pith. sign in

arxiv: 2604.04470 · v1 · submitted 2026-04-06 · 📡 eess.IV · cs.AI

MC-GenRef: Annotation-free mammography microcalcification segmentation with generative posterior refinement

Hyunwoo Cho , Yeeun Kwon , Min Jung Kim , Yangmo Yoo This is my paper

Pith reviewed 2026-05-10 19:54 UTC · model grok-4.3

classification 📡 eess.IV cs.AI
keywords microcalcification segmentationmammographyannotation-freesynthetic datagenerative refinementtest-time adaptationrectified flow
0
0 comments X

The pith

Microcalcification segmentation in mammography works without real dense labels by training on synthetic patterns and refining predictions at test time with a generative prior.

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

The paper shows that a segmentor trained only on synthetic microcalcification examples, created by injecting plausible patterns into real negative mammogram patches, can achieve strong baseline performance. It then introduces test-time generative posterior refinement that uses a trained rectified-flow generator to iteratively adjust the output logits toward more consistent and edge-aware results. This combination targets the practical barriers of expensive and ambiguous pixel-level labels for tiny, sparse targets while aiming to cut missed detections that matter for early cancer screening. If the approach holds, clinics could build and update segmentation tools using existing negative cases plus a one-time synthetic generator rather than repeated expert annotation campaigns.

Core claim

MC-GenRef trains both a base segmentor and a seed-conditioned rectified-flow generator exclusively on synthetic image-mask pairs produced by a lightweight image formation model with local contrast modulation and blur. At inference, test-time generative posterior refinement derives a sparse seed from the initial prediction, forms seed-consistent projections through the generator, converts them into case-specific surrogate targets via the frozen segmentor, and refines the logits under overlap-consistent and edge-aware regularization, yielding higher recall and lower false-negative rates on INbreast and an external Yonsei cohort.

What carries the argument

Test-time generative posterior refinement (TT-GPR), which converts segmentation into iterative approximate posterior inference by conditioning a rectified-flow generator on sparse seeds and feeding its outputs back as surrogate targets.

Load-bearing premise

The lightweight image formation model with local contrast modulation and blur creates microcalcification patterns that are representative enough of real clinical cases to avoid texture-driven false positives or missed puncta in dense tissue.

What would settle it

A new external mammography test set where TT-GPR produces no measurable gain in recall or reduction in false-negative rate over the synthetic-only initializer, or where the refined outputs increase false positives in dense fibroglandular regions.

Figures

Figures reproduced from arXiv: 2604.04470 by Hyunwoo Cho, Min Jung Kim, Yangmo Yoo, Yeeun Kwon.

Figure 1
Figure 1. Figure 1: Overview of MC-GenRef. The framework consists of synthetic MC generation on real mammographic backgrounds, synthetic-only training of a segmentor and a seed-conditioned RF generator, and test-time generative posterior refinement on real mammograms [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
read the original abstract

Microcalcification (MC) analysis is clinically important in screening mammography because clustered puncta can be an early sign of malignancy, yet dense MC segmentation remains challenging: targets are extremely small and sparse, dense pixel-level labels are expensive and ambiguous, and cross-site shift often induces texture-driven false positives and missed puncta in dense tissue. We propose MC-GenRef, a real dense-label-free framework that combines high-fidelity synthetic supervision with test-time generative posterior refinement (TT-GPR). During training, real negative mammogram patches are used as backgrounds, and physically plausible MC patterns are injected through a lightweight image formation model with local contrast modulation and blur, yielding exact image-mask pairs without real dense annotation. Using only these synthetic labeled pairs, MC-GenRef trains a base segmentor and a seed-conditioned rectified-flow (RF) generator that serves as a controllable generative prior. During inference, TT-GPR treats segmentation as approximate posterior inference: it derives a sparse seed from the current prediction, forms seed-consistent RF projections, converts them into case-specific surrogate targets through the frozen segmentor, and iteratively refines the logits with overlap-consistent and edge-aware regularization. On INbreast, the synthetic-only initializer achieved the best Dice without real dense annotations, while TT-GPR improved miss-sensitive performance to Recall and FNR, with strong class-balanced behavior (Bal.Acc., G-Mean). On an external private Yonsei cohort ( n=50 ), TT-GPR consistently improved the synthetic-only initializer under cross-site shift, increasing Dice and Recall while reducing FNR. These results suggest that test-time generative posterior refinement is a practical route to reduce MC misses and improve robustness without additional real dense labeling.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes MC-GenRef, an annotation-free framework for microcalcification (MC) segmentation in mammography. It generates synthetic training pairs by injecting physically plausible MC patterns into real negative patches via a lightweight image formation model (local contrast modulation and blur), trains a base segmentor plus a seed-conditioned rectified-flow (RF) generator as a generative prior, and applies test-time generative posterior refinement (TT-GPR) that derives sparse seeds, forms RF projections, and iteratively refines logits using the frozen segmentor with overlap and edge regularization. On INbreast the synthetic-only initializer yields the highest Dice without real dense labels; TT-GPR further improves recall and reduces FNR. On an external Yonsei cohort (n=50) TT-GPR improves the initializer under cross-site shift.

Significance. If the synthetic patterns prove representative and TT-GPR reliably corrects misses without introducing new biases, the work offers a practical route to high-performance MC segmentation without expensive dense annotations, addressing a clinically relevant problem in screening mammography. The cross-site evaluation and use of a controllable generative prior for test-time refinement are strengths; the approach could generalize to other sparse-lesion tasks if the core assumptions hold.

major comments (2)
  1. [§3] §3 (Synthetic MC Generation / Image Formation Model): the central annotation-free claim rests on the assumption that the lightweight local-contrast-modulation-plus-blur model produces MC patterns whose statistics and appearance match real clinical microcalcifications sufficiently for a segmentor trained only on synthetics to generalize to dense tissue and cross-site data; no quantitative fidelity metrics (feature histograms, perceptual distances, or edge-profile comparisons) between synthetic and real MCs are reported, leaving the weakest link unverified.
  2. [§4] §4 (Experiments and Results): quantitative gains are stated for Dice, Recall, FNR, Bal.Acc. and G-Mean on INbreast and the Yonsei cohort, yet the manuscript provides neither component ablations for TT-GPR, nor statistical significance tests, nor a complete validation protocol (e.g., patient-level cross-validation details or confidence intervals); without these it is unclear whether reported improvements are attributable to the claimed modules or to unstated hyper-parameter choices.
minor comments (2)
  1. [§3.2] Notation for the rectified-flow generator and the overlap-consistent regularization term could be introduced more explicitly with equation numbers to aid reproducibility.
  2. [Figure 2] Figure captions for the TT-GPR pipeline diagram should explicitly label the iterative refinement loop and the role of the frozen segmentor.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive review. The comments identify important areas where additional evidence and rigor can strengthen the manuscript. We address each major comment below and commit to the corresponding revisions.

read point-by-point responses

  1. Referee: [§3] §3 (Synthetic MC Generation / Image Formation Model): the central annotation-free claim rests on the assumption that the lightweight local-contrast-modulation-plus-blur model produces MC patterns whose statistics and appearance match real clinical microcalcifications sufficiently for a segmentor trained only on synthetics to generalize to dense tissue and cross-site data; no quantitative fidelity metrics (feature histograms, perceptual distances, or edge-profile comparisons) between synthetic and real MCs are reported, leaving the weakest link unverified.

    Authors: We agree that direct quantitative fidelity metrics would provide stronger support for the synthetic data assumption. Our primary evidence for the utility of the image formation model is the downstream segmentation performance: the synthetic-only initializer achieves the highest Dice on INbreast and improves under cross-site shift on the Yonsei cohort. Nevertheless, we acknowledge the gap. In the revised manuscript we will add quantitative comparisons, including intensity histogram distances, edge-profile statistics, and perceptual feature distances computed on real MC patches extracted from the INbreast test set versus matched synthetic patches. revision: yes

  2. Referee: [§4] §4 (Experiments and Results): quantitative gains are stated for Dice, Recall, FNR, Bal.Acc. and G-Mean on INbreast and the Yonsei cohort, yet the manuscript provides neither component ablations for TT-GPR, nor statistical significance tests, nor a complete validation protocol (e.g., patient-level cross-validation details or confidence intervals); without these it is unclear whether reported improvements are attributable to the claimed modules or to unstated hyper-parameter choices.

    Authors: We accept that the current experimental section lacks the requested rigor. While the reported gains reflect the full pipeline versus the initializer, we did not include component ablations, statistical tests, or expanded protocol details in the original submission. In the revision we will add: (i) ablations isolating the contributions of seed derivation, RF projection, and the two regularizers; (ii) statistical significance testing (paired Wilcoxon signed-rank tests with p-values) on the metric improvements; and (iii) a more complete validation protocol describing patient-level splitting, confidence intervals, and hyper-parameter sensitivity. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper's chain proceeds from an independent lightweight synthetic MC injection model (applied to real negative patches to create exact image-mask pairs) to training a base segmentor and seed-conditioned RF generator, followed by TT-GPR that uses the frozen segmentor to produce surrogate targets from RF projections during inference. None of these steps reduce by construction to the reported Dice/Recall/FNR metrics or to any fitted parameters defined from the target data; the generative prior and refinement operate without reference to real dense labels or performance numbers. Evaluation on INbreast and the external Yonsei cohort uses standard held-out annotations, providing external validation independent of the training pipeline. No self-citations, uniqueness theorems, or ansatzes are invoked as load-bearing elements in the provided text.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 2 invented entities

The central claim rests on the fidelity of the synthetic data generator and the ability of the rectified-flow model to produce useful surrogate targets during refinement.

free parameters (1)
  • local contrast modulation and blur parameters
    Chosen in the image formation model to create plausible MC patterns; values not specified in abstract.
axioms (2)
  • domain assumption Synthetic MC patterns generated by the lightweight image formation model are representative of real clinical microcalcifications
    Invoked to justify training exclusively on synthetic pairs.
  • domain assumption The seed-conditioned rectified-flow generator provides a controllable prior that improves posterior inference for segmentation
    Core premise of the TT-GPR step.
invented entities (2)
  • Test-time generative posterior refinement (TT-GPR) no independent evidence
    purpose: Iterative refinement of segmentation logits using generative projections and regularization
    New inference procedure introduced in the framework.
  • seed-conditioned rectified-flow (RF) generator no independent evidence
    purpose: Controllable generative prior for creating surrogate targets from sparse seeds
    Introduced as the generative component serving the refinement process.

pith-pipeline@v0.9.0 · 5615 in / 1628 out tokens · 39845 ms · 2026-05-10T19:54:33.092818+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages · 2 internal anchors

  1. [1]

    Breast microcalcifications: Biological and diagnostic perspectives,

    R. Bonfiglio, M. Scimeca, N. Urbano et al. , "Breast microcalcifications: Biological and diagnostic perspectives," 30, Taylor & Francis, 2018, pp. 3097-3099

    work page 2018

  2. [2]

    T detection of breast microcalcifications with medical ,

    M. E. Anderson, M. S. Soo, R. C. Bentley et al., “T detection of breast microcalcifications with medical ,” The Journal of the Acoustical Society of America, vol. 101, no. 1, pp. 29-39, 1997

    work page 1997

  3. [3]

    G. Tse, P. H. Tan, A. L. Pang et al., “ : ’ ,” Journal of clinical pathology, vol. 61, no. 2, pp. 145-151, 2008

    work page 2008

  4. [4]

    morphology and distribution classification for breast mammograms with multi -task graph convolutional neural ,

    H. Du, M. M.-S. Yao, S. Liu et al., “ morphology and distribution classification for breast mammograms with multi -task graph convolutional neural ,” IEEE Journal of Biomedical and Health Informatics, vol. 27, no. 8, pp. 3782-3793, 2023

    work page 2023

  5. [5]

    z ‐ segmentation of microcalcification clusters in ,

    , J Y , E F y J , “ z ‐ segmentation of microcalcification clusters in ,” Medical Physics, vol. 29, no. 4, pp. 475 - 483, 2002

    work page 2002

  6. [6]

    neural networks for the segmentation of microcalcification y ,

    G. Valvano, G. Santini, N. Martini et al., “ neural networks for the segmentation of microcalcification y ,” Journal of healthcare engineering, vol. 2019, no. 1, pp. 9360941, 2019

    work page 2019

  7. [7]

    : Automatic segmentation and classification of breast ,

    A. Gerbasi, G. Clementi, F. Corsi et al., “ : Automatic segmentation and classification of breast ,” Computer Methods and Programs in Biomedicine, vol. 235, pp. 107483, 2023

    work page 2023

  8. [8]

    microcalcifications: A multi- ,

    C. Marasinou, B. Li, J. Paige et al., “ microcalcifications: A multi- ,” arXiv preprint arXiv:2102.00754, 2021

    work page arXiv 2021

  9. [9]

    x -sensitive deep learning ,

    J W , Y Y , “ x -sensitive deep learning ,” Pattern recognition, vol. 78, pp. 12-22, 2018

    work page 2018

  10. [10]

    Improving segmentation of breast arterial calcifications from digital mammography: good annotation is all you need

    K. Wang, M. Hill, S. Knowles -Barley et al. , "Improving segmentation of breast arterial calcifications from digital mammography: good annotation is all you need." pp. 130 - 146

    work page

  11. [11]

    mammography data set for use in computer -aided detection ,

    R. S. Lee, F. Gimenez, A. Hoogi et al. , “ mammography data set for use in computer -aided detection ,” Scientific data, vol. 4, no. 1, pp. 170177, 2017

    work page 2017

  12. [12]

    : toward a full - ,

    I. C. Moreira, I. Amaral, I. Domingues et al., “ : toward a full - ,” Academic radiology, vol. 19, no. 2, pp. 236-248, 2012

    work page 2012

  13. [13]

    Multi -source weak supervision for saliency detection

    Y. Zeng, Y. Zhuge, H. Lu et al. , "Multi -source weak supervision for saliency detection." pp. 6074-6083

    work page

  14. [14]

    Multi-scale mass segmentation for mammograms via cascaded random forests

    H. Min, S. S. Chandra, N. Dhungel et al., "Multi-scale mass segmentation for mammograms via cascaded random forests." pp. 113-117

    work page

  15. [15]

    of microcalcification clusters using hessian matrix and foveal segmentation method on multiscale analysis in digital ,

    T , V , y, “ of microcalcification clusters using hessian matrix and foveal segmentation method on multiscale analysis in digital ,” Journal of digital imaging, vol. 25, no. 5, pp. 607-619, 2012

    work page 2012

  16. [16]

    -scale generative tumor synthesis in computed tomography images for ,

    L. Wu, J. Zhuang, Y. Zhou et al., “ -scale generative tumor synthesis in computed tomography images for ,” Nature Communications, vol. 16, no. 1, pp. 11053, 2025

    work page 2025

  17. [17]

    domain adaptation with optimal transport in multi -site ,

    A. Ackaouy, N. Courty, E. Vallée et al., “ domain adaptation with optimal transport in multi -site ,” Frontiers in computational neuroscience, vol. 14, pp. 19, 2020

    work page 2020

  18. [18]

    -based diffusion models for ,

    H , J Y , “ -based diffusion models for ,” Medical image analysis, vol. 80, pp. 102479, 2022

    work page 2022

  19. [19]

    Score-Based Generative Modeling through Stochastic Differential Equations

    Y. Song, J. Sohl -Dickstein, D. P. Kingma et al., “ - based generative modeling through stochastic differential q ,” arXiv preprint arXiv:2011.13456, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  20. [20]

    A novel focal tversky loss function with improved attention u -net for lesion segmentation

    N. Abraham, and N. M. Khan, "A novel focal tversky loss function with improved attention u -net for lesion segmentation." pp. 683-687

    work page

  21. [21]

    Restormer: Efficient transformer for high -resolution image restoration

    S. W. Zamir, A. Arora, S. Khan et al., "Restormer: Efficient transformer for high -resolution image restoration." pp. 5728-5739

    work page

  22. [22]

    , F H , “ y z ,” arXiv preprint arXiv:1711.05101, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  23. [23]

    Exponential Moving Average of Weights in Deep Learning: Dynamics and Benefits

    D. Morales -Brotons, T. Vogels, and H. Hendrikx, “Ex : y ,” arXiv preprint arXiv:2411.18704 , 2024

    work page arXiv 2024

  24. [24]

    Optuna: A next - generation hyperparameter optimization framework

    T. Akiba, S. Sano, T. Yanase et al. , "Optuna: A next - generation hyperparameter optimization framework." pp. 2623-2631

    work page

  25. [25]

    C. Carr, F. Kitamura, G. Partridge et al., “ y ,” Kaggle, 2022

    work page 2022