Pith. sign in

REVIEW 4 major objections 2 minor 7 references

CoFi: A Fast Coarse-to-Fine Few-Shot Pipeline for Glomerular Basement Membrane Segmentation

T0 review · 4 major / 2 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that segmenting the glomerular basement membrane can be done from three annotated EM images at 74.54% Dice and 1.9 FPS using a coarse-to-fine SAM prompting pipeline.

desk verdict The submission is a different paper than the abstract promises; the CoFi idea is plausible but the evidence is absent. read the letter →

arxiv 2508.11469 v1 pith:YNAGGCV5 submitted 2025-08-15 cs.CV

Hongjin Fang , Daniel Reisenbüchler , Kenji Ikemura , Mert R. Sabuncu , Yihe Yang , Ruining Deng This is my paper · ORCID
classification cs.CV
keywords glomerularbasementmembranefew-shotsegmentationelectronmicroscopySAMpointpromptscoarse-to-finepipelinemedicalimagekidneypathologyannotationefficiency
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 proposes CoFi, a coarse-to-fine few-shot pipeline for segmenting the glomerular basement membrane (GBM) in electron microscopy images. Its claim is that training a lightweight network on only three annotated images, converting the resulting coarse mask into morphology-pruned point prompts, and feeding those prompts to the Segment Anything Model (SAM) yields a final segmentation with a Dice coefficient of 74.54% at 1.9 frames per second. If correct, accurate GBM segmentation—needed for quantifying membrane thickness and supporting kidney-disease diagnosis—would no longer require extensive pixel-level annotation. That would make the measurement practical in clinical and research workflows.

What carries the argument

The load-bearing mechanism is the prompt-generation chain: a lightweight CNN trained on three images outputs a coarse GBM mask; morphology-aware pruning removes implausible fragments and selects representative foreground and background point prompts; these prompts condition SAM's mask decoder to produce the final segmentation. SAM (Segment Anything Model) is a pretrained promptable segmentation model; the paper's contribution is showing how to obtain prompts for an out-of-distribution EM structure with almost no annotation.

What would settle it

On the same test set, compare CoFi against (1) the coarse model alone and (2) CoFi with morphology-aware pruning disabled. If the coarse-only Dice is close to 74.54%, or if removing pruning does not reduce Dice, the reported result is not produced by the claimed prompt-generation mechanism.

Watch

Extended reading notes

Core claim

CoFi's central discovery is that heavy supervised training is not necessary for GBM segmentation: a small model trained on three annotated EM images can produce a coarse membrane mask, and that mask can be automatically transformed into high-quality point prompts through morphology-aware pruning. These prompts, rather than raw pixels, are then given to SAM, whose pretrained ability to follow points lets it recover a sharper boundary. The paper reports 74.54% Dice and 1.9 FPS, attributing the gain to the interaction between the cheap coarse model and SAM's refinement.

Load-bearing premise

The claimed accuracy rests on the coarse model, trained from three images, producing point prompts whose precision and recall survive morphology-aware pruning—and the paper does not directly report that coarse model's accuracy or an ablation of the pruning step.

Editorial extensions

If this is right

  • With only three annotated EM images, a clinically relevant GBM segmentation pipeline can run at interactive speed, 1.9 FPS.
  • The annotation burden for GBM thickness quantification in kidney pathology drops dramatically, bypassing the need for large pixel-level labeled datasets.
  • SAM's general segmentation capability can be adapted to EM ultrastructure through automatically generated point prompts, not through fine-tuning.
  • The combination of speed and annotation efficiency makes the pipeline suitable for research workflows and a candidate for clinical decision support in renal pathology.

Reading between the lines

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

  • The paper does not report the coarse model's standalone Dice or an ablation of the morphology-aware pruning step, so the distinct contribution of SAM refinement is not yet isolated; measuring it would be a direct next step.
  • Because point prompting is structure-agnostic, the same few-shot recipe should transfer to other EM ultrastructures—podocyte foot processes, tubular basement membranes, or mitochondrial membranes—whenever a weak coarse segmenter can be trained cheaply.
  • A human-in-the-loop variant, where a pathologist adds or deletes a few point prompts on failure cases, could improve boundaries at no additional training cost; the static pipeline does not exploit this.
  • Testing with one, three, and five training images would map the annotation-efficiency frontier and show where the method stops improving.
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 / 2 minor

Summary. The submission consists of an abstract for a paper titled 'CoFi: A Fast Coarse-to-Fine Few-Shot Pipeline for Glomerular Basement Membrane Segmentation' and a full text that is an unrelated IEEE journal paper on NOMA-assisted semi-grant-free systems (arXiv:2508.11473v1). The abstract claims a few-shot pipeline that trains a lightweight network on three annotated images, generates morphology-pruned point prompts, and uses SAM to refine GBM segmentation, reporting 74.54% Dice at 1.9 FPS. However, the full text contains no description of the CoFi method, no experimental setup, no dataset details, no code, and no results. As submitted, the central claims cannot be evaluated.

Significance. If the reported result were properly substantiated, the paper could offer a practical contribution to few-shot segmentation in renal pathology by reducing annotation burden while maintaining interactive speed. The intended coarse-to-fine design with SAM is coherent and the public repository link suggests an intention to share code. However, the submission provides only point estimates without error bars, test-set size, patient-level splitting, baseline comparisons, coarse-model accuracy, pruning ablations, or SAM transfer evidence. No machine-checked proofs or reproducible artifacts are present in the manuscript. The significance of the claimed outcome is therefore unassessable from the submitted text.

major comments (4)
  1. [Full text (arXiv:2508.11473v1)] The full text of this submission is a wireless communications paper on 'Reducing AoI and Improving Throughput for NOMA-assisted SGF Systems'. It contains no mention of GBM, SAM, few-shot learning, morphology-aware pruning, or any component of CoFi. The abstract's claims are thus entirely unsupported by the manuscript body. This is a load-bearing mismatch that prevents any technical assessment.
  2. [Abstract] The headline result, 'Dice coefficient of 74.54% and an inference speed of 1.9 FPS', is presented as a single point estimate with no error bars, no number of test images, no patient-level split, and no statistical comparison. The word 'exceptional' is not justified. Without variability and test-set details, this number cannot be interpreted.
  3. [Abstract – coarse-to-fine prompting] The abstract states that the coarse mask is 'automatically processed to generate high-quality point prompts with morphology-aware pruning', but reports neither the coarse model's standalone Dice/recall/precision nor an ablation of the pruning thresholds and radii. Since the final Dice cannot exceed the information available from the coarse mask, this omission is central: the reader cannot tell whether the reported performance depends on the coarse model's quality, the pruning choices, or the prompt placement.
  4. [Abstract – SAM transfer] No evidence is provided that SAM, trained primarily on natural images, transfers to electron microscopy or that refinement by point prompts improves rather than degrades the boundary. Without a comparison to SAM with alternative prompting (e.g., bounding boxes, dense prompts) or to end-to-end few-shot segmentation models, the claimed benefit of the coarse-to-fine scheme is unsupported.
minor comments (2)
  1. [Abstract] The term 'exceptional GBM segmentation performance' should be replaced with quantitative comparisons and variability measures (e.g., mean ± std over multiple runs or cross-validation folds).
  2. [Repository] The GitHub URL is mentioned, but no code or instructions are included in the submission. If the correct manuscript is provided, the repository contents should be summarized and the exact version used for the reported results should be specified.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity derivable: the CoFi full text is an unrelated NOMA paper, so there is no derivation chain to audit; the abstract's empirical claim is unverified but not circular.

full rationale

The submitted full text under arXiv:2508.11469 is not the CoFi paper; it is arXiv:2508.11473v1, a NOMA/SGF wireless communications manuscript. Consequently, the claimed derivation chain (coarse model -> morphology-pruned point prompts -> SAM refinement -> 74.54% Dice) is present only in the abstract, with no equations, no architecture, no pruning algorithm, no dataset split, and no evaluation protocol. There is therefore no paper-internal equation or self-citation chain that can be checked for circularity. The abstract's performance claim is an empirical result, not a first-principles derivation; even if unsupported, unsupportedness is a verification/correctness risk, not circularity. The two-stage prompt-from-mask design means the final Dice inherits coarse-model errors, but that is a known limitation of any such pipeline, not a definitional reduction. No step can be quoted that makes a prediction equal to its input by construction. Score 0.

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

The pipeline introduces no new physical or mathematical entities. The claim rests on standard ML components (a trained CNN, SAM) plus tunable hyperparameters and the domain assumptions listed above. The automated prompting design, if it includes novel algorithmic objects, is not specified in the abstract.

free parameters (3)
  • morphology-aware pruning thresholds and radii = not reported in abstract
    The pruning step that decides which coarse-mask points become SAM prompts presumably relies on tunable morphology thresholds (distance or thickness criteria); values are not given.
  • number and placement of point prompts per GBM = not reported
    SAM point-prompt guidance depends on prompt count and placement; unspecified in the abstract.
  • coarse network architecture and training hyperparameters = not reported
    The lightweight network trained on three images has unspecified architecture, loss, and epochs; these affect the coarse mask quality on which the whole pipeline depends.
assumptions (3)
  • domain assumption A coarse segmenter trained on exactly three annotated GBM images generalizes well enough to produce useful point prompts on unseen EM images.
    Invoked in the abstract's pipeline description; if false, the coarse mask misses the GBM and SAM cannot recover.
  • domain assumption SAM's pretrained representations transfer to electron microscopy images of the glomerular basement membrane.
    The refinement stage relies on SAM responding correctly to point prompts in a domain (EM) far from its natural-image training distribution.
  • domain assumption Point prompts generated from the coarse mask (after pruning) contain more signal than the coarse mask itself, so SAM refinement improves Dice.
    This is the design hypothesis of the pipeline; the abstract reports only the final Dice, not coarse-only or SAM-only ablations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CoFi: A Fast Coarse-to-Fine Few-Shot Pipeline for Glomerular Basement Membrane Segmentation." pith.science (2026). https://pith.science/paper/YNAGGCV5

@misc{pith2026250811469,
  author       = {Pith},
  title        = {Pith review of: CoFi: A Fast Coarse-to-Fine Few-Shot Pipeline for Glomerular Basement Membrane Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YNAGGCV5}},
  note         = {Machine review of arXiv:2508.11469}
}
read the original abstract

Accurate segmentation of the glomerular basement membrane (GBM) in electron microscopy (EM) images is fundamental for quantifying membrane thickness and supporting the diagnosis of various kidney diseases. While supervised deep learning approaches achieve high segmentation accuracy, their reliance on extensive pixel-level annotation renders them impractical for clinical workflows. Few-shot learning can reduce this annotation burden but often struggles to capture the fine structural details necessary for GBM analysis. In this study, we introduce CoFi, a fast and efficient coarse-to-fine few-shot segmentation pipeline designed for GBM delineation in EM images. CoFi first trains a lightweight neural network using only three annotated images to produce an initial coarse segmentation mask. This mask is then automatically processed to generate high-quality point prompts with morphology-aware pruning, which are subsequently used to guide SAM in refining the segmentation. The proposed method achieved exceptional GBM segmentation performance, with a Dice coefficient of 74.54% and an inference speed of 1.9 FPS. We demonstrate that CoFi not only alleviates the annotation and computational burdens associated with conventional methods, but also achieves accurate and reliable segmentation results. The pipeline's speed and annotation efficiency make it well-suited for research and hold strong potential for clinical applications in renal pathology. The pipeline is publicly available at: https://github.com/ddrrnn123/CoFi.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

7 extracted references · 7 canonical work pages

  1. [25]

    Distributed age-of-information scheduling with NOMA via deep reinforcement learning,

    C. Zhang, Y . Zou, Z. Zhang, D. Yu, J. T. G ´omez, T. Lan, F. Dressler, and X. Cheng, “Distributed age-of-information scheduling with NOMA via deep reinforcement learning,” IEEE Trans. Mob. Comput. , vol. 24, no. 1, pp. 30–44, 2025

    work page 2025

  2. [26]

    Deep reinforcement learning based AoI minimization for NOMA-enabled integrated satellite- terrestrial networks,

    X. He, Y . Yang, J. Lee, G. He, and Q. Yan, “Deep reinforcement learning based AoI minimization for NOMA-enabled integrated satellite- terrestrial networks,” IEEE Trans. Veh. Technol., pp. 1–6, 2024

    work page 2024

  3. [27]

    Age of information: Can CR- NOMA help?

    Z. Ding, R. Schober, and H. V . Poor, “Age of information: Can CR- NOMA help?” IEEE Trans. Commun. , 2023

    work page 2023

  4. [28]

    Real-time status: How often should one update?

    S. Kaul, R. Yates, and M. Gruteser, “Real-time status: How often should one update?” in 2012 Proc. IEEE INFOCOM . IEEE, 2012, pp. 2731– 2735

    work page 2012

  5. [29]

    Update or wait: How to keep your data fresh,

    Y . Sun, E. Uysal-Biyikoglu, R. D. Yates, C. E. Koksal, and N. B. Shroff, “Update or wait: How to keep your data fresh,” IEEE Trans. Inf. Theory, vol. 63, no. 11, pp. 7492–7508, 2017

    work page 2017

  6. [30]

    Age of information and throughput in random access-based IoT systems with periodic updating,

    Y . H. Bae and J. W. Baek, “Age of information and throughput in random access-based IoT systems with periodic updating,”IEEE Wirel. Commun. Lett., vol. 11, no. 4, pp. 821–825, 2022

    work page 2022

  7. [31]

    Online altitude control and scheduling policy for minimizing AoI in UA V-assisted IoT wireless networks,

    M. Samir, C. Assi, S. Sharafeddine, and A. Ghrayeb, “Online altitude control and scheduling policy for minimizing AoI in UA V-assisted IoT wireless networks,” IEEE Trans. Mob. , vol. 21, no. 7, pp. 2493–2505, 2022

    work page 2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.