Weakly supervised neural network: segmentation of complex structures in X-ray microCT
Daniele Rusconi, Michela Ascolese, Stephanie Fest-Santini, Alberto Bravin, Maurizio Santini
cs.AI · cs.CV
Quick verdict
Weak-supervision nnU-Net for microCT glomeruli: sparse+expert mix reaches Dice 0.5748 vs 0.7710 fully supervised; evaluation protocol under-specified.
Summary
The paper adapts a 2D nnU-Net to weak supervision for segmenting renal glomeruli in X-ray microCT slices of Microfil-perfused rat kidneys (three animals, 6.81 µm isotropic resolution). Experts fully segmented ~300 slices (15–20 min each) while non-experts placed 8-pixel dot markers on ~800 slices (≤2 min each). Training combines a standard Dice loss on fully labeled slices with a Dice variant restricted to dot-annotation regions, weighted λ=0.5 chosen empirically without a sweep. Reported pseudo Dice scores: 0.3223±0.0765 (weak labels only, 1152 slices), 0.5748±0.1358 (85% weak / 15% high-quality), 0.7710±0.0499 (fully supervised). The authors conclude that sparse annotations give a meaningful learning signal and that a small expert subset approaches fully supervised quality. Main limitations: the train/test protocol and 'pseudo Dice' metric are never defined; Table 1's 1152-label counts conflict with the stated ~300/~800 slices; no classical or point-supervision baselines, no λ ablation; no code or data availability statement.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 18 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Sparse dot annotations placed by non-experts provide a meaningful learning signal, enabling the network to identify likely glomerulus regions without dense labels (weak-only pseudo Dice 0.3223).
- Mixing 85% weakly annotated with 15% fully segmented slices recovers a large portion of the gap to fully supervised training (0.5748 vs 0.7710 pseudo Dice).
- The approach substantially reduces annotation effort (≤2 min/slice dot placement vs 15–20 min/slice expert segmentation) while approaching fully supervised accuracy.
- The Dice loss formulation is suboptimal for dot-based labels, and loss functions tailored to point annotations (centroid- or distance-based) may further improve performance.
Strengths
Concrete annotation-cost trade-off
§3.1 quantifies the economics precisely: expert full segmentation costs 15–20 min/slice (~300 slices) vs ≤2 min/slice dot placement by untrained annotators (~800 slices, 8-pixel ≈ 80 µm dots) — the paper's core value proposition is grounded in real measured effort.
Well-documented imaging pipeline
§2 specifies animal model (three 14-week SD rats, Microfil MV-122 perfusion), scanner settings (120 kV, 20 µA, 3200 projections, 1.8 s/frame), 6.81 µm isotropic voxels with metrological calibration, FDK reconstruction, beam-hardening mitigation, and ethics approval 991/2023-PR.
Honest self-assessment of method limits
Clear three-regime design
Weaknesses
Evaluation protocol undefined
Dataset arithmetic contradiction
No external baselines or ablations
'Reliable localization' claim unsupported at object level
Reproducibility gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (prepared for JINST) investigates whether weakly supervised deep learning can reduce annotation cost for segmenting renal glomeruli — small, low-contrast structures — in X-ray microCT volumes of Microfil-perfused rat kidneys. A 2D nnU-Net is trained under three supervision regimes: dot-only weak labels, a mixture of weak and fully segmented slices, and full supervision. Central claims (Abstract, §4, §5): (1) sparse dot annotations from non-experts provide a meaningful learning signal enabling 'reliable localization'; (2) adding a small expert-labeled subset (15% of labels) recovers a large share of the gap to fully supervised performance; (3) Dice-type loss is suboptimal for point labels, so tailored losses (centroid- or distance-based) could improve results. The scope is deliberately narrow: one organ, one species, 2D slices, one architecture family.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is weakly supervised segmentation in this paper?
Instead of drawing full pixel-level masks (which took experts 15–20 minutes per image), non-expert annotators placed a single small dot marker (8 pixels, ~80 µm) on each glomerulus in under 2 minutes per image. The neural network is trained with a loss that only checks its predictions near those dots, plus a small set of fully segmented images to anchor training.
How close does the mixed weak/expert approach get to full manual annotation quality?
On rat kidney microCT slices, training with 85% dot-annotated and 15% fully segmented slices reached a pseudo Dice score of 0.5748, versus 0.7710 for a fully supervised model trained only on expert segmentations. Dot annotations alone reached only 0.3223.
How much annotation time can this approach save?
In the paper, expert full segmentation of one microCT slice took 15–20 minutes, while an untrained person could dot-annotate a slice in under 2 minutes — roughly a 10× reduction in per-image effort, with about 15% expert-labeled images needed to approach fully supervised accuracy.
Is the code or the dataset from this paper publicly available?
No. The paper contains no code or data availability statement. The data are microCT volumes of kidneys from three laboratory rats, and the evaluation protocol (train/test split, metric definition) is not fully described, so the results cannot currently be independently reproduced.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–3 are present only as captions in the text extraction; the actual image content could not be inspected, so qualitative claims were assessed from text and Table 1 values only.
- Equation 3.2 as extracted shows a numerator Σ p_i with no explicit ground-truth term; this may be an OCR artifact of sub/superscripts, and the loss was interpreted as written (consistent with the surrounding prose).
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.