Pith. sign in

REVIEW 3 major objections 2 minor 68 references

A 3D Gaussian Splatting pipeline can skip Structure-from-Motion entirely and still synthesize new views from as few as two photos, gaining 2.75 dB PSNR over other 3DGS methods.

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 ↗

SfM-free 3D Gaussian Splatting jointly estimates camera poses and geometry from two views and reports a 2.75 dB PSNR improvement over existing 3DGS methods on extremely sparse inputs.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection The abstract describes a plausible 2-view 3DGS advance, but the supplied full text is an unrelated mammography paper, so the central claim is currently unsupported. the 3 major comments →

arxiv 2508.15457 v1 pith:T2Y5N6U3 submitted 2025-08-21 cs.CV

Enhancing Novel View Synthesis from extremely sparse views with SfM-free 3D Gaussian Splatting Framework

Zongqi He , Hanmin Li , Kin-Chung Chan , Yushen Zuo , Hao Xie , Zhe Xiao , Jun Xiao , Kin-Man Lam This is my paper
classification cs.CV
keywords 3D Gaussian Splattingnovel view synthesisSfM-free reconstructionextremely sparse viewscamera pose estimationdense stereo matchingview interpolation supervisionmulti-scale geometry regularization
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.

The reading

This paper targets the weakest link in 3D Gaussian Splatting: its reliance on Structure-from-Motion (SfM) to supply camera poses and an initial point cloud, which breaks down when the input is extremely sparse. The authors claim SfM can be dropped entirely and replaced by a dense stereo module that progressively estimates the camera pose and builds a global dense point cloud that initializes the splats. Around that core they add interpolated intermediate views, generated from the estimated pose trajectory, as extra supervision, plus multi-scale Laplacian and geometry regularizers to keep the geometry stable. Under two-view training they report a 2.75 dB PSNR improvement over other 3DGS-based methods, with noticeably less distortion and preserved high-frequency detail. If correct, this would make realistic novel-view synthesis possible from an ordinary pair of photos, with no SfM preprocessing.

Core claim

3D Gaussian Splatting normally depends on Structure-from-Motion to supply camera poses and an initial point cloud, and SfM degrades when the number of input views is very small. The paper's central claim is that SfM is not a necessary preprocessing stage even in the extreme case of two views: a dense stereo module progressively estimates the relative camera pose and reconstructs a global dense point cloud that directly initializes the splats; a coherent view interpolation module synthesizes viewpoint-consistent intermediate views along the estimated trajectory to pad the sparse training signal; and multi-scale Laplacian and adaptive geometry regularizers hold the geometry together in regions

What carries the argument

The load-bearing component is the dense stereo module that replaces SfM: it takes the two available views, progressively estimates the relative camera pose, and fuses disparity into a global dense point cloud used to initialize the 3D Gaussians. Two supporting mechanisms carry the rest: the coherent view interpolation module, which samples intermediate camera poses along the estimated baseline and renders viewpoint-consistent content as pseudo-supervision, and the two regularizers, a multi-scale Laplacian consistency term and an adaptive spatial-aware multi-scale geometry term, that penalize geometry collapse and floaters where supervision is absent.

Load-bearing premise

The load-bearing premise is that dense stereo matching between two views recovers camera poses and a dense point cloud accurate enough to serve as the 3DGS scaffold; if the stereo result is locally consistent but globally wrong, every downstream regularizer learns from a biased geometry.

What would settle it

Measure the pipeline's pose estimates against known ground-truth camera positions on scenes where stereo correspondence is unreliable (textureless walls, repeating patterns). If the estimated trajectory drifts from ground truth while render PSNR on interpolated viewpoints stays high, the high scores reflect fitting to a globally wrong scaffold rather than true scene geometry; if PSNR collapses on held-out views far from the interpolated arc, the interpolation supervision is masking the SfM failure instead of fixing it.

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

If this is right

  • SfM's failure mode at two views is bypassed rather than patched: pose and initialization come from the same dense stereo pass, so the pipeline no longer waits on a separate, brittle preprocessing step.
  • The interpolated-view supervision is generated entirely from the estimated pose trajectory, so extra training signal does not require ground-truth poses or additional captures.
  • At the reported 2.75 dB PSNR gain in the two-view regime, casual capture (two phone photos of an object) becomes a viable input format for realistic free-viewpoint rendering.
  • The two regularizers suggest that geometric stability, not just photometric appearance, is the limiting factor in sparse-view splatting, so the same regularization design can be reused by later sparse-view radiance-field methods.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable consequence the paper does not report: the stereo-derived pose is only as good as the pixel correspondences, so scenes with low or repetitive texture (blank walls, corridors, foliage) should sharply erode pose accuracy and with it the PSNR advantage over SfM-based baselines.
  • The interpolation supervision inherits any bias in the estimated pose trajectory, so high PSNR on interpolated viewpoints could coexist with globally distorted geometry; comparing the recovered point cloud against ground-truth scans would settle whether the reconstruction is globally correct or only locally consistent.
  • The same SfM-free pose-and-initialization module could plausibly be lifted out of 3DGS into other radiance-field or mesh-based reconstruction pipelines that currently assume SfM inputs, a reuse the paper leaves implicit.
  • An ablation separating the interpolation supervision from the regularizers would reveal which mechanism carries the reported gain when the stereo pose is noisy; the paper presents the components together.
Share X Bluesky LinkedIn Reddit HN

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 / 2 minor

Summary. The abstract describes an SfM-free 3D Gaussian Splatting framework for novel view synthesis from extremely sparse views, with three proposed components: a dense stereo module for pose estimation and global point-cloud initialization, a coherent view interpolation module that generates viewpoint-consistent pseudo-supervision, and multi-scale Laplacian / adaptive spatial-aware geometry regularizers. The abstract reports a 2.75 dB PSNR improvement over state-of-the-art 3DGS methods using only two training views. However, the supplied full text is arXiv:2508.15452v3, the DoSReMC mammography classification paper, which contains no 3DGS content, no dense stereo module, no view interpolation, and no two-view novel-view-synthesis experiments. Consequently, none of the abstract's load-bearing claims are supported by the document as submitted.

Significance. If the abstract's claims are correct, the result would be significant: removing the SfM dependency and achieving a 2.75 dB gain in two-view novel view synthesis would address a practical bottleneck in 3DGS deployment. The proposed design is also plausible and well motivated. However, the manuscript as received provides no derivations, no experimental protocol, no named baselines, no ablations, no code, and no data. It is therefore impossible to verify the central claim or to credit the advertised methodological contributions. The paper currently ships only an abstract whose supporting body is missing.

major comments (3)
  1. [Abstract vs. full text] The supplied full text is the DoSReMC mammography classification paper, arXiv:2508.15452v3, not the 3DGS paper announced by the abstract (arXiv:2508.15457). The body contains no dense stereo module, no coherent view interpolation, no 3DGS, no multi-scale Laplacian or geometry regularization, and no two-view novel-view-synthesis experiments. Every load-bearing premise of the abstract — pose-estimation accuracy, interpolation-supervision consistency, regularizer effectiveness, and the 2.75 dB improvement — therefore lacks supporting derivation, ablation, dataset, or comparison in the document. This is a missing-support finding, not a claim that the abstract's result is false, but as it stands the manuscript cannot be evaluated or published.
  2. [Abstract] The proposed supervision scheme is structurally self-referential. Interpolated views are generated from the training view pairs and then used as 'additional supervision signals,' while the geometry regularizers optimize consistency with the same scaffold. If the intended method trains against its own interpolations, reported gains may partly measure agreement with model-generated geometry rather than with held-out viewpoints. Because the body contains no equations, loss definitions, or validation of the interpolated views (e.g., against held-out poses), this risk cannot be resolved from the submitted document. The full method must specify which losses use real views, which use interpolated views, and how the interpolation module is validated.
  3. [Abstract] The headline quantitative claim, a 2.75 dB PSNR improvement with two training views, is presented without any experimental protocol. No datasets are named, no baseline methods are identified, no standard deviations or per-scene results are reported, and no ablation isolates the contributions of pose initialization, view interpolation, and the two regularizers. Even if the correct full text is recovered, the abstract's single number is not interpretable without these details. The revision should include dataset statistics, baseline configurations, and ablations for each component.
minor comments (2)
  1. [Full text header] The submitted body carries arXiv:2508.15452v3 while the abstract is from arXiv:2508.15457. This ID mismatch should be corrected in any resubmission, and the correct full text should be attached.
  2. [Abstract] The phrase 'other state-of-the-art 3DGS-based approaches' is too vague even for an abstract; at least one concrete named baseline or benchmark family should be indicated so the claimed improvement is falsifiable.

Circularity Check

0 steps flagged

Supplied full text is a different paper (DoSReMC, arXiv:2508.15452v3); no circularity can be established.

full rationale

The provided full text is not the 3DGS paper described in the abstract. The body explicitly carries the header "arXiv:2508.15452v3 [eess.IV] 12 Apr 2026" and the title "DoSReMC: Domain Shift Resilient Mammography Classification using Batch Normalization Adaptation"; its methods, datasets, and experiments concern batch normalization adaptation for mammography, not 3D Gaussian Splatting. The target abstract's derivation chain—dense stereo pose estimation, coherent view interpolation supervision, multi-scale Laplacian/geometry regularization, and the 2.75 dB improvement—appears nowhere in the supplied body. There is therefore no equation, training step, or experiment in the document from which the claimed prediction could reduce to its own inputs. Under the hard rules, an absent derivation is a missing-support problem, not a demonstrable circularity; I cannot quote a specific reduction because none exists in the supplied text. Accordingly, the circularity score is 0.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

Everything here is inferred from the abstract because the supplied full text belongs to a different paper. The method's central premises are two: dense stereo can replace SfM for pose and point-cloud initialization from two views, and interpolated-view content is faithful enough to serve as supervision. No invented entities are introduced; the regularizers are engineering assumptions, not new physical entities.

free parameters (2)
  • Regularization weights for the multi-scale Laplacian and spatial-aware geometry terms
    The abstract introduces two regularizers; their relative weighting is not reported and would be tuned on validation data.
  • Number and density of interpolated views from the coherent view interpolation module
    Interpolating extra camera poses and content adds supervision, but the schedule is unspecified in the abstract and affects the two-view training regime.
axioms (3)
  • domain assumption Dense stereo matching between two sparse views yields camera poses and a global dense point cloud accurate enough to initialize 3DGS
    The abstract states the dense stereo module 'progressively estimates camera pose information and reconstructs a global dense point cloud for initialization'; if this geometry is biased, the entire pipeline inherits the bias.
  • domain assumption Interpolated views are viewpoint-consistent and faithful enough to serve as supervision without ground-truth poses
    The abstract says the coherent view interpolation module 'generates viewpoint-consistent content as additional supervision signals for training'; this assumes consistency with the real scene, otherwise the renderer is trained on its own artifacts.
  • ad hoc to paper Multi-scale Laplacian and adaptive spatial-aware geometry regularizers improve geometry without suppressing high-frequency detail
    These terms are introduced for this method; their benefit over plain 3DGS losses is an unproven design assumption in the abstract.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Novel View Synthesis from extremely sparse views with SfM-free 3D Gaussian Splatting Framework." pith.science (2026). https://pith.science/paper/T2Y5N6U3

@misc{pith2026250815457,
  author       = {Pith},
  title        = {Pith review of: Enhancing Novel View Synthesis from extremely sparse views with SfM-free 3D Gaussian Splatting Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T2Y5N6U3}},
  note         = {Machine review of arXiv:2508.15457}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

3D Gaussian Splatting (3DGS) has demonstrated remarkable real-time performance in novel view synthesis, yet its effectiveness relies heavily on dense multi-view inputs with precisely known camera poses, which are rarely available in real-world scenarios. When input views become extremely sparse, the Structure-from-Motion (SfM) method that 3DGS depends on for initialization fails to accurately reconstruct the 3D geometric structures of scenes, resulting in degraded rendering quality. In this paper, we propose a novel SfM-free 3DGS-based method that jointly estimates camera poses and reconstructs 3D scenes from extremely sparse-view inputs. Specifically, instead of SfM, we propose a dense stereo module to progressively estimates camera pose information and reconstructs a global dense point cloud for initialization. To address the inherent problem of information scarcity in extremely sparse-view settings, we propose a coherent view interpolation module that interpolates camera poses based on training view pairs and generates viewpoint-consistent content as additional supervision signals for training. Furthermore, we introduce multi-scale Laplacian consistent regularization and adaptive spatial-aware multi-scale geometry regularization to enhance the quality of geometrical structures and rendered content. Experiments show that our method significantly outperforms other state-of-the-art 3DGS-based approaches, achieving a remarkable 2.75dB improvement in PSNR under extremely sparse-view conditions (using only 2 training views). The images synthesized by our method exhibit minimal distortion while preserving rich high-frequency details, resulting in superior visual quality compared to existing techniques.

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

68 extracted references · 57 canonical work pages

  1. [1]

    F. Bray, M. Laversanne, H. Sung, J. Ferlay, R. L. Siegel, I. Soerjomataram, A. Jemal, Global can- cer statistics 2022: Globocan estimates of inci- dence and mortality worldwide for 36 cancers in 185 countries, CA: a cancer journal for clinicians (2021).����������������������

    work page 2022

  2. [2]

    R. L. Siegel, A. N. Giaquinto, A. Jemal, Cancer statistics, 2024, CA: a cancer journal for clini- cians 74 (1) (2024) 12–49.����������������� �����

    work page 2024

  3. [3]

    W. Ren, M. Chen, Y . Qiao, F. Zhao, Global guide- lines for breast cancer screening: a systematic review, The Breast 64 (2022) 85–99.������� �������������������������

    work page 2022

  4. [4]

    Katalinic, N

    A. Katalinic, N. Eisemann, K. Kraywinkel, M. R. Noftz, J. Hübner, Breast cancer incidence and mortality before and after implementation of the german mammography screening program, Inter- national journal of cancer 147 (3) (2020) 709–718. ���������������������

    work page 2020

  5. [5]

    S. W. Duffy, L. Tabár, H.-H. Chen, M. Holmqvist, M.-F. Yen, S. Abdsalah, B. Epstein, E. Frodis, E. Ljungberg, C. Hedborg-Melander, et al., The impact of organized mammography service screening on breast carcinoma mortality in seven swedish counties: a collaborative evaluation, Can- cer: Interdisciplinary International Journal of the American Cancer Societ...

    work page 2002

  6. [6]

    Schopper, C

    D. Schopper, C. de Wolf, How effective are breast cancer screening programmes by mammography? review of the current evidence, European journal of cancer 45 (11) (2009) 1916–1923.������� �����������������������

    work page 2009

  7. [7]

    Kalager, M

    M. Kalager, M. Zelen, F. Langmark, H.-O. Adami, Effect of screening mammography on breast- cancer mortality in norway, New England Journal of Medicine 363 (13) (2010) 1203–1210.���� ���������������������

    work page 2010

  8. [8]

    Brodersen, V

    J. Brodersen, V . D. Siersma, Long-term psychoso- cial consequences of false-positive screening mammography, The Annals of Family Medicine 11 (2) (2013) 106–115.���������������� ����

    work page 2013

  9. [9]

    H. D. Nelson, E. S. O’meara, K. Kerlikowske, S. Balch, D. Miglioretti, Factors associated with rates of false-positive and false-negative re- sults from digital mammography screening: an analysis of registry data, Annals of internal medicine 164 (4) (2016) 226–235.������� �������������

    work page 2016

  10. [10]

    Zhang, Y

    X. Zhang, Y . Zhang, E. Y . Han, N. Jacobs, Q. Han, X. Wang, J. Liu, Classification of whole mammo- gram and tomosynthesis images using deep con- volutional neural networks, IEEE transactions on nanobioscience 17 (3) (2018) 237–242.������� ���������������������

    work page 2018

  11. [11]

    Y . Tan, K. Sim, F. F. Ting, Breast cancer de- tection using convolutional neural networks for mammogram imaging system, in: 2017 Interna- tional Conference on Robotics, Automation and Sciences (ICORAS), IEEE, 2017, pp. 1–5.���� ���������������������������

    work page 2017

  12. [12]

    Lu, E.-W

    H.-C. Lu, E.-W. Loh, S.-C. Huang, The classifi- cation of mammogram using convolutional neu- ral network with specific image preprocessing for breast cancer detection, in: 2019 2nd International Conference on Artificial Intelligence and Big Data (ICAIBD), IEEE, 2019, pp. 9–12.������������ �������������������

    work page 2019

  13. [13]

    K. J. Geras, S. Wolfson, Y . Shen, N. Wu, S. Kim, E. Kim, L. Heacock, U. Parikh, L. Moy, K. Cho, High-resolution breast cancer screening with multi-view deep convolutional neural net- works, arXiv preprint arXiv:1703.07047 (2017). �����������������������������

    Pith/arXiv arXiv 2017

  14. [14]

    H. N. Khan, A. R. Shahid, B. Raza, A. H. Dar, H. Alquhayz, Multi-view feature fusion based four views model for mammogram classification us- ing convolutional neural network, IEEE Access 7 (2019) 165724–165733.������������������� ������������

    work page 2019

  15. [15]

    H. Wang, J. Feng, Z. Zhang, H. Su, L. Cui, H. He, L. Liu, Breast mass classification via deeply in- tegrating the contextual information from multi- view data, Pattern Recognition 80 (2018) 42–52. ��������������������������������

    work page 2018

  16. [16]

    Kyono, F

    T. Kyono, F. J. Gilbert, M. van der Schaar, Mammo: A deep learning solution for facilitating radiologist-machine collaboration in breast can- cer diagnosis, arXiv preprint arXiv:1811.02661 (2018).�����������������������������. 22

    Pith/arXiv arXiv 2018

  17. [17]

    Lopez, E

    E. Lopez, E. Grassucci, M. Valleriani, D. Com- miniello, Hypercomplex neural architectures for multi-view breast cancer classification, arXiv preprint arXiv:2204.05798 (2022).������� ����������������������

    arXiv 2022

  18. [18]

    Pathak, J

    S. Pathak, J. Schlötterer, J. Geerdink, J. Veltman, M. van Keulen, N. Strisciuglio, C. Seifert, Breast cancer prediction using mammography exams for real hospital settings, Computers in Biology and Medicine 198 (2025) 111136.�������������� ����������������������

    work page 2025

  19. [19]

    M. G. Ertosun, D. L. Rubin, Probabilistic visual search for masses within mammography images using deep learning, in: 2015 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), IEEE, 2015, pp. 1310–1315.������� ����������������������

    work page 2015

  20. [20]

    W. E. Fathy, A. S. Ghoneim, A deep learning approach for breast cancer mass detection, Inter- national Journal of Advanced Computer Science and Applications 10 (1) (2019).������������� �������������������

    work page 2019

  21. [21]

    A. J. Barnett, F. R. Schwartz, C. Tao, C. Chen, Y . Ren, J. Y . Lo, C. Rudin, A case-based inter- pretable deep learning model for classification of mass lesions in digital mammography, Nature Ma- chine Intelligence 3 (12) (2021) 1061–1070.���� ��������������������������

    work page 2021

  22. [22]

    Zhou, L.-Y

    J. Zhou, L.-Y . Luo, Q. Dou, H. Chen, C. Chen, G.- J. Li, Z.-F. Jiang, P.-A. Heng, Weakly supervised 3d deep learning for breast cancer classification and localization of the lesions in mr images, Jour- nal of Magnetic Resonance Imaging 50 (4) (2019) 1144–1151.����������������������

    work page 2019

  23. [23]

    Agarwal, O

    R. Agarwal, O. Diaz, X. Lladó, M. H. Yap, R. Martí, Automatic mass detection in mammo- grams using deep convolutional neural networks, Journal of Medical Imaging 6 (3) (2019) 031409– 031409.����������������������������

    work page 2019

  24. [24]

    H. Cao, S. Pu, W. Tan, J. Tong, Breast mass detec- tion in digital mammography based on anchor-free architecture, Computer methods and programs in biomedicine 205 (2021) 106033.������������ ������������������

    work page 2021

  25. [25]

    Y . Su, Q. Liu, W. Xie, P. Hu, Yolo-logo: A transformer-based yolo segmentation model for breast mass detection and segmentation in digi- tal mammograms, Computer Methods and Pro- grams in Biomedicine 221 (2022) 106903.���� ��������������������������

    work page 2022

  26. [26]

    Zheng, F

    T. Zheng, F. Lin, X. Li, T. Chu, J. Gao, S. Zhang, Z. Li, Y . Gu, S. Wang, F. Zhao, et al., Deep learning-enabled fully automated pipeline system for segmentation and classifica- tion of single-mass breast lesions using contrast- enhanced mammography: a prospective, multi- centre study, EClinicalMedicine 58 (2023).���� ����������������������������

    work page 2023

  27. [27]

    A. A. Yurdusev, K. Adem, M. Hekim, Detection and classification of microcalcifications in mam- mograms images using difference filter and yolov4 deep learning model, Biomedical Signal Process- ing and Control 80 (2023) 104360.������� �����������������������

    work page 2023

  28. [28]

    K. Wang, N. Khan, R. Highnam, Automated seg- mentation of breast arterial calcifications from dig- ital mammography, in: 2019 International Con- ference on Image and Vision Computing New Zealand (IVCNZ), IEEE, 2019, pp. 1–6.���� �������������������������������

    work page 2019

  29. [29]

    Sakaida, T

    M. Sakaida, T. Yoshimura, M. Tang, S. Ichikawa, H. Sugimori, Development of a mammography calcification detection algorithm using deep learn- ing with resolution-preserved image patch divi- sion, Algorithms 16 (10) (2023) 483.������� ��������������

    work page 2023

  30. [30]

    A. Yala, P. G. Mikhael, F. Strand, G. Lin, K. Smith, Y .-L. Wan, L. Lamb, K. Hughes, C. Lehman, R. Barzilay, Toward robust mammography-based models for breast cancer risk, Science Transla- tional Medicine 13 (578) (2021) eaba4373.���� ����������������������������

    work page 2021

  31. [31]

    A. Yala, P. G. Mikhael, F. Strand, G. Lin, S. Sat- uluru, T. Kim, I. Banerjee, J. Gichoya, H. Trivedi, C. D. Lehman, et al., Multi-institutional val- idation of a mammography-based breast can- cer risk model, Journal of Clinical Oncology 40 (16) (2022) 1732–1740.���������������� ��������

    work page 2022

  32. [32]

    Dembrower, P

    K. Dembrower, P. Lindholm, F. Strand, A multi-million mammography image dataset and population-based screening cohort for the train- ing and evaluation of deep neural networks—the 23 cohort of screen-aged women (csaw), Journal of digital imaging 33 (2) (2020) 408–413.���� ��������������������������

    work page 2020

  33. [33]

    K. Liu, Y . Shen, N. Wu, J. Chł˛ edowski, C. Fernandez-Granda, K. J. Geras, Weakly- supervised high-resolution segmentation of mam- mography images for breast cancer diagnosis, Pro- ceedings of machine learning research 143 (2021) 268.�����������������������������

    work page 2021

  34. [34]

    Y . Tang, Z. Cao, Y . Zhang, Z. Yang, Z. Ji, Y . Wang, M. Han, J. Ma, J. Xiao, P. Chang, Leveraging large-scale weakly labeled data for semi-supervised mass detection in mammograms, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 3855–3864.���������������������� ����������

    work page 2021

  35. [35]

    Bakalo, R

    R. Bakalo, R. Ben-Ari, J. Goldberger, Classifica- tion and detection in mammograms with weak su- pervision via dual branch deep neural net, in: 2019 IEEE 16th International Symposium on Biomedi- cal Imaging (ISBI 2019), IEEE, 2019, pp. 1905– 1909.�����������������������������

    work page 2019

  36. [36]

    Y . Shen, N. Wu, J. Phang, J. Park, K. Liu, S. Tyagi, L. Heacock, S. G. Kim, L. Moy, K. Cho, et al., An interpretable classifier for high-resolution breast cancer screening images utilizing weakly super- vised localization, Medical image analysis 68 (2021) 101908.������������������������� ������

    work page 2021

  37. [37]

    Jiménez-Sánchez, M

    A. Jiménez-Sánchez, M. Tardy, M. A. G. Ballester, D. Mateus, G. Piella, Memory-aware curriculum federated learning for breast cancer classification, Computer Methods and Programs in Biomedicine 229 (2023) 107318.������������������� �����������

    work page 2023

  38. [38]

    H. R. Roth, K. Chang, P. Singh, N. Neumark, W. Li, V . Gupta, S. Gupta, L. Qu, A. Ihsani, B. C. Bizzo, et al., Federated learning for breast density classification: A real-world implementation, in: MICCAI Workshop on Domain Adaptation and Representation Transfer, Springer, 2020, pp. 181– 191.��������������������������������

    work page 2020

  39. [39]

    C. S. Perone, P. Ballester, R. C. Barros, J. Cohen- Adad, Unsupervised domain adaptation for med- ical imaging segmentation with self-ensembling, NeuroImage 194 (2019) 1–11.�������������� ����������������������

    work page 2019

  40. [40]

    Karani, K

    N. Karani, K. Chaitanya, C. Baumgartner, E. Konukoglu, A lifelong learning approach to brain mr segmentation across scanners and pro- tocols, in: International conference on medical image computing and computer-assisted interven- tion, Springer, 2018, pp. 476–484.������� �������������������������

    work page 2018

  41. [41]

    H. Guan, M. Liu, Domain adaptation for medical image analysis: a survey, IEEE Transactions on Biomedical Engineering 69 (3) (2021) 1173–1185. �����������������������������

    work page 2021

  42. [42]

    Ghafoorian, A

    M. Ghafoorian, A. Mehrtash, T. Kapur, N. Karsse- meijer, E. Marchiori, M. Pesteie, C. R. Guttmann, F.-E. De Leeuw, C. M. Tempany, B. Van Gin- neken, et al., Transfer learning for domain adap- tation in mri: Application in brain lesion seg- mentation, in: International conference on med- ical image computing and computer-assisted in- tervention, Springer, ...

    work page 2017

  43. [43]

    Y . Gu, Z. Ge, C. P. Bonnington, J. Zhou, Pro- gressive transfer learning and adversarial domain adaptation for cross-domain skin disease classi- fication, IEEE journal of biomedical and health informatics 24 (5) (2019) 1379–1393.������� ����������������������

    work page 2019

  44. [44]

    P. Laiz, J. Vitrià, S. Seguí, Using the triplet loss for domain adaptation in wce, in: 2019 IEEE/CVF In- ternational Conference on Computer Vision Work- shop (ICCVW), 2019, pp. 399–405.������� ���������������������

    work page 2019

  45. [45]

    Garrucho, K

    L. Garrucho, K. Kushibar, S. Jouide, O. Diaz, L. Igual, K. Lekadir, Domain generalization in deep learning based mass detection in mammog- raphy: A large-scale multi-center study, Artifi- cial Intelligence in Medicine 132 (2022) 102386. ��������������������������������

    work page 2022

  46. [46]

    Kumar, C

    D. Kumar, C. Kumar, M. Shao, Cross-database mammographic image analysis through unsuper- vised domain adaptation, in: 2017 IEEE interna- tional conference on big data (Big Data), IEEE, 2017, pp. 4035–4042.�������������������� ������������

    work page 2017

  47. [47]

    F. Ryan, K. L.-L. Román, B. Z. Gerbolés, K. M. Rebescher, M. S. Txurio, R. C. Ugarte, M. J. G. González, I. M. Oliver, Unsupervised domain adaptation for the segmentation of breast tissue 24 in mammography images, Computer Methods and Programs in Biomedicine 211 (2021) 106368. ������������������������������

    work page 2021

  48. [48]

    G. I. Quintana, V . Jugnon, L. Vancamberg, A. Des- olneux, M. Mougeot, Contrastive learning: an effi- cient domain adaptation strategy for 2d mammog- raphy image classification, in: 2024 IEEE Interna- tional Symposium on Biomedical Imaging (ISBI), IEEE, 2024, pp. 1–5.���������������������� �������������

    work page 2024

  49. [49]

    A. D. Lauritzen, M. C. von Euler-Chelpin, E. Lynge, I. Vejborg, M. Nielsen, N. Karssemeijer, M. Lillholm, Robust cross-vendor mammographic texture models using augmentation-based domain adaptation for long-term breast cancer risk, Jour- nal of Medical Imaging 10 (5) (2023) 054003– 054003.�����������������������������

    work page 2023

  50. [50]

    H. T. Nguyen, H. Q. Nguyen, H. H. Pham, K. Lam, L. T. Le, M. Dao, V . Vu, Vindr-mammo: A large- scale benchmark dataset for computer-aided di- agnosis in full-field digital mammography, Sci- entific Data 10 (1) (2023) 277.������������ ������������������

    work page 2023

  51. [51]

    Strand, CSAW-CC (mammography) – a dataset for AI research to improve screening, diagnostics and prognostics of breast cancer (2022).������� ��������������

    F. Strand, CSAW-CC (mammography) – a dataset for AI research to improve screening, diagnostics and prognostics of breast cancer (2022).������� ��������������. URL���������������������������������

    work page 2022

  52. [52]

    H. Lim, B. Kim, J. Choo, S. Choi, Ttn: A domain- shift aware batch normalization in test-time adap- tation, arXiv preprint arXiv:2302.05155 (2023). �����������������������������

    Pith/arXiv arXiv 2023

  53. [53]

    Y . Li, N. Wang, J. Shi, J. Liu, X. Hou, Revisit- ing batch normalization for practical domain adap- tation, arXiv preprint arXiv:1603.04779 (2016). �����������������������������

    Pith/arXiv arXiv 2016

  54. [54]

    Schneider, E

    S. Schneider, E. Rusak, L. Eck, O. Bringmann, W. Brendel, M. Bethge, Improving robustness against common corruptions by covariate shift adaptation, Advances in neural information pro- cessing systems 33 (2020) 11539–11551.���� �������������������������

    work page 2020

  55. [55]

    Frankle, D

    J. Frankle, D. J. Schwab, A. S. Morcos, Train- ing batchnorm and only batchnorm: On the expressive power of random features in cnns, arXiv preprint arXiv:2003.00152 (2020).���� �������������������������

    Pith/arXiv arXiv 2003

  56. [56]

    Ioffe, C

    S. Ioffe, C. Szegedy, Batch normalization: Ac- celerating deep network training by reducing in- ternal covariate shift, in: International conference on machine learning, pmlr, 2015, pp. 448–456. �����������������������������

    work page 2015

  57. [57]

    Santurkar, D

    S. Santurkar, D. Tsipras, A. Ilyas, A. Madry, How does batch normalization help optimization?, Ad- vances in neural information processing systems 31 (2018).�����������������������������

    work page 2018

  58. [58]

    Bjorck, C

    N. Bjorck, C. P. Gomes, B. Selman, K. Q. Wein- berger, Understanding batch normalization, Ad- vances in neural information processing systems 31 (2018).�����������������������������

    work page 2018

  59. [59]

    N. Wu, J. Phang, J. Park, Y . Shen, S. G. Kim, L. Heacock, L. Moy, K. Cho, K. J. Geras, The nyu breast cancer screening dataset v1. 0, New York Univ., New York, NY , USA, Tech. Rep (2019)

    2019

  60. [60]

    K. He, X. Zhang, S. Ren, J. Sun, Deep resid- ual learning for image recognition, in: Proceed- ings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778.���� ��������������������

    work page 2016

  61. [61]

    Ganin, E

    Y . Ganin, E. Ustinova, H. Ajakan, P. Ger- main, H. Larochelle, F. Laviolette, M. March, V . Lempitsky, Domain-adversarial training of neural networks, Journal of Machine Learning Research 17 (59) (2016) 1–35. �����������������������������. URL��������������������������� �����������

    work page 2016

  62. [62]

    R. Liaw, E. Liang, R. Nishihara, P. Moritz, J. E. Gonzalez, I. Stoica, Tune: A research plat- form for distributed model selection and training, arXiv preprint arXiv:1807.05118 (2018).���� �������������������������

    Pith/arXiv arXiv 2018

  63. [63]

    R. M. French, Catastrophic forgetting in con- nectionist networks, Trends in cognitive sci- ences 3 (4) (1999) 128–135.������������ ���������������������

    work page 1999

  64. [64]

    Y . Wu, K. He, Group normalization, in: Proceed- ings of the European conference on computer vi- sion (ECCV), 2018, pp. 3–19.������������ �����������������������������������

    work page 2018

  65. [65]

    J. L. Ba, J. R. Kiros, G. E. Hinton, Layer normalization, arXiv preprint arXiv:1607.06450 (2016).����������������������������� ����������������. 25

    Pith/arXiv arXiv 2016

  66. [66]

    S. Niu, J. Wu, Y . Zhang, Z. Wen, Y . Chen, P. Zhao, M. Tan, Towards stable test-time adap- tation in dynamic wild world, arXiv preprint arXiv:2302.12400 (2023).������������� ����������������

    Pith/arXiv arXiv 2023

  67. [67]

    B. Zhao, C. Chen, S.-T. Xia, Delta: degradation- free fully test-time adaptation, arXiv preprint arXiv:2301.13018 (2023).������������� ����������������

    Pith/arXiv arXiv 2023

  68. [68]

    Sendra-Balcells, V

    C. Sendra-Balcells, V . M. Campello, C. Martín- Isla, D. Viladés, M. L. Descalzo, A. Guala, J. F. Rodriguez-Palomares, K. Lekadir, Domain gen- eralization in deep learning for contrast-enhanced imaging, Computers in Biology and Medicine 149 (2022) 106052.������������������������� �����������. 26 Appendix A. Kernel Density Estimations Appendix A.1. KDEs of...

    work page 2022

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.