REVIEW 3 major objections 4 minor 25 references
False Promises in Medical Imaging AI? Assessing Validity of Outperformance Claims
T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash
Pith's one-line read A Bayesian re-analysis finds most medical-imaging AI outperformance claims unsubstantiated: >5% chance of a false claim in 86% of classification and 53% of segmentation papers.
desk verdict A useful, important audit of MICCAI reporting practices, but the headline 86%/53% false-claim rates are worst-case estimates that depend on an unverified congruence assumption. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is model congruence, the unmeasured degree to which two competing methods' per-sample errors align: for classification, the proportion of test items both methods classify correctly, $\hat{p}_{1,1}$; for segmentation, the correlation $r_{AB}$ between the two methods' per-image Dice scores. Congruence enters two Bayesian identities that carry the argument: the classification posterior is $\mathrm{Dirichlet}(x_1+1,\,x_2+1,\,n-x_1-x_2+2)$ with $x_1=n(\hat{p}_A-\hat{p}_{1,1})$ and $x_2=n(\hat{p}_B-\hat{p}_{1,1})$, and the segmentation posterior for the winner-minus-runner-up mean Dice is a Student-$t$ with $n-1$ degrees of freedom and spread $\sqrt{s_A^2+s_B^2-2s_A s_B r_{AB}}$. Since no analyzed paper reports congruence, the authors take its empirical distribution from external tasks, replace each paper's value with the median (or Q1/Q3 in sensitivity analysis), and clamp into the feasible bounds implied by the reported accuracies; the clamped classification values are often at the lower bound $p_A+p_B-1$, the point of maximal disagreement between the top two methods.
What would settle it
Take any set of medical-imaging papers that report per-test-sample predictions or confidence intervals, compute the actual accuracy congruence or Dice correlation between the top two methods, and recompute the false-claim probabilities; if the true congruence values approach the upper feasible bound, the share of papers exceeding the 5% threshold would drop well below the reported 86% and 53%.
Extended reading notes
Core claim
The paper's central claim is that most outperformance claims in medical imaging AI are unsubstantiated: given only the means, test-set sizes, and a reasonable assumption about how much the two leading methods agree, the probability that the runner-up is actually as good as or better than the winner exceeds conventional thresholds in the majority of papers. To get this probability it derives posterior distributions for the two true performances under non-informative priors — a Dirichlet posterior over accuracies for classification, a Student-$t$ posterior for the difference in mean Dice for segmentation — and evaluates them with Monte Carlo sampling. Because the agreement quantity (model congruence) is never reported, it is imported from external experiments at median values of 0.67 for both tasks, with each paper's value clamped to the feasible range. The headline numbers are 86% of classification and 53% of segmentation papers exceeding a 5% false-claim probability, and 58% and 25% exceeding 30%; the paper reads this as evidence that typical benchmarking claims do not survive a chance test.
Load-bearing premise
The results stand or fall on the assumed model congruence — the unobserved agreement between the two methods' predictions — which is imported from external datasets and, for most classification papers, clamped to its lowest feasible value, a choice that maximizes the estimated probability of false claims.
Editorial extensions
If this is right
- A reported accuracy gain of 0.01 needs a test set roughly eight times the current median of 500 images to get the false-claim probability below 5%.
- A reported Dice gain of 0.01 needs roughly ten times the median segmentation test set of 62 images.
- With fewer than 15% of papers doing any statistical significance testing, most published superiority claims currently contain no internal check that the winning method is distinguishable from noise.
- Reviewers can use the paper's lookup tables or formulas to flag a candidate claim as likely false from the reported test size and performance gap alone.
- If larger test sets and uncertainty reporting became standard, many 'state-of-the-art' results would need to be reframed, redirecting attention to methods whose gains actually exceed sampling noise.
Reading between the lines
- The same calculation could be run on other venues and on metrics such as AUC or F1; because an image-level classification yields only one bit of evidence per sample, false-claim rates on such metrics are likely at least as high as those reported for accuracy.
- A direct policy test of the paper's thesis: if the conference required authors to report per-sample predictions or confidence intervals, one could measure whether the distribution of claimed gains and test sizes shifted and whether replication rates improved.
- The formulas also suggest a cheap reviewer-side check — treating any reported improvement below about 0.01–0.02 at test sizes of a few hundred samples as unproven without paired statistics — which the paper supports implicitly but does not itself state.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper studies the validity of outperformance claims in medical imaging AI by analyzing a cohort of 347 MICCAI 2023 classification and segmentation papers. The authors document reporting practices (e.g., >80% of papers bold the best mean metric, yet <15% perform statistical testing), then introduce a Bayesian framework to compute the probability that the top-ranked method is not truly better than the second-ranked method given reported test set sizes, reported metric differences, and externally estimated 'model congruence' parameters. Using median congruence assumptions, they report that >5% false-claim probability occurs in 86% of classification papers and 53% of segmentation papers, and that demonstrating a 0.01 accuracy or DSC improvement requires test sets roughly 8-10 times larger than current medians. The paper provides detailed derivations in the Supplementary Information, as well as code and practical lookup tables for reviewers.
Significance. If the central prevalence numbers are robust, this is an important and field-relevant contribution: it quantifies a suspected weakness in benchmarking practices and supplies concrete, reusable formulas and sample-size guidance. Strengths include the carefully documented manual data extraction with double annotation and conflict resolution, the transparent Bayesian derivations in the Supplementary Information, the availability of code, and the practical framing for reviewers. The finding that typical reported deltas of 0.01 are often within sampling noise given common test set sizes is compelling and likely robust. However, the headline percentages (86% and 53%) depend critically on model-congruence values that are not directly observable in the analyzed papers; the current sensitivity analysis does not explore the full feasible range of these parameters, and in the classification case the procedure clamps the congruence to the endpoint that maximizes false-claim probabilities.
major comments (3)
- [Methods, Data analysis subset; Supplementary C.2] The classification false-claim probability depends on the model congruence p11, which is constrained to the interval [pA+pB-1, min(pA,pB)] as derived in Supplementary C.2. For the analyzed MICCAI papers, reported accuracies are often around 0.9, so the lower bound pA+pB-1 is typically 0.79-0.89. The externally estimated median congruence (0.67) and Q1 (0.47) fall below this lower bound, and the procedure replaces them with the lower bound itself, which is the maximal-disagreement endpoint and therefore maximizes the resulting false-claim probability. The sensitivity analysis in Figure 5 uses Q3=0.83, which is still below the upper bound min(pA,pB) for many high-accuracy papers; the upper half of the feasible interval (e.g., congruence values 0.9-0.99) is never explored. As a concrete illustration, for the median classification paper (delta=0.01, n=500, accuracies about 0.91 vs. 0.90), the upper endpoint yields a false-claim probability of roughly 1%, while the lower endpoint yields roughly 30%, so whether such a paper is counted as a false claim depends entirely on the clamping choice. The paper should either report results across the full feasible congruence interval for each paper, or provide evidence that near-maximal disagreement is realistic for the top-two methods in MICCAI 2023 papers.
- [Figure 5 and Methods, segmentation congruence] For segmentation, the congruence parameter rAB is varied only up to the 75th percentile of the MSD-derived correlation (Q3=0.82), while the endpoint rAB=1 (or values close to 1) is never considered. Since the segmentation false-claim formula in Supplementary C.3 uses (sA^2+sB^2-2 sA sB rAB)^(-1/2) in the Student t-statistic, higher positive correlation can dramatically reduce the estimated probability of a false claim, particularly at the small test set sizes (median 62) common in these papers. The current sensitivity analysis therefore does not bound the full range of possible prevalence estimates. At minimum, the paper should present results for rAB over the entire feasible range and clarify whether the external MSD correlation estimates are representative of the correlation between the top-two methods in MICCAI 2023 papers, which may be more aligned than the median of all pairwise method correlations.
- [Data availability] The extracted dataset from 347 papers is not publicly released; the authors state that it 'can be made available from the corresponding author upon reasonable request' due to concerns about individual author reputation. Because the headline percentages (86% and 53%) are computed from this specific extraction and depend on the congruence handling, the current manuscript does not allow independent verification of the central claims. A de-identified version with only test set sizes, reported metric values, and method rankings for the included papers would preserve privacy while enabling full reproducibility of the prevalence analysis.
minor comments (4)
- [Methods, Data analysis subset] In the sentence listing exclusion criteria, 'we had to exclude papers which only only cross-validation' contains a duplicated 'only'; please correct.
- [Supplementary C.2] The phrase 'The integral 1 is not tractable analytically' should refer to 'Equation (1)' or 'the integral in Equation (1)' for clarity.
- [Figure 6] Some Q3 values in Figure 6 appear non-monotonic (for example, the cell for test set size 500-1,000 and delta 0.005-0.01 lists Q3=0.38 while smaller-delta cells show higher Q3 values); please proofread the table for internal consistency.
- [Main text, Discussion] The paper states that 'our sensitivity analyses confirm that the methodology remains robust across varying parameter assumptions,' but the sensitivity analysis only varies congruence up to Q3 (0.83 for classification, 0.82 for segmentation). Please temper this claim or extend the sensitivity analysis to the full feasible range.
Circularity Check
No significant circularity: the false-claim probabilities are computed from first-principles Bayesian formulas with externally estimated congruence parameters, not from fitted or self-defined targets.
full rationale
The derivation chain is self-contained. The probability-of-false-claims formulas are derived from standard Bayesian models: a Dirichlet-multinomial conjugate model for classification and a normal/Student model for segmentation, both with stated non-informative priors. The reported accuracies, test set sizes, and imputed standard deviations are inputs from the MICCAI 2023 papers, while model congruence values are estimated from external datasets (MSD for segmentation, the authors' private dataset for classification) and are not fitted to the MICCAI results being evaluated. The sensitivity analysis varies congruence between Q1 and Q3 values and separately tests SD imputation, so the headline percentages are not definitionally equal to a single assumed input. The clamping of congruence to its feasible lower bound for high-accuracy classification papers is a modeling and robustness concern that could inflate the estimated false-claim probabilities, but it is not circular: the reported probability is a derived quantity, not a restatement of the congruence assumption. The self-citations used for the private dataset and SD imputation provide empirical estimates that are externally reproducible or falsifiable rather than being used as an unverified uniqueness theorem or ansatz. The paper openly states the limitation that true congruence parameters were unavailable. Thus no circular step meeting the specified evidentiary standard is present.
Assumptions & free parameters
free parameters (3)
- classification model congruence p11 =
median 0.67, Q1 0.47, Q3 0.83, clamped to feasible bounds
- segmentation model congruence rAB =
median 0.67, Q1 0.44, Q3 0.82
- imputed standard deviations sA, sB for segmentation =
imputed from DSC means via model in Christodoulou et al. 2024
assumptions (4)
- standard math Uniform Dirichlet prior over multinomial proportions for classification
- standard math Non-informative joint prior p(mu, sigma^2) proportional to sigma^-2 for segmentation
- domain assumption Per-image DSC values for the two methods are jointly normally distributed
- domain assumption Reported mean performances are treated as unbiased sample means with the given test set size
Cite this review
Pith. "Pith review of False Promises in Medical Imaging AI? Assessing Validity of Outperformance Claims." pith.science (2026). https://pith.science/paper/PHYULR4W
@misc{pith2026250504720,
author = {Pith},
title = {Pith review of: False Promises in Medical Imaging AI? Assessing Validity of Outperformance Claims},
year = {2026},
howpublished = {\url{https://pith.science/paper/PHYULR4W}},
note = {Machine review of arXiv:2505.04720}
}
read the original abstract
Performance comparisons are fundamental in medical imaging Artificial Intelligence (AI) research, often driving claims of superiority based on relative improvements in common performance metrics. However, such claims frequently rely solely on empirical mean performance. In this paper, we investigate whether newly proposed methods genuinely outperform the state of the art by analyzing a representative cohort of medical imaging papers. We quantify the probability of false claims based on a Bayesian approach that leverages reported results alongside empirically estimated model congruence to estimate whether the relative ranking of methods is likely to have occurred by chance. According to our results, the majority (>80%) of papers claims outperformance when introducing a new method. Our analysis further revealed a high probability (>5%) of false outperformance claims in 86% of classification papers and 53% of segmentation papers. These findings highlight a critical flaw in current benchmarking practices: claims of outperformance in medical imaging AI are frequently unsubstantiated, posing a risk of misdirecting future research efforts.
Figures
Figures from the paper (3 more)
Reference graph
Works this paper leans on
-
[1]
& Narayanan, A
Kapoor, S. & Narayanan, A. Leakage and the reproducibility crisis in machine-learning-based science. Patterns 4 (2023)
2023
-
[2]
Cross-validation failure: Small sample sizes lead to large error bars
Varoquaux, G. Cross-validation failure: Small sample sizes lead to large error bars. Neuroimage 180, 68–77 (2018)
work page 2018
-
[3]
Ioannidis, J. P. Why most published research findings are false. PLoS medicine 2, e124 (2005)
work page 2005
-
[4]
Antonelli, M. et al. The medical segmentation decathlon. Nature communications 13, 4128 (2022)
work page 2022
-
[5]
Godau, P., Srivastava, A., Adler, T. & Maier-Hein, L. Beyond knowledge silos: Task fingerprinting for democrati- zation of medical imaging ai. arXiv preprint arXiv:2412.08763 (2024)
-
[6]
Medical meta learner (version 1.0.0) [computer software]
Godau, P. Medical meta learner (version 1.0.0) [computer software]. https://github.com/IMSY-DKFZ/mml (2024)
work page 2024
-
[7]
Analysis of images to detect abnormalities in endoscopy (aida-e) challenge. https://aidasub-clebarrett. grand-challenge.org/home/ (2024)
work page 2024
-
[8]
Yang, X. et al. Covid-ct-dataset: a ct scan dataset about covid-19. arXiv preprint arXiv:2003.13865 (2020)
arXiv 2020
Show all 25 references
-
[9]
Kawahara, J. et al. Seven-point checklist and skin lesion classification using multitask multimodal neural nets. IEEE journal of biomedical and health informatics 23, 538–546 (2018)
2018
-
[10]
SIIM-ISIC 2020 Challenge Dataset
International Skin Imaging Collaboration. SIIM-ISIC 2020 Challenge Dataset. https://doi.org/10.34970/ 2020-ds01 (2020). Accessed: 2025-04-10
2020
-
[11]
Rotemberg, V. et al. A patient-centric dataset of images and metadata for identifying melanomas using clinical context. Scientific data 8, 1–20 (2021)
2021
-
[12]
Borgli, H. et al. Hyperkvasir, a comprehensive multi-class image and video dataset for gastrointestinal endoscopy. Scientific data 7, 1–14 (2020)
2020
-
[13]
Al-Dhabyani, W. et al. Dataset of breast ultrasound images. Data in brief 28, 104863 (2020)
2020
-
[14]
Rajpurkar, P. et al. Mura dataset: Towards radiologist-level abnormality detection in musculoskeletal radiographs. In Medical imaging with deep learning (2017)
2017
-
[15]
Irvin, J. et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI conference on artificial intelligence , vol. 33, 590–597 (2019)
2019
-
[16]
Yu, W. et al. Metaformer baselines for vision. IEEE Transactions on Pattern Analysis and Machine Intelligence 46, 896–912 (2023)
2023
-
[17]
Wu, K. et al. Tinyvit: Fast pretraining distillation for small vision transformers. In European conference on computer vision, 252–270 (Springer, 2022)
2022
-
[18]
Liu, Z. et al. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision , 10012–10022 (2021)
2021
-
[19]
Dai, Z. et al. Coatnet: Marrying convolution and attention for all data sizes. Advances in neural information processing systems 34, 3965–3977 (2021)
2021
-
[20]
Yalniz, I. Z. et al. Billion-scale semi-supervised learning for image classification. arXiv preprint arXiv:1905.00546 (2019)
2019 arXiv
-
[21]
Christodoulou, E. et al. Confidence intervals uncovered: Are we ready for real-world medical imaging ai? In Medical Image Computing and Computer Assisted Intervention – MICCAI 2024 , 124–132 (Springer Nature Switzerland, Cham, 2024)
2024
-
[22]
France 2030
Wasef Hattab, M. A derivation of prediction intervals for gamma regression. Journal of Statistical Computation and Simulation 86, 3512–3526 (2016). Acknowledgments This publication received funding from the European Research Council (ERC) under the European Union’s Horizon 202...
2016
-
[23]
& Grandvalet, Y
Bengio, Y. & Grandvalet, Y. No unbiased estimator of the variance of k-fold cross-validation. Journal of machine learning research 5, 1089–1105 (2004)
2004
-
[24]
& Lawson, A
Lesaffre, E. & Lawson, A. B. Bayesian biostatistics (John Wiley & Sons, 2012)
2012
-
[25]
B., Stern, H
Gelman, A., Carlin, J. B., Stern, H. S. & Rubin, D. B. Bayesian data analysis, Second Edition (Chapman and Hall/CRC, 2004). 22
2004
Reviewed August 15, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.