXt). The model is trained on a local clinical dataset of six patients (~1,050 original MRI slices, augmented to 6,080 images) using Kaggle GPUs with checkpointing. The authors report a Dice score of 0.7636 and IoU of 0.7357, and they argue that these results demonstrate competitive performance despite limited data. The paper includes a comparison with published methods on BraTS/LiTS and an ablation of three model variants.

Significance. If the reported results were obtained on a proper held-out test set, the paper would be a useful feasibility study on site-specific model development for resource-constrained clinical settings. The data pipeline, privacy considerations, and use of pretrained encoders and checkpointing are practical contributions. However, the central quantitative claim is currently unsupported because the reported Dice and IoU values are explicitly described as training-performance metrics. The lack of any out-of-sample evaluation, patient-level split, cross-validation, or error bars means the paper does not establish generalization, which is the core premise of the study. The architectural novelty is also incremental, as it combines existing modules (transformer bottleneck, SE, CBAM, ResNeXt, efficient attention) without a systematic analysis of their individual contributions beyond a small ablation.

major comments (5)
  1. [Section 3, Figure 5] The only reported quantitative results (Dice 0.7636, IoU 0.7357, precision 0.9736, recall 0.9756) are presented in Figure 5, whose caption reads 'Training performance metrics over epochs.' The text states that 'by epoch 29, these metrics improved' without ever describing a held-out test set, a validation split, a patient-level split, or cross-validation. Training-set metrics do not support the abstract's claim of 'competitive performance' on a local dataset, because they can reflect memorization rather than generalization. This is a load-bearing issue: the paper's central claim is empirical, and the evidence for it is absent.
  2. [Section 2.1 and Section 3] The dataset consists of MRI scans from six patients, with approximately 1,050 original slices. The augmentation pipeline expands this to 6,080 images, and Section 3 states that the model was 'trained end-to-end on a clinically sourced MRI dataset comprising 6,080 images.' The paper does not explain how the data were partitioned into training, validation, and test sets. If the 6,080 augmented images were used for both training and the reported evaluation, augmented copies of the same original slice could appear in both partitions, causing data leakage. A patient-level split is essential with only six patients, and no such split is described.
  3. [Table 2] The comparison in Table 2 is not meaningful for the claims made. HTTU-Net, H-DenseUNet, and MM-BiFPN are evaluated on BraTS or LiTS datasets with 58,000–221,400 slices, while the proposed hybrid model is evaluated on 6,080 local slices (and apparently on training metrics only). These comparisons involve different datasets, different modalities (CT for LiTS), and different evaluation protocols, so the reported Dice/IoU values cannot support the conclusion that the proposed model is 'competitive' or that it demonstrates a 'generalization gap.'
  4. [Table 3] The ablation in Table 3 compares 'U-Net (Baseline),' 'U-Net with Transformer Bottleneck,' and the 'Final Hybrid Model,' but no evaluation methodology is described for these variants. Since the paper never mentions a test split, all values in Table 3 likely reflect training performance. The improvements from 0.7190 to 0.7286 to 0.7636 Dice could be due to increased model capacity fitting the training set better, rather than improved generalization. Additionally, with only six patients, reporting a single point estimate without error bars or patient-level cross-validation is insufficient to establish that the architectural modifications are beneficial.
  5. [Section 2.3, Eq. (3)] The composite loss in Eq. (3) includes a weight λ that is described as 'empirically optimized,' but its value is never reported. The paper also omits details such as the optimizer, learning rate, learning-rate schedule, number of training epochs (29 appears only in Section 3), batch size details beyond 'maximum batch size of 8,' and any early-stopping criterion. These omissions substantially reduce the reproducibility of the experiments, which is especially important given that the central evaluation is already under-specified.
minor comments (6)
  1. [Section 2.1] The sentence 'All images were acquired at a voxel resolution of acquired at a voxel resolution of 0.84 × 0.84 × 2.00mm' contains a duplicated phrase; it should be corrected to a single 'acquired at a voxel resolution of 0.84 × 0.84 × 2.00 mm.'
  2. [Figure 5 and Section 3] The in-text references to Figure 5 are inconsistent. Section 3 says 'precision and recall increased ... as shown in Figure 5b' and 'the hybrid loss also steadily decreased ... as depicted in Figure 5c,' but the figure caption lists (a) Dice/IoU, (b) Precision/Recall, and (c) Loss. The text and caption should be reconciled.
  3. [Section 2.2] In the description of the transformer bottleneck, the text says the 32×32×1024 tensor 'is reshaped into a sequence of 1024 tokens (i.e., 1024 × 1024).' The parenthetical is confusing: a 32×32 tensor has 1024 spatial positions, each with 1024 channels, so the token sequence should be 1024 tokens of dimension 1024. The notation '1024 × 1024' should be clarified.
  4. [References] References [8], [9], and [12] are authored or co-authored by Asifullah Khan, and the Introduction leans on these preprints. Self-citation is not inherently problematic, but the authors should consider citing published peer-reviewed versions if they exist, to strengthen the literature basis.
  5. [Table 2] The naming is inconsistent: the text in Section 2.2 refers to 'Two-Track UNet' [24], while Table 2 calls it 'HTTU-Net.' Please use the same name consistently, matching the cited paper.
  6. [Abstract and Section 3] The abstract reports Dice '76.4%' and IoU '73.6%,' while the body uses 0.7636 and 0.7357. Use one consistent format (either proportions or percentages) throughout.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported Dice/IoU are training metrics, so the central performance claim reduces to the optimized training objective; self-citations are not load-bearing.

  1. fitted input called prediction [Section 3, Figure 5 caption; Tables 2 and 3; Abstract]
    "As depicted in Figure 5a, the model exhibited consistent convergence over 29 epochs with a batch size of 8. The initial Dice Score and IoU were 0.4548 and 0.3519, respectively. By epoch 29, these metrics improved to 0.7636 and 0.7357, respectively. Figure 5: Training performance metrics over epochs."

    The only reported outcome metrics (Dice 0.7636, IoU 0.7357) are the final-epoch values from Figure 5, whose caption explicitly says 'Training performance metrics over epochs.' Section 3 describes no held-out test set, validation split, patient-level split, or cross-validation, so these numbers are computed on the same augmented 6,080-image data used to train the model. Furthermore, Eq. (3) trains the network with L_overall = L_BCE + lambda*L_Dice, meaning the soft Dice on the training set is the very quantity being optimized by gradient descent. Reporting final training Dice/IoU as 'competitive performance despite limited data' is therefore reporting the fitted value of the training objective, not an independent prediction of generalization.

full rationale

The paper is an empirical pipeline rather than a derivation chain, and its architectural pieces (SE, CBAM, ResNeXt, efficient attention, transformer bottleneck) are standard modules defined independently of the target result. The self-citations in references [8], [9], and [12] are used for motivation and for naming existing hybrid models; none of them supplies the reported Dice/IoU or forces the architecture choice in a way that reduces the result to those citations. However, the central performance claim is circular in a narrower sense: the reported Dice/IoU are explicitly training metrics from Figure 5, and no out-of-sample evaluation is described. Because the training loss directly optimizes BCE plus soft Dice on the same augmented dataset, the reported values are the optimized training objective, not a held-out prediction. This is a fitted-input-called-prediction issue and makes the central generalization claim unsupported; the remainder of the paper does not exhibit additional circular steps.

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

The central claim rests on the assumption that training metrics reflect generalization, plus standard deep learning components. No new physical entities are introduced. The only fitted parameter is the loss weight lambda, which is not reported.

free parameters (3)
  • loss weighting lambda = not reported
    Empirically optimized in the composite loss L_overall = L_BCE + lambda * L_Dice (Equation 3), but the chosen value is not stated.
  • SE reduction ratio r = 16 (standard, not tuned)
    Channel reduction ratio in Squeeze-and-Excitation blocks; affects channel recalibration and is chosen by hand.
  • transformer blocks and heads = 4 blocks, 8 heads
    Hand-chosen architecture hyperparameters; no ablation study justifies these values.
assumptions (3)
  • domain assumption The reported Dice and IoU on the training set are valid proxies for clinical segmentation performance on unseen patients.
    The paper relies on this to claim competitive performance without a held-out test set or cross-validation.
  • domain assumption The augmented local dataset is representative of clinical heterogeneity and does not introduce leakage between patients.
    The dataset comes from six patients, and augmentation multiplies it; no patient-level split is described.
  • standard math Standard softmax attention and U-Net operations are accepted as background.
    Equation 1 for attention and the U-Net encoder-decoder structure are standard, cited results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Automated MRI Tumor Segmentation using hybrid U-Net with Transformer and Efficient Attention." pith.science (2026). https://pith.science/paper/JGVUTASI

@misc{pith2026250615562,
  author       = {Pith},
  title        = {Pith review of: Automated MRI Tumor Segmentation using hybrid U-Net with Transformer and Efficient Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JGVUTASI}},
  note         = {Machine review of arXiv:2506.15562}
}
read the original abstract

Cancer is an abnormal growth with potential to invade locally and metastasize to distant organs. Accurate auto-segmentation of the tumor and surrounding normal tissues is required for radiotherapy treatment plan optimization. Recent AI-based segmentation models are generally trained on large public datasets, which lack the heterogeneity of local patient populations. While these studies advance AI-based medical image segmentation, research on local datasets is necessary to develop and integrate AI tumor segmentation models directly into hospital software for efficient and accurate oncology treatment planning and execution. This study enhances tumor segmentation using computationally efficient hybrid UNet-Transformer models on magnetic resonance imaging (MRI) datasets acquired from a local hospital under strict privacy protection. We developed a robust data pipeline for seamless DICOM extraction and preprocessing, followed by extensive image augmentation to ensure model generalization across diverse clinical settings, resulting in a total dataset of 6080 images for training. Our novel architecture integrates UNet-based convolutional neural networks with a transformer bottleneck and complementary attention modules, including efficient attention, Squeeze-and-Excitation (SE) blocks, Convolutional Block Attention Module (CBAM), and ResNeXt blocks. To accelerate convergence and reduce computational demands, we used a maximum batch size of 8 and initialized the encoder with pretrained ImageNet weights, training the model on dual NVIDIA T4 GPUs via checkpointing to overcome Kaggle's runtime limits. Quantitative evaluation on the local MRI dataset yielded a Dice similarity coefficient of 0.764 and an Intersection over Union (IoU) of 0.736, demonstrating competitive performance despite limited data and underscoring the importance of site-specific model development for clinical deployment.

Figures

Figures reproduced from arXiv: 2506.15562 by the authors.

Figure 1
Figure 1. Data Processing pipeline Given the modest size of the original dataset (˜1,000 images), an extensive data augmentation strategy was implemented to enhance both data diversity and model robustness. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Data Augmentation Pipeline representations of increasingly complex features [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. (left) The Encoder block that downsamples the image and passes it through a bottleneck [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Overview of the proposed architecture. The Encoder (blue) uses a ResNet50-based U [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Training performance metrics over epochs. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 23 canonical work pages

  1. [1]

    Global cancer observatory: cancer today

    Jacques Ferlay et al. “Global cancer observatory: cancer today”. In: Lyon: International agency for research on cancer20182020 (2020)

    work page 2020

  2. [2]

    MR imaging of neoplastic central nervous system lesions: review and recom- mendations for current practice

    M. Essig et al. “MR imaging of neoplastic central nervous system lesions: review and recom- mendations for current practice”. In: AJNR. American Journal of Neuroradiology33.5 (2012), pp. 803–817. doi: 10.3174/ajnr.A2640. url: https://doi.org/10.3174/ajnr.A2640

    work page doi:10.3174/ajnr.a2640 2012

  3. [3]

    Classification of subtype of acute ischemic stroke: Definitions for use in a multicenter clinical trial. TOAST. Trial of Org 10172 in Acute Stroke Treatment

    Jr. Adams H. P. et al. “Classification of subtype of acute ischemic stroke: Definitions for use in a multicenter clinical trial. TOAST. Trial of Org 10172 in Acute Stroke Treatment”. In: Stroke 24.1 (1993), pp. 35–41. doi: 10.1161/01.STR.24.1.35 . url: https://doi.org/10. 1161/01.STR.24.1.35. 13

    work page doi:10.1161/01.str.24.1.35 1993

  4. [4]

    Simultaneous Truth and Performance Level Estimation (STAPLE): an algorithm for the validation of image segmentation

    Simon K. Warfield, Kelly H. Zou, and William M. Wells. “Simultaneous Truth and Performance Level Estimation (STAPLE): an algorithm for the validation of image segmentation”. In:IEEE Transactions on Medical Imaging23.7 (2004), pp. 903–921. doi: 10.1109/TMI.2004.828354

    arXiv 2004

  5. [5]

    A survey on deep learning in medical image analysis

    Geert Litjens et al. “A survey on deep learning in medical image analysis”. In: Medical Image Analysis 42 (2017), pp. 60–88. doi: 10.1016/j.media.2017.07.005

    doi:10.1016/j.media.2017.07.005 2017

  6. [6]

    The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS)

    Bjoern H Menze et al. “The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS)”. In: IEEE Transactions on Medical Imaging34.10 (2015), pp. 1993–2024

    work page 2015

  7. [7]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. “U-net: Convolutional networks for biomedical image segmentation”. In:Medical image computing and computer-assisted intervention– MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceed- ings, part III 18. Springer. 2015, pp. 234–241

    work page 2015

  8. [8]

    A survey of the self supervised learning mechanisms for vision trans- formers

    Asifullah Khan et al. “A survey of the self supervised learning mechanisms for vision trans- formers”. In: arXiv preprint arXiv:2408.17059(2024)

    arXiv 2024

Show all 34 references
  1. [9]

    A recent survey of vision transformers for medical image segmentation

    Asifullah Khan et al. “A recent survey of vision transformers for medical image segmentation”. In: arXiv preprint arXiv:2312.00634(2023)

    2023 arXiv

  2. [10]

    Domain Adaptation for Medical Image Analysis: A Survey

    H. Guan and M. Liu. “Domain Adaptation for Medical Image Analysis: A Survey”. In: IEEE Transactions on Biomedical Engineering69.3 (Mar. 2022). Epub 2022 Feb 18. PMID: 34606445; PMCID: PMC9011180, pp. 1173–1185. doi: 10.1109/TBME.2021.3117407

    2022

  3. [11]

    Official Journal of the European Union, L 119, 4 May 2016, pp

    Regulation (EU) 2016/679 of the European Parliament (General Data Protection Regulation). Official Journal of the European Union, L 119, 4 May 2016, pp. 1–88. http://data.europa. eu/eli/reg/2016/679/oj. 2016

    2016

  4. [12]

    MaxViT-UNet: Multi-axis attention for medical image segmentation

    Abdul Rehman Khan and Asifullah Khan. “MaxViT-UNet: Multi-axis attention for medical image segmentation”. In: arXiv preprint arXiv:2305.08396(2023)

    2023 arXiv

  5. [13]

    nnU-Net: A Self-Configuring Method for Deep Learning-Based Biomed- ical Image Segmentation

    Fabian Isensee et al. “nnU-Net: A Self-Configuring Method for Deep Learning-Based Biomed- ical Image Segmentation”. In: Nature Methods 18.2 (2021), pp. 203–211

    2021

  6. [14]

    UNETR: Transformers for 3D Medical Image Segmentation

    Amir Hatamizadeh et al. “UNETR: Transformers for 3D Medical Image Segmentation”. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2022, pp. 9312–9321

    2022

  7. [15]

    The Liver Tumor Segmentation Benchmark (LiTS)

    Patrick Bilic et al. “The Liver Tumor Segmentation Benchmark (LiTS)”. In: arXiv preprint arXiv:1901.04056 (2019)

    2019 arXiv

  8. [16]

    H-DenseUNet: Hybrid Densely Connected UNet for Liver and Tumor Segmentation From CT Volumes

    Xiaomeng Li et al. “H-DenseUNet: Hybrid Densely Connected UNet for Liver and Tumor Segmentation From CT Volumes”. In: IEEE Transactions on Medical Imaging37.12 (2018), pp. 2663–2674. doi: 10.1109/TMI.2018.2845918

    2018

  9. [17]

    pydicom: An Open Source DICOM Library

    pydicom Contributors. pydicom: An Open Source DICOM Library. https://pydicom.github. io. Accessed: Feb. 2025. 2025

    2025

  10. [18]

    A Survey on Image Data Augmentation for Deep Learning

    Connor Shorten and Taghi M. Khoshgoftaar. “A Survey on Image Data Augmentation for Deep Learning”. In: Journal of Big Data6.1 (2019), p. 60. 14

    2019

  11. [19]

    Attention is all you need

    A Vaswani. “Attention is all you need”. In: Advances in Neural Information Processing Systems (2017)

    2017

  12. [20]

    A Stacked Multi-Connection Simple Reducing Net for Brain Tumor Segmen- tation

    Yi Ding et al. “A Stacked Multi-Connection Simple Reducing Net for Brain Tumor Segmen- tation”. In: IEEE AccessPP (July 2019), pp. 1–1. doi: 10.1109/ACCESS.2019.2926448

    2019

  13. [21]

    Transunet: Transformers make strong encoders for medical image seg- mentation

    Jieneng Chen et al. “Transunet: Transformers make strong encoders for medical image seg- mentation”. In: arXiv preprint arXiv:2102.04306(2021)

    2021 arXiv

  14. [22]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy et al. “An image is worth 16x16 words: Transformers for image recognition at scale”. In: arXiv preprint arXiv:2010.11929(2020)

    2020 arXiv

  15. [23]

    HUT: Hybrid UNet transformer for brain lesion and tumour segmentation

    Wei Kwek Soh, Hing Yee Yuen, and Jagath C Rajapakse. “HUT: Hybrid UNet transformer for brain lesion and tumour segmentation”. In: Heliyon 9.12 (2023)

    2023

  16. [24]

    HTTU-Net: Hybrid Two Track U-Net for Automatic Brain Tumor Segmentation

    Nagwa M. Aboelenein et al. “HTTU-Net: Hybrid Two Track U-Net for Automatic Brain Tumor Segmentation”. In: IEEE Access 8 (2020), pp. 101406–101415. doi: 10.1109/ACCESS.2020. 2998601

    2020 doi

  17. [25]

    Deep residual learning for image recognition

    Kaiming He et al. “Deep residual learning for image recognition”. In: Proceedings of the IEEE conference on computer vision and pattern recognition. 2016, pp. 770–778

    2016

  18. [26]

    Segmentation Models

    Pavel Yakubovskiy. Segmentation Models. https : / / github . com / qubvel / segmentation _ models. Accessed: 2025-06-09. 2019

    2025

  19. [27]

    Imagenet: A large-scale hierarchical image database

    Jia Deng et al. “Imagenet: A large-scale hierarchical image database”. In: 2009 IEEE confer- ence on computer vision and pattern recognition. Ieee. 2009, pp. 248–255

    2009

  20. [28]

    Squeeze-and-excitation networks

    Jie Hu, Li Shen, and Gang Sun. “Squeeze-and-excitation networks”. In: Proceedings of the IEEE conference on computer vision and pattern recognition. 2018, pp. 7132–7141

    2018

  21. [29]

    Cbam: Convolutional block attention module

    Sanghyun Woo et al. “Cbam: Convolutional block attention module”. In: Proceedings of the European conference on computer vision (ECCV). 2018, pp. 3–19

    2018

  22. [30]

    Efficient attention: Attention with linear complexities

    Zhuoran Shen et al. “Efficient attention: Attention with linear complexities”. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. 2021, pp. 3531–3539

    2021

  23. [31]

    Aggregated residual transformations for deep neural networks

    Saining Xie et al. “Aggregated residual transformations for deep neural networks”. In: Pro- ceedings of the IEEE conference on computer vision and pattern recognition. 2017, pp. 1492– 1500

    2017

  24. [32]

    V-Net: Fully Convolutional Neu- ral Networks for Volumetric Medical Image Segmentation

    Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. “V-Net: Fully Convolutional Neu- ral Networks for Volumetric Medical Image Segmentation”. In: 2016 Fourth International Conference on 3D Vision (3DV). 2016, pp. 565–571. doi: 10.1109/3DV.2016.79

    2016 doi

  25. [33]

    Unet++: A nested u-net architecture for medical image segmentation

    Zongwei Zhou et al. “Unet++: A nested u-net architecture for medical image segmentation”. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML- CDS 2018, He...

    2018

  26. [34]

    MM-BiFPN: multi-modality fu- sion network with Bi-FPN for MRI brain tumor segmentation

    Nur Suriza Syazwany, Ju-Hyeon Nam, and Sang-Chul Lee. “MM-BiFPN: multi-modality fu- sion network with Bi-FPN for MRI brain tumor segmentation”. In: IEEE Access 9 (2021), pp. 160708–160720. 16

    2021

Pith tools

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