Pith. sign in

REVIEW 5 major objections 5 minor 62 references

Confidence-Uncertainty Boundary Calibration for Bayesian Deep Learning in Medical Image Analysis

T0 review · 5 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read The paper argues that well-calibrated medical AI must place predictive entropy on an extreme boundary—low for confident predictions, high for uncertain ones—and builds a loss and a dual-temperature post-hoc scaling to force that alignment.

desk verdict A sensible, clearly written extension of the AvU/TS line with a real but unproven normative assumption: the target curve is stipulated, not derived, so the headline BCCE gains are partly fitted. Referee it, but expect major revision. read the letter →

arxiv 2602.11973 v2 pith:B5J5HKAJ submitted 2026-02-12 cs.CV cs.LG

Hua Xu , Julián D. Arias-Londoño , Juan I. Godino-Llorente This is my paper · ORCID
classification cs.CVcs.LG
keywords Bayesiandeeplearninguncertaintycalibrationconfidence-uncertaintyboundarymedicalimageanalysisout-of-distributiondetectionvariationalinferencetemperaturescalingAvUmetric
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

This paper tries to establish that uncertainty calibration in Bayesian deep learning for medical imaging can be reduced to a geometric constraint: every prediction should sit on a single 'confidence-uncertainty boundary curve' defined by the theoretical entropy bounds. On this curve, high-confidence predictions are assigned the minimum possible entropy and low-confidence predictions the maximum possible entropy. The authors derive this curve from entropy bounds for a fixed confidence, then build a training loss (CUB-Loss) that measures each sample's distance to the curve and a post-hoc scaling method (DTS) with two temperatures that pushes points toward the curve. They report large improvements in accuracy-uncertainty alignment (AvU 0.95, up from 0.43) and a 56.8% reduction in their own boundary calibration error, along with best near-OOD detection among the methods compared, on pneumonia screening, diabetic retinopathy, and skin lesion tasks.

What carries the argument

The Confidence-Uncertainty Boundary Curve (CUBC): for a given max-probability confidence p̂, predictive entropy U(p) is bounded between Umin(p̂) (residual mass concentrated in one alternative class) and Umax(p̂) (residual mass spread uniformly over all other classes). The paper collapses this feasible band into a single target Uideal(p̂) = Umin if p̂>γ, Umax if p̂≤γ. CUB-Loss measures geometric deviation from this target, and Dual Temperature Scaling (DTS) applies T_high to sharpen high-confidence/low-uncertainty samples and T_low to soften low-confidence/high-uncertainty samples, with temperatures chosen on a validation set to minimize BCCE.

What would settle it

Train a plain VI BNN on the pneumonia dataset, group correct and incorrect test predictions by confidence bin, and compare the average entropy to Umin/Umax predicted by the paper's CUBC; if correct high-confidence predictions have mean entropy far above Umin (or incorrect low-confidence predictions far below Umax), the boundary target is not the natural operating point. Alternatively, compare NLL or Brier score of the CUB-Loss+DTS pipeline against standard temperature scaling on the same validation split; if standard TS attains better NLL while having worse BCCE, the BCCE metric is not decisio

Watch

Extended reading notes

Core claim

The central claim is that the confidence-uncertainty plane has a single ideal curve—minimum entropy Umin for confidence above a threshold γ and maximum entropy Umax below it—and that calibration is achieved by a distance-based loss to that curve during training plus a post-hoc two-temperature scaling at inference. The paper derives Umin and Umax from entropy bounds given a fixed confidence, defines BCCE to measure deviation from the curve, and shows the combination improves AvU and OOD AUROC on three medical datasets.

Load-bearing premise

The paper assumes that a perfectly calibrated model should have the lowest possible entropy for every high-confidence prediction and the highest possible entropy for every low-confidence prediction; this is stipulated, not derived, and if the true ideal lies inside the feasible band, the loss and the metric actively push the model away from good calibration.

Editorial extensions

If this is right

  • Calibration is reframed as a distance to a curve rather than a count of categorical states, so the loss gradient is proportional to the severity of miscalibration; this should make training more stable and sample-efficient than count-based AvU-style losses.
  • Two-temperature scaling sharpens confident predictions and softens uncertain ones simultaneously without changing the rank order of logits, so classification accuracy is preserved while uncertainty is recalibrated.
  • The same BCCE objective guides both training-time regularization and post-hoc tuning, giving the pipeline a single principled target from end to end.
  • Reported gains across X-ray, fundus, and dermoscopy images—including data-scarce and class-imbalanced settings—suggest the method transfers across modalities and architectures.
  • Near-OOD detection results (AUROC 74.2%) outperform the compared output-based and feature-based baselines, indicating that calibrated uncertainty can flag out-of-scope chest X-rays for review.

Reading between the lines

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

  • Because Uideal is stipulated rather than derived from decision theory, the strong empirical gains should be read as fit to that specific curve; a truly decision-relevant calibration target would be validated against expected loss (e.g., Brier score or NLL) on the same tasks.
  • The near-OOD AUROC gap between the proposed uncertainty (74.2%) and the best feature-based method (65.0%) is large; an independent replication on a separate near-OOD test set would clarify whether the advantage comes from the calibration method or from the particular architecture and backbone.
  • The framework could transfer to other closed-set confidence problems (segmentation, regression) wherever an analogous entropy bound can be derived; a natural test is to replace predictive entropy with another dispersion measure and check whether the boundary-curve form still holds.
  • If the CUBC target is wrong for some realistic regime (e.g., a correct but genuinely ambiguous low-confidence prediction), the loss will push that sample's entropy to the maximum, potentially overstating uncertainty and triggering unnecessary expert review—costly in a clinical workflow.
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

5 major / 5 minor

Summary. The paper proposes a Bayesian deep learning framework for medical image classification that couples predictive confidence and uncertainty via a Confidence-Uncertainty Boundary Curve (CUBC). The target curve Uideal in Eq. (15) assigns the minimum feasible entropy to high-confidence predictions and the maximum feasible entropy to low-confidence predictions. A training-time CUB-Loss (Eq. 17-18) penalizes deviations from this curve, and a post-hoc Dual Temperature Scaling (DTS, Eq. 24) is optimized with a newly proposed Boundary Curve Calibration Error (BCCE, Eq. 21) to further align outputs with Uideal. The method is evaluated on three medical imaging tasks: pneumonia screening from chest X-rays, diabetic retinopathy detection from fundus photos, and skin lesion classification from dermoscopic images. Reported results include improved AvU and BCCE, plus OOD detection on a Near-OOD chest X-ray set.

Significance. If the central claim were fully supported, the paper would offer an interesting geometric perspective on uncertainty calibration for Bayesian deep learning, with practical appeal for medical decision support. The authors provide public code, detailed experimental protocols, and use multiple real-world medical datasets. The derivations of Umin and Umax are mathematically correct, and the proposed CUB-Loss has a plausible intuition. However, the current evidence is incomplete because the ideal curve is stipulated rather than derived or independently validated, and key implementation details are missing. The reported BCCE improvements are largely fitting the model to the same target used in the metric, so external validation is essential before the method's clinical value can be accepted.

major comments (5)
  1. [§3.2.3, Eq. (15)] The load-bearing assumption is that perfect calibration is equivalent to U(p) = Uideal(phat), with all high-confidence predictions at Umin and all low-confidence predictions at Umax. This is stipulated, not derived from any standard calibration principle. Standard confidence calibration (Guo et al., 2017) constrains only the confidence-accuracy relationship, and for K>2 the entropy of a calibrated model can legitimately lie in the interior of [Umin,Umax]. Because BCCE (Eq. 21), CUB-Loss (Eq. 18), and DTS (Eq. 25) all optimize toward this same Uideal, the reported BCCE reductions and improved alignment in Figure 3 are tautological unless Uideal is independently justified. I request either a first-principles derivation of Uideal, or a demonstration that boundary-aligned uncertainty improves decision-relevant calibration as measured by external metrics such as ECE, NLL, or a clearly specifi
  2. [§3.3.2, Eq. (17)] The CUB-Loss is not implementable as written. The paper says 'a linear normalization is applied' to map δi to [0,1], but the normalization formula and the constants are never specified. Eq. (17) defines LCUB in terms of the normalized deviation δ̃i, yet no definition of δ̃i is given. Without the exact normalization, no reader can reproduce the loss. Please provide the full formula, including the min/max values used for normalization, and describe how the normalization is computed per region and per batch.
  3. [§3.3.3 and §4.2.2] The joint optimization strategy describes β as 'dynamically adjusted' so that β·LCUB stays on the same order as the ELBO terms, but all experiments fix β at constant values: 0.1 for pneumonia, 1.2 for diabetic retinopathy, and 0.6 for skin lesions. The procedure for selecting these constants is not described, and no sensitivity analysis is provided. Since CUB-Loss is the main training-time contribution, the effect of β on accuracy and uncertainty calibration should be reported. Please clarify whether β is dynamic or fixed, and justify the chosen values.
  4. [Table 5] In Table 5, the rows 'LCUB (confidence)' and 'LCUB (uncertainty)' report identical values (AUROC 71.1, AUPR 73.7) for the Near-OOD test set. This is suspicious and likely a typographical error, since confidence and uncertainty are different scoring functions and should not give exactly identical results. The table also does not state how the scoring functions are defined for these rows (e.g., softmax confidence vs. predictive entropy) or how thresholds are chosen. Please correct the table and specify the scoring protocol for both confidence-based and uncertainty-based OOD detection.
  5. [§2.2, Eq. (1) and results tables] The AvU metric in Eq. (1) depends on an uncertainty threshold that separates 'certain' from 'uncertain' states, but the paper never specifies this threshold in the experiments. The text mentions γ=0.9 and η=0.325 in the DTS setting, but it is not clear whether the same threshold is used to compute AvU for all baselines and all tables. Without a precise definition, the AvU values in Tables 4, 7, and 10 are not reproducible, and comparisons across methods may be misleading. Please state the threshold(s) used for AvU and report sensitivity to that threshold.
minor comments (5)
  1. [§3.2.2, Eq. (11)] The summation over j in Eq. (11) is notationally redundant since the term is constant; consider simplifying to a single expression after the summation. Also, Umax for K=1 is not defined; the paper could mention that the derivation assumes K≥2.
  2. [§3.4.1, Eq. (21)] BCCE depends on the number of bins M and the binning scheme, but M is never specified in the experiments. Please report the value of M and whether the results are robust to changes in M.
  3. [§3.4.2, Eq. (24)] The DTS condition for samples in the intermediate confidence region uses the uncertainty threshold η, but the relationship between η and the CUBC-derived thresholds γlow and γhigh is only briefly described. A more formal specification of which samples enter Step 1 vs. Step 2 would improve clarity.
  4. [§4.2.2] The text states γ=0.9 'corresponds to a theoretical uncertainty lower bound of approximately 0.325.' This value depends on the number of classes K; since the pneumonia experiment has K=3, it is correct, but the same statement is repeated for DR (K=5) and HAM (K=7) without noting that Umin(0.9) is different for those K. Please correct or qualify.
  5. [General] The manuscript contains several grammatical issues and inconsistent capitalization (e.g., 'Bayesian Deep Learning' vs. 'Bayesian deep learning', 'BNNs' vs. 'BNN'). A thorough proofreading pass is recommended.

Circularity Check

2 steps flagged · score 6.0 of 10

The 'ideal' CUBC is stipulated; the loss, the calibration definition, and BCCE all target that same curve, and the headline BCCE reduction is the value of the objective DTS directly minimizes.

  1. self definitional [§3.2.3 Eq. (15); §3.4.1 Eqs. (19) and (21); §3.3.3 Eq. (18)]
    "Given a confidence threshold hyperparameter γ, the ideal uncertainty U ideal is defined as: Uideal(ˆp) = Umin(ˆp) if ˆp > γ; Umax(ˆp) if ˆp ≤ γ (15) ... Based on the CUBC derived above, the perfect uncertainty calibration is defined as: U(p) = U ideal(ˆp), ∀p (19) ... BCCE = Σ |Bm|/N |Ū(Bm) − Ūideal(Bm)| (21)"

    The normative target Uideal in Eq. (15) is stipulated, not derived from a calibration principle. Eq. (19) then defines 'perfect uncertainty calibration' as U(p)=Uideal(ˆp), Eq. (18) trains the model toward that same Uideal via CUB-Loss, and Eq. (21) measures calibration error as distance to Uideal. Thus the training objective, the definition of calibration, and the evaluation metric all point to the same user-chosen curve. Any BCCE improvement is evidence of fit to Uideal, not independent evidence that Uideal is the correct calibration target. The central claim of 'well-calibrated' reduces to 'close to the curve we defined as ideal.'

  2. fitted input called prediction [§3.4.3 Eq. (25); §4.2.3, Figure 3]
    "T ∗ = argmin Thigh,Tlow BCCE(Dval;Thigh,Tlow) (25) ... The optimal temperature parameters (Thigh = 0.874, Tlow = 2.991) were obtained by minimizing BCCE on the validation set using the L-BFGS optimizer. After applying DTS with these optimized temperatures, the empirical curve aligns closely with the theoretical bound, with BCCE reduced from 78.08 to 33.70 (a 56.8% reduction)."

    DTS's two temperatures are selected by directly minimizing BCCE on the validation set (Eq. 25). The reported 56.8% reduction is the difference between the initial and optimized values of that same objective, so it is an optimization artifact rather than a free prediction. Moreover, BCCE itself is defined relative to the stipulated Uideal curve (Eq. 21), so the reduction only demonstrates that DTS can move predictions closer to the curve it was explicitly fitted to match. Reporting this as the main calibration evidence is therefore a fitted quantity presented as a validated improvement.

full rationale

The paper contains no load-bearing self-citation or imported uniqueness theorem; the methodology and comparisons are mostly self-contained. However, the central calibration claim is partially circular. The CUBC target is defined, not derived: Eq. (15) stipulates Uideal as the extreme boundary, Eq. (19) defines 'perfect calibration' as equality to that curve, and Eq. (21) defines BCCE as distance to that curve. CUB-Loss (Eq. 18) and DTS (Eqs. 24–25) both optimize toward the same Uideal, so the reported BCCE drop (78.08 to 33.70, a 56.8% reduction) is the value of the objective minimized by DTS. AvU and OOD AUROC are partially external, but the AvU threshold is never specified and the CUB-Loss's four regions mirror the AvU state decomposition, so the AvU gains are also a designed consequence of the loss. Accuracy preservation and OOD detection against external labels provide some independent content, which is why the score is 6 rather than higher.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The framework's core is a stipulated ideal curve and hand-set constants. The entropy bounds are externally justified mathematics; the remaining pieces are choices made for this paper. The CUBC has no independent verification, and because it drives both the loss and the metric, it is the main unvalidated premise.

free parameters (6)
  • γ (confidence threshold) = 0.9
    Partitions high/low confidence regimes in CUBC; set by hand, not tuned or justified by data.
  • β (CUB-Loss weight) = 0.1 (pneumonia), 1.2 (retinopathy), 0.6 (skin)
    Controls regularization strength; described as dynamic in §3.3.3 but fixed per task in experiments.
  • η (DTS uncertainty threshold) = 0.325
    Set equal to Umin(0.9); determines the DTS region split.
  • T_high, T_low (DTS temperatures) = 0.874, 2.991 (pneumonia)
    Fitted on validation set by minimizing BCCE.
  • δ̃ normalization constants = unspecified
    The linear normalization of δi in §3.3.2 is not defined; necessary to compute Eq. (17).
  • BCCE bin count M = unspecified
    The number of bins in Eq. (21) is not reported; it changes the metric values.
assumptions (5)
  • standard math The feasible entropy region for fixed confidence is exactly [Umin(p̂), Umax(p̂)] (Eqs. 12, 14)
    Correct consequence of Shannon entropy; used to define CUBC.
  • domain assumption Predictive entropy is the appropriate uncertainty measure for clinical decision support
    The paper states entropy is adopted as the primary metric without comparing alternatives (e.g., mutual information, variance).
  • ad hoc to paper The piecewise ideal curve Uideal in Eq. (15) is the optimal calibration target
    No decision-theoretic or empirical justification for choosing Umin above γ and Umax below γ; the threshold γ is arbitrary.
  • domain assumption Mean-field Gaussian variational posterior is a sufficient approximation for the weights
    Standard in BNNs but known to under-estimate uncertainty; the method's calibration is conditioned on this approximation.
  • domain assumption MC sampling with 5 training / 80 inference samples is sufficient for the predictive distribution
    No convergence analysis; chosen for computational convenience.
invented entities (1)
  • CUBC / Uideal curve
    purpose: Defines the training target for CUB-Loss, the metric BCCE, and the DTS objective; the paper's central construct.
    The curve is a stipulated normative model with no falsifiable handle outside the paper; its validity is assumed throughout.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Confidence-Uncertainty Boundary Calibration for Bayesian Deep Learning in Medical Image Analysis." pith.science (2026). https://pith.science/paper/B5J5HKAJ

@misc{pith2026260211973,
  author       = {Pith},
  title        = {Pith review of: Confidence-Uncertainty Boundary Calibration for Bayesian Deep Learning in Medical Image Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B5J5HKAJ}},
  note         = {Machine review of arXiv:2602.11973}
}
read the original abstract

In critical decision support systems based on medical imaging, the reliability of AI-assisted decision-making is as relevant as predictive accuracy. Although deep learning models have demonstrated significant accuracy, they frequently suffer from miscalibration, manifested as overconfidence in erroneous predictions. To facilitate clinical acceptance, it is imperative that models quantify uncertainty in a manner that correlates with prediction correctness, allowing clinicians to identify unreliable outputs for further review. To address this necessity, this paper proposes a probabilistic optimization framework grounded in Bayesian deep learning. Specifically, the Confidence-Uncertainty Boundary Curve (CUBC) is first explored as an intermediate operational target. Grounded in this target, a novel Confidence-Uncertainty Boundary Loss (CUB-Loss) is proposed to regularize the alignment between prediction confidence and uncertainty estimates during training, imposing penalties on high-certainty errors and low-certainty correct predictions. Upon completion of training optimization, a Boundary Curve Calibration Error (BCCE) metric is further introduced to measure the degree of boundary alignment in the calibrated model. Building on this measurement, a Dual Temperature Scaling (DTS) strategy is devised to perform post-hoc refinement, further adjusting the posterior predictive distribution across different confidence-uncertainty regions. The proposed framework is validated on three distinct medical imaging tasks: automatic screening of pneumonia, diabetic retinopathy detection, and identification of skin lesions. Empirical results demonstrate that the proposed approach improves uncertainty calibration across diverse modalities, maintains robust performance in data-scarce scenarios, and remains effective on severely imbalanced datasets, underscoring its potential for real clinical deployment.

Figures

Figures reproduced from arXiv: 2602.11973 by the authors.

Figure 1
Figure 1. Visualization of the Confidence-Uncertainty Boundary Curve. [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Partitioning of the confidence-uncertainty space by thresholds [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Boundary Curve Calibration Error before DTS (a) and after DTS (b). [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

62 extracted references · 3 linked inside Pith

  1. [1]

    On calibration of modern neural networks

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International Conference on Machine Learning (ICML), pages 1321–1330, 2017

    2017

  2. [2]

    Leveraging uncertainty information from deep neural networks for disease detection.Scientific Reports, 7(1):1–14, 2017

    Christian Leibig, Vaneeda Allken, Murat Seçkin Ayhan, Philipp Berens, and Siegfried Wahl. Leveraging uncertainty information from deep neural networks for disease detection.Scientific Reports, 7(1):1–14, 2017. 21 APREPRINT- FEBRUARY13, 2026

    2017

  3. [3]

    Uncertainty estimation in deep neural networks for dermoscopic image classification

    Marc Combalia, Ferran Hueto, Susu Puig, Josep Malvehy, and Veronica Vilaplana. Uncertainty estimation in deep neural networks for dermoscopic image classification. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 744–745, 2020

    2020

  4. [4]

    David J. C. MacKay.Bayesian Methods for Adaptive Models. PhD thesis, California Institute of Technology, Pasadena, CA, 1992

    1992

  5. [5]

    Neal.Bayesian Learning for Neural Networks, volume 118 ofLecture Notes in Statistics

    Radford M. Neal.Bayesian Learning for Neural Networks, volume 118 ofLecture Notes in Statistics. Springer, New York, 1996

    1996

  6. [6]

    Dropout as a Bayesian approximation: Representing model uncertainty in deep learning

    Yarin Gal and Zoubin Ghahramani. Dropout as a Bayesian approximation: Representing model uncertainty in deep learning. InInternational Conference on Machine Learning (ICML), pages 1050–1059, 2016

    2016

  7. [7]

    Weight uncertainty in neural networks

    Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural networks. InInternational Conference on Machine Learning (ICML), pages 1613–1622, 2015

    2015

  8. [8]

    Practical variational inference for neural networks

    Alex Graves. Practical variational inference for neural networks. InAdvances in Neural Information Processing Systems (NIPS), volume 24, pages 2348–2356, 2011

    2011

Show all 62 references
  1. [9]

    Concrete dropout

    Yarin Gal, Jiri Hron, and Alex Kendall. Concrete dropout. InAdvances in Neural Information Processing Systems (NIPS), volume 30, pages 3581–3590, 2017

    2017

  2. [10]

    Uncertainty estimation in medical image classification: systematic review.JMIR Medical Informatics, 10(8):e36427, 2022

    Alexander Kurz, Katja Hauser, Hendrik Alexander Mehrtens, Eva Krieghoff-Henning, Achim Hekler, Jakob Niko- las Kather, Stefan Fröhling, Christof V on Kalle, Titus Josef Brinker, et al. Uncertainty estimation in medical image classification: systematic review.JMIR Medical Infor...

    2022

  3. [11]

    Can your ai differentiate cats from covid-19? sample efficient uncertainty estimation for deep learning safety

    Ankur Mallick, Chaitanya Dwivedi, Bhavya Kailkhura, Gauri Joshi, and T Yong-Jin Han. Can your ai differentiate cats from covid-19? sample efficient uncertainty estimation for deep learning safety. InICML 2020 Workshop on Uncertainty & Robustness in Deep Learning, 2020

    2020

  4. [12]

    Improving model calibration with accuracy versus uncertainty optimiza- tion

    Ranganath Krishnan and Omesh Tickoo. Improving model calibration with accuracy versus uncertainty optimiza- tion. InAdvances in Neural Information Processing Systems, volume 33, pages 18237–18248, 2020

    2020

  5. [13]

    Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases

    Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In2017 IEEE Conference on Computer Vision and...

    2017

  6. [14]

    Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison

    Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilcus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. InProceedings of the AAAI C...

    2019

  7. [15]

    Deep transfer learning models for medical diabetic retinopathy detection.Acta Informatica Medica, 27(5):327, 2019

    Nour Khalifa, Mohamed Loey, Mohamed Taha, and Hamed Mohamed. Deep transfer learning models for medical diabetic retinopathy detection.Acta Informatica Medica, 27(5):327, 2019

    2019

  8. [16]

    The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions.Scientific Data, 5(1):180161, 2018

    Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions.Scientific Data, 5(1):180161, 2018

    2018

  9. [17]

    Skin lesion classification and detection using machine learning techniques: A systematic review.Diagnostics, 13(19):3147, 2023

    Taye Girma Debelee. Skin lesion classification and detection using machine learning techniques: A systematic review.Diagnostics, 13(19):3147, 2023

    2023

  10. [18]

    Unreliable Monte Carlo dropout uncertainty estimation

    Aslak Djupskås, Signe Riemer-Sørensen, and Alexander Johannes Stasik. Unreliable Monte Carlo dropout uncertainty estimation. InProceedings of the Northern Lights Deep Learning Conference (NLDL), Proceedings of Machine Learning Research. PMLR, 2026

    2026

  11. [19]

    Variational dropout and the local reparameterization trick

    Diederik P Kingma, Tim Salimans, and Max Welling. Variational dropout and the local reparameterization trick. InAdvances in Neural Information Processing Systems, volume 28, pages 2575–2583, 2015

    2015

  12. [20]

    Well-calibrated model uncertainty with temperature scaling for dropout variational inference

    Max-Heinrich Laves, Sontje Ihler, Karl-Philipp Kortmann, and Tobias Ortmaier. Well-calibrated model uncertainty with temperature scaling for dropout variational inference. In4th Workshop on Bayesian Deep Learning at NeurIPS, 2019

    2019

  13. [21]

    Calibration of model uncertainty for dropout variational inference.arXiv preprint arXiv:2006.11584, 2020

    Max-Heinrich Laves, Sontje Ihler, Karl-Philipp Kortmann, and Tobias Ortmaier. Calibration of model uncertainty for dropout variational inference.arXiv preprint arXiv:2006.11584, 2020

    2006 arXiv

  14. [22]

    Staib, and John A

    Tal Zeevi, Ravid Shwartz-Ziv, Yann LeCun, Lawrence H. Staib, and John A. Onofrey. Rate-In: Information-driven adaptive dropout rates for improved inference-time uncertainty estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), ...

    2025

  15. [23]

    Dropout injection at test time for post hoc uncertainty quantification in neural networks.Information Sciences, 645:119356, 2023

    Emanuele Ledda, Giorgio Fumera, and Fabio Roli. Dropout injection at test time for post hoc uncertainty quantification in neural networks.Information Sciences, 645:119356, 2023. 22 APREPRINT- FEBRUARY13, 2026

    2023

  16. [24]

    Can you trust your model’s uncertainty? Evaluating predictive uncertainty under dataset shift

    Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model’s uncertainty? Evaluating predictive uncertainty under dataset shift. InAdvances in Neural Information Proces...

    2019

  17. [25]

    ’in-between’uncertainty in bayesian neural networks.arXiv preprint arXiv:1906.11537, 2019

    Andrew YK Foong, Yingzhen Li, José Miguel Hernández-Lobato, and Richard E Turner. ’in-between’uncertainty in bayesian neural networks.arXiv preprint arXiv:1906.11537, 2019

    1906 arXiv

  18. [26]

    Radial Bayesian neural networks: Beyond discrete support in large-scale Bayesian deep learning

    Sebastian Farquhar, Michael Osborne, and Yarin Gal. Radial Bayesian neural networks: Beyond discrete support in large-scale Bayesian deep learning. InProceedings of the 23rd International Conference on Artificial Intelligence and Statistics (AISTATS), volume 108 ofProceedings ...

    2020

  19. [27]

    How good is the bayes posterior in deep neural networks really? InInternational Conference on Machine Learning (ICML), pages 10248–10259, 2020

    Florian Wenzel, Kevin Roth, Bastiaan S Veeling, Jakub Swiatkowski, Linh Tran, Stephan Mandt, Jasper Snoek, and Sebastian Nowozin. How good is the bayes posterior in deep neural networks really? InInternational Conference on Machine Learning (ICML), pages 10248–10259, 2020

    2020

  20. [28]

    Soft calibration objectives for neural networks

    Archit Karandikar, Nicholas Cain, Dustin Tran, Balaji Lakshminarayanan, Jonathon Shlens, Michael C Mozer, and Becca Weber. Soft calibration objectives for neural networks. InAdvances in Neural Information Processing Systems (NeurIPS), volume 34, pages 29768–29779, 2021

    2021

  21. [29]

    Well-calibrated regression uncer- tainty in medical imaging with deep learning

    Max-Heinrich Laves, Sontje Ihler, Karl-Ingo Kortmann, and Tobias Ortmaier. Well-calibrated regression uncer- tainty in medical imaging with deep learning. InMedical Imaging with Deep Learning (MIDL), pages 393–412, 2020

    2020

  22. [30]

    David J. C. MacKay. A practical Bayesian framework for backpropagation networks.Neural Computation, 4(3):448–472, 1992

    1992

  23. [31]

    Hinton and Drew van Camp

    Geoffrey E. Hinton and Drew van Camp. Keeping the neural networks simple by minimizing the description length of the weights. InProceedings of the 6th Annual Conference on Computational Learning Theory (COLT), pages 5–13, 1993

    1993

  24. [32]

    Simple and scalable predictive uncertainty estimation using deep ensembles

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. InAdvances in Neural Information Processing Systems (NeurIPS), volume 30, pages 6402–6413, 2017

    2017

  25. [33]

    Claude E. Shannon. A mathematical theory of communication.The Bell System Technical Journal, 27(3):379–423, 1948

    1948

  26. [34]

    A baseline for detecting misclassified and out-of-distribution examples in neural networks

    Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. InProceedings of the International Conference on Learning Representations (ICLR), 2017

    2017

  27. [35]

    The role of chest radiography in confirming covid-19 pneumonia.BMJ, 370:m2426, 2020

    Joanne Cleverley, James Piper, and Melvyn M Jones. The role of chest radiography in confirming covid-19 pneumonia.BMJ, 370:m2426, 2020

    2020

  28. [36]

    Automated detection of covid-19 cases using deep neural networks with x-ray images.Computers in Biology and Medicine, 121:103792, 2020

    Tulin Ozturk, Muhammed Talo, Eylul Azra Yildirim, Ulas Baran Baloglu, Ozal Yildirim, and U Rajendra Acharya. Automated detection of covid-19 cases using deep neural networks with x-ray images.Computers in Biology and Medicine, 121:103792, 2020

    2020

  29. [37]

    Distance-based detection of out-of-distribution silent failures for covid-19 lung lesion segmentation.Medical Image Analysis, 82:102596, 2022

    Camila González, Karol Gotkowski, Moritz Fuchs, Andreas Bucher, Armin Dadras, Ricarda Fischbach, Isabel Jas- min Kaltenborn, and Anirban Mukhopadhyay. Distance-based detection of out-of-distribution silent failures for covid-19 lung lesion segmentation.Medical Image Analysis, ...

    2022

  30. [38]

    MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reports.Scientific Data, 6:317, 2019

    Alistair EW Johnson, Tom J Pollard, et al. MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reports.Scientific Data, 6:317, 2019

    2019

  31. [39]

    Covid-net: a tailored deep convolutional neural network design for detection of covid-19 cases from chest x-ray images.Scientific Reports, 10(1):19549, 2020

    Linda Wang, Zhong Qiu Lin, and Alexander Wong. Covid-net: a tailored deep convolutional neural network design for detection of covid-19 cases from chest x-ray images.Scientific Reports, 10(1):19549, 2020

    2020

  32. [40]

    Covid-19 chest x-ray dataset initiative

    Joseph Paul Cohen, Paul Morrison, Lan Dao, Karsten Roth, Tim Q Duong, and Marzyeh Ghassemi. Covid-19 chest x-ray dataset initiative. https://github.com/ieee8023/covid-chestxray-dataset , 2020. Dataset

    2020

  33. [41]

    Actualmed covid-19 chest x-ray dataset initiative

    Agam Chung et al. Actualmed covid-19 chest x-ray dataset initiative. https://github.com/agchung/Actua lmed-COVID-chestxray-dataset, 2020. Dataset

    2020

  34. [42]

    Muhammad E. H. Chowdhury, Tawsif R. Rahman, Amith Khandakar, Rashid Mazhar, et al. Covid-19 radiography database – version 3. https://www.kaggle.com/datasets/tawsifurrahman/covid19-radiography-d atabase, 2020. Dataset

    2020

  35. [43]

    Rsna pneumonia detection challenge dataset

    RSNA and Society of Thoracic Radiology. Rsna pneumonia detection challenge dataset. https://www.kaggle .com/c/rsna-pneumonia-detection-challenge, 2018. Dataset. 23 APREPRINT- FEBRUARY13, 2026

    2018

  36. [44]

    Saborit, Joaquim Á

    María de la Iglesia Vayá, José M. Saborit, Joaquim Á. Montell, Antonio Pertusa, Aurelia Bustos, Miguel Cazorla, Joaquin Galant, Xavier Barber, Domingo Orozco-Beltrán, Francisco García-García, Marisa Caparrós, Germán González, and Jose M. Salinas. Bimcv-covid19+: A large annota...

    2020

  37. [45]

    Stony brook university covid-19 positive cases (covid-19-ny-sbu).https://doi.org/10.7937/TCIA.BBAG-2923, 2021

    Jacob Saltz, Michael Saltz, Pranav Prasanna, Richard Moffitt, John Hajagos, et al. Stony brook university covid-19 positive cases (covid-19-ny-sbu).https://doi.org/10.7937/TCIA.BBAG-2923, 2021. Dataset

    2021 doi

  38. [46]

    Analysis of the clever hans effect in covid-19 detection using chest x-ray images and bayesian deep learning.Biomedical Signal Processing and Control, 90:105831, 2024

    Julián D Arias-Londoño and Juan I Godino-Llorente. Analysis of the clever hans effect in covid-19 detection using chest x-ray images and bayesian deep learning.Biomedical Signal Processing and Control, 90:105831, 2024

    2024

  39. [47]

    OpenOOD v1.5: Enhanced benchmark for out-of-distribution detection.Journal of Data-centric Machine Learning Research, 2, 2024

    Jingyang Zhang, Jingkang Yang, Pengyun Wang, Haoqi Wang, Yueqian Lin, Haoran Zhang, Yiyou Sun, Xuefeng Du, Yixuan Li, Ziwei Liu, Yiran Chen, and Hai Li. OpenOOD v1.5: Enhanced benchmark for out-of-distribution detection.Journal of Data-centric Machine Learning Research, 2, 2024

    2024

  40. [48]

    Terrance DeVries and Graham W. Taylor. Learning confidence for out-of-distribution detection in neural networks. arXiv preprint arXiv:1802.04865, 2018

    2018 arXiv

  41. [49]

    David Macêdo, Tsang Ing Ren, Cleber Zanchettin, Adriano L. I. Oliveira, and Teresa Ludermir. Entropic out-of- distribution detection: Seamless detection of unknown examples.IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 33(6):2350–2364, 2022

    2022

  42. [50]

    On the Use of Mahalanobis Distance for Out-of-distribution Detection with Neural Networks for Medical Imaging

    Harry Anthony and Konstantinos Kamnitsas. On the Use of Mahalanobis Distance for Out-of-distribution Detection with Neural Networks for Medical Imaging. In Carole H. Sudre, Christian F. Baumgartner, Adrian Dalca, Raghav Mehta, Chen Qin, and William M. Wells, editors,Uncertaint...

    2023

  43. [51]

    Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. InInternational Conference on Learning Representations (ICLR), 2018

    2018

  44. [52]

    Kernel PCA for out-of-distribution detection

    Kun Fang, Qinghua Tao, Kexin Lv, Mingzhen He, Xiaolin Huang, and Jie Yang. Kernel PCA for out-of-distribution detection. InAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  45. [53]

    Alshalali, and Zaffar Ahmed Shaikh

    Mohsin Akram, Muhammad Adnan, Syed Farooq Ali, Jameel Ahmad, Amr Yousef, Tagrid Abdullah N. Alshalali, and Zaffar Ahmed Shaikh. Uncertainty-aware diabetic retinopathy detection using deep learning enhanced by bayesian approaches.Scientific Reports, 15(1):1342, 2025

    2025

  46. [54]

    APTOS 2019 blindness detection.https://www.kaggle.com/c/a ptos2019-blindness-detection, 2019

    Asia Pacific Tele-Ophthalmology Society. APTOS 2019 blindness detection.https://www.kaggle.com/c/a ptos2019-blindness-detection, 2019

    2019

  47. [55]

    Diabetic retinopathy 224x224 gaussian filtered

    Sovit Rath. Diabetic retinopathy 224x224 gaussian filtered. https://www.kaggle.com/datasets/sovitrat h/diabetic-retinopathy-224x224-gaussian-filtered, 2019

    2019

  48. [56]

    PhD thesis, University of Cambridge, 2016

    Yarin Gal.Uncertainty in deep learning. PhD thesis, University of Cambridge, 2016

    2016

  49. [57]

    Automatic detection and classification of diabetic retinopathy using the improved pooling function in the convolution neural network.Diagnostics, 13(15):2606, 2023

    Usharani Bhimavarapu, Nalini Chintalapudi, and Gopi Battineni. Automatic detection and classification of diabetic retinopathy using the improved pooling function in the convolution neural network.Diagnostics, 13(15):2606, 2023

    2023

  50. [58]

    Alwakid, W

    G. Alwakid, W. Gouda, M. Humayun, and N.Z. Jhanjhi. Enhancing diabetic retinopathy classification using deep learning.Digital Health, 9:20552076231203676, 2023

    2023

  51. [59]

    Analysis of pre-trained convolutional neural network models in diabetic retinopathy detection through retinal fundus images

    José Escorcia-Gutierrez, Jose Cuello, Carlos Barraza, Margarita Gamarra, Pere Romero-Aroca, Eduardo Caicedo, Aida Valls, and Domenec Puig. Analysis of pre-trained convolutional neural network models in diabetic retinopathy detection through retinal fundus images. In Khalid Sae...

    2022

  52. [60]

    Dynamically weighted balanced loss: Class imbalanced learning and confidence calibration of deep neural networks.Diagnostics, 11(8):1242, 2021

    K Ruwani M Fernando and Chris P Tsokos. Dynamically weighted balanced loss: Class imbalanced learning and confidence calibration of deep neural networks.Diagnostics, 11(8):1242, 2021

    2021

  53. [61]

    Specifying Weight Priors in Bayesian Deep Neural Networks with Empirical Bayes

    Ranganath Krishnan, Mahesh Subedar, and Omesh Tickoo. Specifying Weight Priors in Bayesian Deep Neural Networks with Empirical Bayes. InProceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4477–4484, 2020

    2020

  54. [62]

    Pre-train your loss: Easy bayesian transfer learning with informative priors.Advances in Neural Information Processing Systems, 35:27706–27715, 2022

    Ravid Shwartz-Ziv, Micah Goldblum, Hossein Souri, Sanyam Kapoor, Chen Zhu, Yann LeCun, and Andrew G Wilson. Pre-train your loss: Easy bayesian transfer learning with informative priors.Advances in Neural Information Processing Systems, 35:27706–27715, 2022. 24

    2022

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.