Pith. sign in

REVIEW 4 major objections 4 minor 33 references

Dual-attention ResNet outperforms transformers in HER2 prediction on DCE-MRI

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A dual-attention ResNet predicts HER2 status from DCE-MRI more accurately than transformer models under matched preprocessing.

desk verdict Useful normalization benchmark buried under a shaky architecture comparison: the THDA-versus-transformer claim does not survive matched evaluation. read the letter →

arxiv 2510.13897 v1 pith:5WMXNNSQ submitted 2025-10-14 q-bio.QM cs.AI

Naomi Fridman , Anat Goldstein This is my paper · ORCID
classification q-bio.QMcs.AI
keywords HER2predictionDCE-MRIdual-attentionResNetintensitynormalizationvisiontransformerbreastcancerexternalvalidationmedicalimageclassification
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a convolutional architecture with dual attention—spatial gating on features plus channel weighting over three contrast phases—predicts HER2 status from DCE-MRI more accurately than transformer-based models, as long as intensity preprocessing is chosen deliberately. The winning recipe is per-channel upper clipping at the 98th percentile with no bias-field correction; N4 correction slightly hurt performance. The model reaches 0.75 accuracy and 0.74 AUC on a held-out multicenter test cohort and 0.66 AUC on an independent external set without fine-tuning. The authors themselves note that part of the gap over transformers may lie in how temporal information is presented, not just in the architecture. A sympathetic reader should care because the result suggests that for small medical imaging datasets, preprocessing and data presentation may matter as much as model choice.

What carries the argument

The central mechanism is the Triple-Head Dual-Attention ResNet: three weight-shared ResNet34 backbones, one per temporal phase, with multi-scale spatial attention modules on the Layer3 and Layer4 outputs that softly gate feature maps (including an edge-suppression step), followed by an adaptive channel-attention layer that weights and fuses the temporal branches. The paired preprocessing identity is per-channel 98th-percentile upper clipping, which maps 12–16 bit intensities into the 0–255 range while preserving within-phase contrast.

What would settle it

Run all three architectures with the same training protocol and test every slice-aggregation function (mean, median, minimum, percentiles) at a shared decision threshold; if the ResNet's AUC advantage over ViT and CvT disappears or reverses for any aggregation, the claimed architectural superiority is an artifact of unmatched evaluation choices.

Watch

Extended reading notes

Core claim

Under matched preprocessing and training protocols, the paper's Triple-Head Dual-Attention ResNet processes each DCE-MRI temporal phase (pre-contrast, early post-contrast, late post-contrast) through a weight-shared ResNet34 backbone, applies multi-scale spatial attention to gate salient regions, then uses channel attention to fuse the phases before classification. This model reports 0.75 accuracy and 0.74 AUC for HER2 status on the multicenter test cohort, ahead of Vision Transformer (AUC 0.64–0.67) and Convolutional Vision Transformer (AUC 0.61–0.63) under the same preprocessing and training settings; on an external 43-lesion cohort, without fine-tuning, it reaches 0.66 AUC. The paper also

Load-bearing premise

The load-bearing premise is that the per-model evaluation choices—median slice aggregation for the ResNet, quantile for ViT, minimum for CvT, and the 0.7 decision threshold—are interchangeable, so the reported gap reflects architecture rather than evaluation design.

Editorial extensions

If this is right

  • If the claim holds, transformer architectures are not inherently superior for DCE-MRI HER2 prediction at this dataset scale; the way temporal phases are encoded into the input may be the decisive factor.
  • Per-channel 98th-percentile upper clipping is a cheap, reproducible normalization that can be adopted immediately, and N4 bias correction can be dropped without sacrificing accuracy.
  • A decision threshold of 0.7 (rather than 0.5) shifts the model toward high specificity (0.86) with modest sensitivity (0.41), a trade-off that could be tuned for clinical screening contexts.
  • The external-validation result, while lower (AUC 0.61–0.66), suggests the learned features transfer across imaging protocols well enough to be worth testing in prospective studies.

Reading between the lines

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

  • Going beyond the paper's claims: because the ResNet keeps temporal phases separate until late fusion while the transformers receive RGB-fused images as a single input, a fairer transformer baseline would tokenize each temporal phase separately; if the gap then narrows, the paper's headline result is about input representation, not attention vs convolution.
  • Also editorial: the comparison uses different slice-level aggregation functions per model (median, quantile, minimum), so an independent reader should rerun all models on a shared aggregation grid; if the ordering changes, the architectural claim is an artifact of evaluation design rather than a property of the models.
  • A further extension: since per-channel clipping preserved within-phase contrast and bias correction hurt, the model likely relies on fine intensity patterns; one could test this by perturbing intensities within the clipped range and measuring sensitivity.
  • The paper's own hypothesis that transformers underperform partly because of dataset size could be tested by training the same transformer on increasing training-set sizes to see if its AUC catches up; that would separate data-hunger from representation.
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 / 4 minor

Summary. The paper presents a Triple-Head Dual-Attention ResNet (THDA-ResNet) for HER2 status prediction from DCE-MRI, benchmarks seven intensity-normalization strategies, and compares the proposed architecture against Vision Transformer (ViT) and Convolutional Vision Transformer (CvT) baselines. The model is trained on 885 I-SPY cases, tuned on 132 validation cases, and tested on 132 I-SPY test cases, with additional external validation on the BreastDCEDL_AMBL dataset (43 lesions). The authors report that THDA-ResNet with channel-wise upper clipping at q0.98 achieves 0.75 accuracy and 0.74 AUC on the I-SPY test set at a decision threshold of 0.7, and AUC 0.61–0.66 on the external set without fine-tuning. The paper's central claim is that THDA-ResNet outperforms transformer-based architectures 'under matched preprocessing and training protocols.' However, the evaluation protocol does not fully support this claim: slice-level aggregation differs across architectures, the optimal threshold is introduced after inspecting threshold-0.5 results, no confidence intervals are reported, and several tables contain internal inconsistencies.

Significance. If the central claim were established, the paper would provide a useful benchmark for DCE-MRI preprocessing and architecture selection for HER2 prediction from a multicenter cohort, particularly the finding that N4 bias-field correction degrades deep-learning performance. The systematic comparison of normalization strategies, use of a public multicenter dataset, and external validation without fine-tuning are positive features. However, the paper's headline 'outperforms transformers' is not supported by the evidence as presented because the comparison is not matched: different slice-aggregation functions are used for different architectures, and the THDA accuracy is reported at a threshold selected after examining test-set results. These issues, combined with missing confidence intervals and row-count inconsistencies in Table 3, mean the quantitative claims should be treated as provisional.

major comments (4)
  1. [Section 2.5, Table 2] The 'matched protocols' claim is contradicted by the slice-aggregation choices. Section 2.5 states that several aggregation methods were tested and that 'the median gave the best AUC,' yet Table 2 reports THDA with median aggregation, ViT with quantile aggregation, and CvT with min aggregation. If each model's aggregation was selected after inspecting test predictions, the reported AUCs are optimistically selected estimates and the architectures are not compared under a common protocol. Please either use one pre-specified aggregation for all models, or justify per-model aggregation with an independent selection procedure (e.g., validation set), and report the sensitivity of the comparison to the aggregation choice.
  2. [Section 3.2, Tables 2 and 3] The decision threshold of 0.7 is introduced after the threshold-0.5 results are presented. Section 3.2 states that the higher threshold 'improved the specificity-sensitivity balance' on the complete test cohort. This makes the THDA accuracy of 0.75 at threshold 0.7 a test-selected estimate, while the transformer accuracies reported in Section 3.1 are at threshold 0.5. Since accuracy is threshold-dependent, this does not support the claim that THDA 'outperforms' transformers. Please report threshold-independent metrics (e.g., AUC with confidence intervals) as the primary comparison, or preselect the threshold on the validation set and report full operating points for all models.
  3. [Table 3] Table 3 contains internally inconsistent row counts. For the 'Complete Test Cohort - 132 cases,' the Channel upper clipping (q0.95) row sums to 99 (TP=8, TN=65, FP=12, FN=14), not 132, while the other two rows in the same block sum to 132. For the 'ISPY2 Dataset Test cohort - 99 cases,' all three rows sum to 103, not 99. These discrepancies suggest either a labeling error or a data-handling error and must be corrected and reconciled with the cohort definitions.
  4. [Section 3.3, Table 4] The external validation set is small (43 lesions), yet no confidence intervals or uncertainty measures are provided for the reported AUC values. The differences between preprocessing methods on this set (AUC 0.60–0.66) are well within typical sampling error for n=43, so the claim that 'global min-max normalization yielded the highest accuracy and AUC' is not supported without intervals. Additionally, Section 2.6 refers to 42 patients while Section 3.3 refers to 43 lesions; please clarify the cohort definition. Reporting bootstrap or DeLong confidence intervals would also help calibrate the generalizability claim.
minor comments (4)
  1. [Abstract/Table 1] Table 1's title says '(n=150)' but the column totals sum to 1,149. This appears to be a typo and should be corrected.
  2. [Throughout] Several typographical errors: 'classification.classification' (last line of Section 1), 'maintained and maintained' (end of Section 1), 'Layer4 outputs.Layer4 outputs' (Section 2.4.1), and 'ISPY2' used inconsistently with 'I-SPY 2'. Additionally, Table 3's expansion 'Threshold-based Deep Hashing Association Network' does not match the paper's 'Triple-Head Dual-Attention ResNet' definition and should be corrected.
  3. [Section 5/References] The ethics statement mentions the Duke Breast Cancer MRI dataset as included, but Section 2.1 and the external validation only mention I-SPY and AMBL. Please clarify which datasets are actually used. Also consider stating whether the 'external' AMBL dataset (ref [31]) was curated by the same authors, since this bears on the interpretation of 'external' validation.
  4. [Section 2.4.2] The paper says transformers were fine-tuned using 'identical preprocessing and training protocols as THDA-ResNet,' but Section 2.5 then assigns different slice-aggregation functions to each model. The apparent contradiction should be resolved by specifying the protocol before model-specific aggregation choices are made.

Circularity Check

3 steps flagged · score 4.0 of 10

Threshold and aggregation are selected on the test cohort, so the headline accuracy advantage is partly an evaluation artifact; the external benchmark is a self-cited dataset.

  1. fitted input called prediction [Section 3.2 (Table 3)]
    "This higher threshold improved the specificity-sensitivity balance compared to threshold 0.5. The three best preprocessing methods—channel upper clipping q0.98, channel upper clipping q0.95, and global upper clipping q0.99—achieved accuracy of 0.74–0.75 with AUCs of 0.72–0.74 on the complete test cohort."

    The decision threshold 0.7 is selected after inspecting the complete test cohort, and accuracy is threshold-dependent. Reporting THDA accuracy at this test-selected threshold while transformer accuracy is reported at threshold 0.5 (Table 2) makes the accuracy comparison an artifact of post-hoc threshold choice rather than a matched architectural comparison. The reported 0.75 accuracy is a fitted value, not a pre-specified prediction.

  2. fitted input called prediction [Section 2.5 (Table 2)]
    "We tested several aggregation methods to combine slice-level predictions into a patient-level score: mean, median, minimum, 10th (q0.1) and 90th (q0.9) percentiles, and the proportion of HER2-positive predictions. The median gave the best AUC, while the 90th percentile provided the highest accuracy."

    The patient-level aggregation is itself a free parameter. THDA is reported with median aggregation, ViT with quantile aggregation, and CvT with min aggregation (Table 2). If each aggregation was chosen because it gave the best result on the same test set, then the reported AUCs (including THDA 0.74 vs ViT 0.64–0.67 and CvT 0.61–0.63) are optimistically selected estimates, not the outcome of a single matched protocol. This weakens the claim of 'matched preprocessing and training protocols' underlying the architecture comparison.

1 more flagged steps
  1. self citation load bearing [Section 2.6 (reference [31])]
    "we performed external validation on the public BreastDCEDL_AMBL dataset, which provides segmented tumor masks and HER2 status annotations for 42 patients [31]."

    Reference [31] is authored by the same two authors of this paper (Naomi Fridman and Anat Goldstein, 'Transformer classification of breast lesions: The breastdcedl_ambl benchmark dataset and 0.92 auc baseline, 2025'). The cross-institutional generalizability claim (AUC 0.61–0.66) is therefore evaluated on a dataset whose curation, labels, and benchmark definition come from the authors' own prior work. While the images are from a different institution, the benchmark is not independent of the present authors, making the 'external validation' partially self-referential.

full rationale

This is an empirical benchmarking paper with no analytic derivation, so there is no equation-level EqX=EqY circularity. The central numerical claims are nevertheless partly self-referential. The decision threshold (0.7) and per-model slice-aggregation rules are selected on the same test cohort used to report the headline accuracy, so the THDA-vs-transformer accuracy comparison is not a matched, pre-specified evaluation. The externally validating dataset is cited to the authors' own prior benchmark paper, weakening the 'independent' label. These issues inflate or confound the reported metrics but do not make the AUC computation itself equivalent to its inputs: the AUC comparison (0.74 vs 0.61–0.67) is threshold-independent and remains an empirical result, and the model does learn from training labels rather than from the test labels directly. Therefore the overall circularity score is 4, not 6+.

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

The central claim is empirical, not derivational, so the ledger captures the hand-chosen evaluation parameters and domain assumptions the result depends on. The main cost is the non-matched aggregation/threshold choices and the reliance on self-curated benchmarks.

free parameters (4)
  • Decision threshold = 0.7
    Chosen after threshold 0.5 results to improve specificity; no independent threshold-selection protocol is reported (Section 3.2, Table 3).
  • Slice-aggregation statistics per model = THDA: median; ViT: quantile; CvT: min
    Non-matched aggregation choices in Table 2; 'median gave best AUC' was determined for THDA, and the same grid was not run for the transformers (Section 2.5).
  • Upper-clipping percentile = q0.98 for best I-SPY config; global min-max best on AMBL
    Selected from seven normalization strategies based on performance; the optimal strategy differs between I-SPY and external data (Sections 2.3, 3.2, 3.3).
  • Edge suppression factor = 0.30
    Hand-set deweighting of border pixels in the attention module; no ablation is provided (Section 2.4.1).
assumptions (4)
  • domain assumption HER2 status is learnable from three selected DCE-MRI phases fused as RGB (pre-contrast, early post-contrast, late post-contrast).
    Section 2.4 follows phase selection from the authors' prior BreastDCEDL work; no ablation of phase timing is reported.
  • domain assumption Patient-level aggregation of slice-level predictions preserves patient-level ranking.
    Section 2.5 aggregates slice scores with mean/median/minimum; within-patient correlation and number of slices per patient are not modeled or reported.
  • domain assumption ImageNet-pretrained weights transfer useful features to 8-bit DCE-MRI despite strong domain shift.
    All models, including THDA-ResNet, rely on pretrained backbones (Sections 2.4.1 and 2.4.2).
  • domain assumption HER2 labels in I-SPY and AMBL are accurate ground truth with negligible label noise.
    The study uses publicly available labels; no pathology re-review or label-quality audit is described (Section 2.6, Ethics).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dual-attention ResNet outperforms transformers in HER2 prediction on DCE-MRI." pith.science (2026). https://pith.science/paper/5WMXNNSQ

@misc{pith2026251013897,
  author       = {Pith},
  title        = {Pith review of: Dual-attention ResNet outperforms transformers in HER2 prediction on DCE-MRI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5WMXNNSQ}},
  note         = {Machine review of arXiv:2510.13897}
}
read the original abstract

Breast cancer is the most diagnosed cancer in women, with HER2 status critically guiding treatment decisions. Noninvasive prediction of HER2 status from dynamic contrast-enhanced MRI (DCE-MRI) could streamline diagnostics and reduce reliance on biopsy. However, preprocessing high-dynamic-range DCE-MRI into standardized 8-bit RGB format for pretrained neural networks is nontrivial, and normalization strategy significantly affects model performance. We benchmarked intensity normalization strategies using a Triple-Head Dual-Attention ResNet that processes RGB-fused temporal sequences from three DCE phases. Trained on a multicenter cohort (n=1,149) from the I-SPY trials and externally validated on BreastDCEDL_AMBL (n=43 lesions), our model outperformed transformer-based architectures, achieving 0.75 accuracy and 0.74 AUC on I-SPY test data. N4 bias field correction slightly degraded performance. Without fine-tuning, external validation yielded 0.66 AUC, demonstrating cross-institutional generalizability. These findings highlight the effectiveness of dual-attention mechanisms in capturing transferable spatiotemporal features for HER2 stratification, advancing reproducible deep learning biomarkers in breast cancer imaging.

Figures

Figures reproduced from arXiv: 2510.13897 by the authors.

Figure 1
Figure 1. Intensity normalization strategies for DCE-MRI (dynamic contrast-enhanced magnetic resonance imaging) [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Triple-Head Dual-Attention ResNet architecture for DCE-MRI analysis. (a) Overall architecture showing [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Triple-Head Dual-Attention ResNet mechanism visualization for DCE-MRI (dynamic contrast-enhanced [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 2 linked inside Pith

  1. [1]

    Turnbull

    Lindsay W. Turnbull. Dynamic contrast-enhanced mri in the diagnosis and management of breast cancer.NMR in Biomedicine, 22(1):28–39, 2009

    2009

  2. [2]

    Current status of breast mr imaging part 2

    Christiane K Kuhl. Current status of breast mr imaging part 2. clinical applications.Radiology, 244(3):672–691, 2007

    2007

  3. [3]

    F. M. Blows et al. Subtyping of breast cancer by immunohistochemistry to investigate a relationship between subtype and short and long term survival: A collaborative analysis of data for 10,159 cases from 12 studies.PLoS Medicine, 7(5):e1000279, 2010

    2010

  4. [4]

    Weigelt, F

    B. Weigelt, F. L. Baehner, and J. S. Reis-Filho. The contribution of gene expression profiling to breast cancer classification, prognostication and prediction: a retrospective of the last decade.Journal of Pathology, 220(2):263– 280, 2010

    2010

  5. [5]

    M. C. U. Cheang et al. Ki67 index, her2 status, and prognosis of patients with luminal b breast cancer.JNCI: Journal of the National Cancer Institute, 101(10):736–750, 2009

    2009

  6. [6]

    C. S. Vallejos et al. Breast cancer classification according to immunohistochemistry markers: Subtypes and association with clinicopathologic variables in a peruvian hospital database.Clinical Breast Cancer, 10(4):294– 300, 2010

    2010

  7. [7]

    Use of chemotherapy plus a monoclonal antibody against her2 for metastatic breast cancer that overexpresses her2.New England Journal of Medicine, 344(11):783–792, 2001

    Dennis J Slamon, Brenda Leyland-Jones, Sidney Shak, Hermann Fuchs, Virginia Paton, Tejada Bajamonde, Thomas Fleming, Wolfgang Eiermann, Joerg Wolter, Mark Pegram, Jose Baselga, and Larry Norton. Use of chemotherapy plus a monoclonal antibody against her2 for metastatic breast cancer that overexpresses her2.New England Journal of Medicine, 344(11):783–792, 2001

    2001

  8. [8]

    Antonio C Wolff, M Elizabeth H Hammond, Kimberly H Allison, Brian E Harvey, Pamela B Mangu, John M S Bartlett, Michael Bilous, Ian O Ellis, Noreen U Lin, Michael F Press, et al. Human epidermal growth factor receptor 2 testing in breast cancer: American society of clinical oncology/college of american pathologists clinical practice guideline focused updat...

    2018

Show all 33 references
  1. [9]

    Luo, J.-L

    H.-J. Luo, J.-L. Ren, L. mei Guo, J. liang Niu, and X.-L. Song. Mri-based machine learning radiomics for prediction of her2 expression status in breast invasive ductal carcinoma.European Journal of Radiology Open, 13:100592, 2024

    2024

  2. [10]

    Fan et al

    Y . Fan et al. Deep learning predicts her2 status in invasive breast cancer from multimodal ultrasound and mri. Biomolecules and Biomedicine, 2025

    2025

  3. [11]

    V . N. Thatha et al. Histopathological image based breast cancer diagnosis using deep learning and bio inspired optimization.Scientific Reports, 15(1):19034, 2025

    2025

  4. [12]

    Behar and M

    N. Behar and M. Shrivastava. Resnet50-based effective model for breast cancer classification using histopathology images.Computational Modeling in Engineering & Sciences, 130(2):823–839, 2021

    2021

  5. [13]

    Dosovitskiy et al

    A. Dosovitskiy et al. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv, 2020

    2020

  6. [14]

    Fridman, B

    N. Fridman, B. Solway, T. Fridman, I. Barnea, and A. Goldshtein. Breastdcedl: Curating a comprehensive dce-mri dataset and developing a transformer implementation for breast cancer treatment response prediction.arXiv, 2025

    2025

  7. [15]

    Wang et al

    F. Wang et al. Residual attention network for image classification. In2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6450–6458, 2017

    2017

  8. [16]

    X. Wei, M. Z. Hossain, and K. A. Ahmed. A resnet attention model for classifying mosquitoes from wing-beating sounds.Scientific Reports, 12(1):10334, 2022

    2022

  9. [17]

    S. Liu. Breast cancer classification with enhanced interpretability: Dalaresnet50 and dt grad-cam.arXiv, 2023

    2023

  10. [18]

    J. Hu, W. Huang, H. Zhang, et al. Brain image denoising using dual-channel attentional residual network.Critical Reviews in Biomedical Engineering, 52(5):17–27, 2024

    2024

  11. [19]

    Zhu and Y

    B. Zhu and Y . Yang. Quality assessment of abdominal ct images: an improved resnet algorithm with dual-attention mechanism.American Journal of Translational Research, 16(7):3099–3107, 2024. 12 DUAL-ATTENTION RESNET OUTPERFORMS TRANSFORMERS IN HER2 PREDICTION ON DCE-MRI

    2024

  12. [20]

    Xu, Y .-L

    W. Xu, Y .-L. Fu, and D. Zhu. Resnet and its application to medical image processing: Research progress and challenges.Computer Methods and Programs in Biomedicine, 240:107660, 2023

    2023

  13. [21]

    Cheng et al

    J. Cheng et al. Resganet: Residual group attention network for medical image classification and segmentation. Medical Image Analysis, 76:102313, 2022

    2022

  14. [22]

    Breastdcedl: A deep learning–ready breast dce-mri dataset, 2025

    Naomi Fridman, Bubby Solway, Tomer Fridman, Itamar Barnea, and Anat Goldstein. Breastdcedl: A deep learning–ready breast dce-mri dataset, 2025. Resource type: Dataset. Language: English

    2025

  15. [23]

    Newitt, N

    D. Newitt, N. Hylton, on behalf of the I-SPY 1, and ACRIN 6657 Trial Team. Breast dce-mri data and segmentations from patients in the i-spy 1/acrin 6657 trials. ISPY1 - The Cancer Imaging Archive (TCIA), 2016. Accessed: Aug. 09, 2024

    2016

  16. [24]

    L. W. et al. I-spy 2 breast dynamic contrast enhanced mri trial (ispy2) (version 1) [data set]. The Cancer Imaging Archive (TCIA), 2022

    2022

  17. [25]

    Wang and D

    H. Wang and D. Yee. I-spy 2: a neoadjuvant adaptive clinical trial designed to improve outcomes in high-risk breast cancer.Current Breast Cancer Reports, 11(4):303–310, 2019

    2019

  18. [26]

    Quantitative mri analysis: developing software applications for the healthcare environment.Magnetic Resonance Imaging, 25(8):1245–1260, 2007

    Olga Kubassova, Mikkel Boesen, and Roger D Boyle. Quantitative mri analysis: developing software applications for the healthcare environment.Magnetic Resonance Imaging, 25(8):1245–1260, 2007

    2007

  19. [27]

    N4itk: improved n3 bias correction.IEEE transactions on medical imaging, 29(6):1310–1320, 2010

    Nicholas J Tustison, Brian B Avants, Philip A Cook, Yuanjie Zheng, Alexander Egan, Paul A Yushkevich, and James C Gee. N4itk: improved n3 bias correction.IEEE transactions on medical imaging, 29(6):1310–1320, 2010

    2010

  20. [28]

    Cvt: Introducing convolutions to vision transformers.arXiv preprint arXiv:2103.15808, 2021

    Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers.arXiv preprint arXiv:2103.15808, 2021

    2021 arXiv

  21. [29]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at...

    2010 arXiv

  22. [30]

    Delving deep into rectifiers: Surpassing human-level performance on imagenet classification

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. InProceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1026–1034, 2015

    2015

  23. [31]

    Transformer classification of breast lesions: The breastdcedl_ambl benchmark dataset and 0.92 auc baseline, 2025

    Naomi Fridman and Anat Goldstein. Transformer classification of breast lesions: The breastdcedl_ambl benchmark dataset and 0.92 auc baseline, 2025

    2025

  24. [32]

    Lirong Song, Chunli Li, and Jiandong Yin. Texture analysis using semiquantitative kinetic parameter maps from dce-mri: Preoperative prediction of her2 status in breast cancer.BioMed Research International, 2021:6693437, 2021

    2021

  25. [33]

    W. Li, D. C. Newitt, J. Gibbs, L. J. Wilmes, E. F. Jones, V . A. Arasu, F. Strand, N. Onishi, A. A.-T. Nguyen, J. Kornak, B. N. Joe, E. R. Price, H. Ojeda-Fournier, M. Eghtedari, K. W. Zamora, S. A. Woodard, H. Umphrey, W. Bernreuter, M. Nelson, and N. M. Hylton. Predicting br...

    2020

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.