Pith. sign in

REVIEW 3 major objections 6 minor 50 references

Deep Learning-Based BMD Estimation from Radiographs with Conformal Uncertainty Quantification

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Conformal prediction gives knee-X-ray BMD estimates error bars that are statistically guaranteed: 94.8% empirical coverage at nominal 95%, despite moderate point accuracy ($R = 0.68$).

desk verdict A clean, honest proof-of-concept for conformal UQ on BMD from knee X-rays, but the headline coverage guarantee hinges on a patient-level split the paper never confirms. read the letter →

arxiv 2505.22551 v1 pith:6DMQGYKC submitted 2025-05-28 cs.CV stat.AP

Long Hui , Wai Lok Yeung This is my paper
classification cs.CVstat.AP
keywords bonemineraldensityestimationconformalpredictionsplituncertaintyquantificationkneeradiographsEfficientNettest-timeaugmentationosteoporosisscreening
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 tries to establish that opportunistic osteoporosis screening can be built on ordinary knee radiographs: an EfficientNet model predicts femoral-neck BMD from bilateral knee X-rays, and split conformal prediction wraps each prediction in a patient-specific interval with a distribution-free coverage guarantee. The point predictions are moderate — Pearson correlation $R = 0.68$ with traditional test-time augmentation — but the paper's primary claim is about the uncertainty layer: the intervals cover the true BMD at the stated rate (94.8% empirical coverage for nominal 95%), and interval width rises with prediction error, so the model signals when it is out of its depth. The payoff, if the claim holds, is a triage tool: routine X-rays already taken for other reasons could flag patients who need a confirmatory DXA scan, while the conformal intervals decide who gets flagged. The paper is explicit that this is a proof of concept, not a clinical device, because of the anatomical mismatch between knee radiographs and the DXA reference site.

What carries the argument

The load-bearing mechanism is split conformal prediction: a calibration set kept separate from training supplies absolute-residual non-conformity scores $s_j = |Y_j - f_\theta(X_j)|$; the $(1-\alpha)(N_{\text{calib}}+1)$-th smallest such score becomes $q_{1-\alpha}$, and every test prediction receives the symmetric interval $[f_\theta(X_{\text{test}}) - q_{1-\alpha},\, f_\theta(X_{\text{test}}) + q_{1-\alpha}]$. This single quantile construction converts the regressor's raw residuals into a distribution-free marginal coverage guarantee, the property the clinical framing depends on. The base regressor is an ImageNet-pretrained EfficientNetV2-M with a regression head of one fully connected layer (768 units), dropout at rate 0.4, and a single output unit, trained with Huber loss ($\delta = 0.5$) on 384×384 knee radiographs from the OAI dataset with DXA-derived femoral-neck BMD as the target. Two test-time augmentation strategies are compared inside the framework: averaging augmented predictions before conformalizing, versus treating each augmented sample as its own conformal instance.

What would settle it

Re-run the calibration/test construction with a strict patient-level split, so every image of a given subject stays in one partition, and recompute empirical coverage on the test set; if coverage drops materially below nominal 95%, the exchangeability assumption was violated and the reported 94.8% is not a valid demonstration of the conformal guarantee. A complementary probe: hold out an entire imaging site or acquisition protocol and check whether coverage collapses, which would show that the guarantee does not survive distribution shift.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that split conformal prediction can be bolted onto a radiograph-based BMD regressor and deliver what plain deep learning cannot: a statistically valid statement about each prediction's reliability. Using absolute residuals on a held-out calibration set, the method computes a quantile $q_{1-\alpha}$ from the $(1-\alpha)(N_{\text{calib}}+1)$-th smallest non-conformity score and forms intervals $[\hat{Y} - q_{1-\alpha}, \hat{Y} + q_{1-\alpha}]$ that cover the true BMD with probability at least $1-\alpha$ under exchangeability. Empirically the intervals hit their targets — 94.8% coverage at nominal 95% and 98.7% at nominal 99% for the bilateral-knee model with TTA — and wider intervals track larger prediction errors. The paper also reports a methodological comparison: conventional average-then-conformalize TTA gives better point predictions ($R = 0.68$), while the multi-sample conformalize-each-sample variant yields slightly tighter 95% intervals (radius 0.2776 vs 0.2792 for the combined model) at the same coverage. The authors frame the contribution as a template for trustworthy AI-assisted BMD screening rather than a ready diagnostic, since the knee-to-femoral-neck site mismatch caps the clinical value of the point estimates.

Load-bearing premise

The guarantee stands or falls on whether calibration and test images are truly exchangeable — in plain terms, whether images from the same patient never appear in both the calibration set and the test set; the paper doubles its data by using both knees of each subject but does not state whether the split was done by patient, so unstated same-patient leakage would quietly void the coverage guarantee.

Editorial extensions

If this is right

  • If the coverage claim holds, clinicians can threshold on interval width: a wide interval routes the patient to a confirmatory DXA scan, while a narrow interval supports same-visit reassurance — turning a moderate point predictor into a usable triage signal.
  • The conformal wrapper is model-agnostic: any improved base regressor (better architecture, larger or more diverse training data) inherits the same coverage guarantee from the same calibration set, with no retraining of the calibration step.
  • The empirical numbers — 94.8% coverage at nominal 95% and 98.7% at nominal 99% on the bilateral-knee model with TTA — mean the method's stated uncertainty is honest in aggregate on this test distribution.
  • The multi-sample TTA variant, if it holds up, delivers slightly tighter 95% intervals (radius 0.2776 vs 0.2792 for the combined model) at the same coverage, a direct efficiency gain for clinical use.
  • The authors position the pipeline as a foundation for other targets and sites: the same training-plus-calibration recipe can be reused for T-scores, fracture risk, or different anatomical regions.

Reading between the lines

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

  • A strict patient-level split — never letting the left and right knees of one subject fall on opposite sides of the calibration/test boundary — would settle whether the 94.8% figure is genuinely exchangeable; if the current split leaks same-patient images, coverage on truly patient-independent data could fall below nominal.
  • The same calibration machinery is model-agnostic and modality-agnostic: it could be dropped onto any image-regression task, from hip or chest X-ray BMD to other continuous biomarkers, because conformal prediction only needs a calibration set and a residual score.
  • Marginal coverage is a population average; patients with unusual anatomy or imaging artifacts could still get miscalibrated intervals, which is exactly the case where conditional methods such as Conformalized Quantile Regression (named in the paper as future work) would be the natural next experiment.
  • A deployment-level extension the paper does not develop: combine the point estimate with interval width to assign a WHO screening category (normal, osteopenic, osteoporotic) only when the interval is narrow enough to separate the T-score thresholds, and defer to DXA otherwise.
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

3 major / 6 minor

Summary. This paper proposes an EfficientNetV2-M-based deep learning model to estimate femoral-neck bone mineral density (BMD) from knee radiographs using the OAI dataset, and integrates split conformal prediction to provide patient-specific prediction intervals with claimed marginal coverage guarantees. Two test-time augmentation (TTA) strategies are compared: traditional averaging-then-conformalization and a multi-sample approach that conformalizes each augmented version. Results report a Pearson correlation of 0.68 for the bilateral-knee model with traditional TTA, empirical coverage of 94.8% for nominal 95% intervals, and slightly tighter intervals for the multi-sample TTA approach. The paper is framed as a proof-of-concept for trustworthy uncertainty quantification in opportunistic BMD screening.

Significance. If the coverage guarantee is valid, the paper makes a useful methodological contribution by demonstrating the application of split conformal prediction to a practical medical imaging regression task, with a clear clinical motivation. The point prediction performance is moderate (R=0.68), so the primary value lies in the uncertainty quantification framework rather than in the accuracy of the BMD estimates. The method is simple and reproducible in principle, though no code or data are provided. The comparison of two TTA strategies is interesting, but the claimed advantage of multi-sample TTA is small and not statistically substantiated. The central risk is the exchangeability assumption underpinning the conformal coverage guarantee; if the data split is not patient-level, the headline coverage result is invalid.

major comments (3)
  1. [Section 3.2.1 and 3.3] The dataset split is not described as patient-level. The paper states that left and right knee images were separated from bilateral radiographs, and the regression target is a single femoral-neck BMD value per patient. If the split into training/validation/test/calibration is performed at image level, the same patient's knees can appear in both calibration and test sets, inducing dependence between calibration nonconformity scores and test residuals. This violates the exchangeability assumption stated in Section 3.1.4 and voids the conformal coverage guarantee. The reported empirical coverage of 94.8% in Section 4.4 could then be an artifact of leakage rather than a valid property of the method. Please clarify whether the split was performed at the patient level; if not, re-run the analysis with a patient-level split and report the resulting coverage values and interval widths.
  2. [Section 4.4 and Table 1] The paper claims that the multi-sample TTA approach produces slightly tighter prediction intervals "while maintaining proper coverage," but no empirical coverage values are reported for the multi-sample TTA variant. Table 1 lists only CP radii (interval half-widths) for all methods; Section 4.4 reports coverage only for the combined model with traditional TTA. Without coverage numbers for the multi-sample method, the comparison is incomplete: a tighter interval is only meaningful if coverage is similar to the traditional approach. Please report empirical coverage for all TTA variants at each nominal level (90%, 95%, 99%).
  3. [Section 3.2.4] The multi-sample TTA conformalization procedure is not defined precisely. The text says it "treats each augmented version as a separate sample for conformal prediction," but it does not specify how the conformal quantile is computed from the calibration scores. If each augmented version of a calibration image is treated as an independent calibration sample, these augmented versions are not exchangeable with test samples because they are deterministic transformations of the same image and are dependent across the calibration and test splits. This could itself violate the exchangeability condition. Please describe the exact algorithm for the multi-sample case and justify why the calibration scores and test scores are exchangeable under that procedure.
minor comments (6)
  1. [Section 2, final paragraph] The text states that the work leverages conformal prediction for "combined hand and knee radiographs," but the methods section only describes the bilateral knee dataset and model. Hand data appear only in Table 1 without any description of preprocessing, training, or evaluation. Please either provide the missing methods for the hand model or remove the hand results to avoid an unexplained discrepancy.
  2. [Section 4.4] The empirical coverage of 94.8% for nominal 95% intervals is reported as a single point estimate without the test set size or a standard error. Given the finite-sample variability of coverage estimates, this value may not be statistically distinguishable from 95%. Please report the number of test samples and, if possible, a confidence interval for the coverage.
  3. [Section 4.5] The claim that the model "appropriately expresses higher uncertainty for cases where its predictions are less accurate" is not quantified. Please report the correlation (e.g., Pearson or Spearman) between the absolute prediction error and the interval width, with a significance measure, to support this statement.
  4. [References] References [29] and [30] are both cited in the text as "Liu et al.," but they have different first authors (Ren Wei Liu and Meng-Chen Liu). Please disambiguate the citations to avoid confusion.
  5. [Table 1 caption] The notation "CP 99%↓" appears to denote the conformal prediction radius (half-width of the interval), but the caption does not explicitly define "radius." Please clarify that these columns report the average interval radius, not the coverage rate.
  6. [Section 3.1.2 heading] The heading "BMD Regression T ask" contains a typo; it should read "BMD Regression Task."

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; conformal intervals derive from held-out calibration data and are evaluated on a separate test set.

full rationale

The derivation chain is not circular. The conformal prediction intervals are constructed exactly per the standard split-conformal recipe (Section 3.1.4): nonconformity scores s_j = |Y_j - f_theta(X_j)| are computed on a dedicated calibration set, the quantile q_{1-alpha} is taken from those scores, and the interval [f_theta(X_test) - q_{1-alpha}, f_theta(X_test) + q_{1-alpha}] is applied to held-out test inputs. The reported 94.8% empirical coverage (Section 4.4) is a measurement on the test set, not a quantity fitted to that set; the finite-sample coverage guarantee is imported from external literature (Vovk, Gammerman and Shafer; Shafer and Vovk), and no self-citation is used as load-bearing evidence. The model is trained with Huber loss and early stopping on the validation set; no parameter is fitted to the test-set coverage or to the between-TTA interval-width comparison. The comparison of traditional versus multi-sample TTA is an empirical observation from the same experimental run, which may be statistically fragile given the small differences in Table 1 and the paper's own acknowledgment of augmentation randomness, but fragility is not circularity. The one substantive validity risk, that the dataset construction (Section 3.2.1) separates left and right knee images without stating a patient-level split (Section 3.3), could violate the exchangeability assumption behind the coverage guarantee; this is a correctness and external-validity concern rather than a circular reduction, because the conformal quantile is still computed from a held-out calibration set and not from test labels. Therefore no circular step can be exhibited.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

No new entities are introduced. The main free parameters are standard training hyperparameters. The critical axiom is exchangeability of calibration and test data, which may be violated by the data split.

free parameters (5)
  • Learning rate = 0.0005
    Chosen by hand for AdamW optimizer; affects model predictions and therefore conformal scores.
  • Weight decay = 0.01
    L2 regularization strength, hand-chosen.
  • Huber loss delta = 0.5
    Threshold between quadratic and linear loss, hand-chosen.
  • Dropout rate = 0.4
    Dropout probability in regression head, hand-chosen.
  • Cosine annealing T0 = 10
    Initial restart period, hand-chosen.
assumptions (3)
  • domain assumption Calibration and test data are exchangeable (e.g., drawn i.i.d. from the same distribution)
    Required for conformal coverage guarantee; uncertain due to possible image-level split of bilateral images. See Section 3.1.4.
  • domain assumption The OAI dataset's femoral neck BMD values are a valid ground truth for knee-derived predictions
    The clinical relevance relies on this, though the paper acknowledges anatomical mismatch. See Section 5.4.
  • domain assumption ImageNet pre-training transfers to medical radiographs
    The model's feature extraction depends on transferable features, a common but unproven assumption. See Section 3.2.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Learning-Based BMD Estimation from Radiographs with Conformal Uncertainty Quantification." pith.science (2026). https://pith.science/paper/6DMQGYKC

@misc{pith2026250522551,
  author       = {Pith},
  title        = {Pith review of: Deep Learning-Based BMD Estimation from Radiographs with Conformal Uncertainty Quantification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6DMQGYKC}},
  note         = {Machine review of arXiv:2505.22551}
}
read the original abstract

Limited DXA access hinders osteoporosis screening. This proof-of-concept study proposes using widely available knee X-rays for opportunistic Bone Mineral Density (BMD) estimation via deep learning, emphasizing robust uncertainty quantification essential for clinical use. An EfficientNet model was trained on the OAI dataset to predict BMD from bilateral knee radiographs. Two Test-Time Augmentation (TTA) methods were compared: traditional averaging and a multi-sample approach. Crucially, Split Conformal Prediction was implemented to provide statistically rigorous, patient-specific prediction intervals with guaranteed coverage. Results showed a Pearson correlation of 0.68 (traditional TTA). While traditional TTA yielded better point predictions, the multi-sample approach produced slightly tighter confidence intervals (90%, 95%, 99%) while maintaining coverage. The framework appropriately expressed higher uncertainty for challenging cases. Although anatomical mismatch between knee X-rays and standard DXA limits immediate clinical use, this method establishes a foundation for trustworthy AI-assisted BMD screening using routine radiographs, potentially improving early osteoporosis detection.

Figures

Figures reproduced from arXiv: 2505.22551 by the authors.

Figure 1
Figure 1. Conceptual illustration of a symmet￾ric conformal prediction interval for a regression task. Yˆ test is the model’s point prediction, and q1−α is derived from the calibration set to en￾sure at least 1 − α coverage. The prediction in￾terval [Yˆ test − q1−α, Yˆ test + q1−α] guarantees that P(Ytest ∈ Interval) ≥ 1 − α. For a new test sample Xtest and a desired signif￾icance level α (e.g., α = 0.1 for 90% prediction int… view at source ↗
Figure 2
Figure 2. Example of a knee radiograph from the OAI dataset before preprocessing. (a) Left knee (b) Right knee [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Example of left and right knee radio￾graphs after splitting the original image during pre￾processing. 3.2.2 BMD Regression Model Architecture The core of our BMD estimation system is a deep convolutional neural network (CNN) based on the EfficientNet architecture [35]. Specifically, we em￾ployed the tf_efficientnetv2_m.in21k_ft_in1k model variant, obtained using the timm library and pre-trained on the ImageNet datas… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Architecture of the EfficientNetV2-based [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Scatter plot of true vs. predicted BMD values using the combined model with TTA. Points are color-coded by absolute error, with lighter col￾ors indicating larger prediction errors. The per￾fect prediction line (red dashed) represents where predictions would exactly mat…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 40 canonical work pages

  1. [1]

    McCloskey, Anders Oden, Nicholas C

    Eugene V. McCloskey, Anders Oden, Nicholas C. Harvey, William D. Leslie, Didier Hans, Helena Johansson, Reinhard Barkmann, Stephanie Boutroy, Jacques Brown, Roland Chapurlat, et al. A meta- analysis of trabecular bone score in fracture risk prediction and its relationship to frax. Journal of Bone and Mineral Research, 36(2): 280–291, 2021. doi: 10.1002/jbmr.4172

    work page doi:10.1002/jbmr.4172 2021

  2. [2]

    Osteoporosis in the european union: medical management, epidemiology and economic burden.Archives of Osteoporosis, 8(1):136, 2013

    Emma Hernlund, Axel Svedbom, Markus Ivergård, Juliet Compston, Cyrus Cooper, John Stenmark, Eugene V McCloskey, Bengt Jönsson, and John A Kanis. Osteoporosis in the european union: medical management, epidemiology and economic burden.Archives of Osteoporosis, 8(1):136, 2013. doi: 10.1007/ s11657-013-0136-1

    work page 2013

  3. [3]

    Pathogenesis of glucocorticoid-induced osteoporosis and options for treatment.Nature Reviews Endocrinology, 16(8):437–447, 2020

    Pojchong Chotiyarnwong and Eu- gene V McCloskey. Pathogenesis of glucocorticoid-induced osteoporosis and options for treatment.Nature Reviews Endocrinology, 16(8):437–447, 2020. doi: 10.1038/s41574-020-0341-0

    work page doi:10.1038/s41574-020-0341-0 2020

  4. [4]

    Assessment of osteoporosis at the primary health care level, 2007

    World Health Organization Scientific Group. Assessment of osteoporosis at the primary health care level, 2007. URL https://www.sheffield.ac.uk/FRAX/pdfs/ WHO_Technical_Report.pdf. Technical Report, WHO Collaborating Centre for Metabolic Bone Diseases, University of Sheffield, UK

    work page 2007

  5. [5]

    Dermatologist- level classification of skin cancer with deep neural networks.Nature, 542(7639):115–118,

    Andre Esteva, Brett Kuprel, Roberto A Novoa, Justin Ko, Susan M Swetter, Helen M Blau, and Sebastian Thrun. Dermatologist- level classification of skin cancer with deep neural networks.Nature, 542(7639):115–118,

    work page

  6. [6]

    Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs.JAMA, 316(22):2402–2410,

    Varun Gulshan, Lily Peng, Marc Coram, Martin C Stumpe, Derek Wu, Arunachalam Narayanaswamy, Subhashini Venugopalan, Kasumi Widner, Tom Madams, Jorge Cuadros, et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs.JAMA, 316(22):2402–2410,

    work page

  7. [7]

    Deep learning for osteoporosis classifi- cation using hip radiographs and patient clin- ical covariates.Scientific Reports, 10(1):1534,

    Norio Yamamoto, Shintaro Sukegawa, Akira Kitamura, Ryosuke Goto, Tomoyuki Noda, 9 Keisuke Nakano, Kiyofumi Takabatake, Ho- taka Kawai, Hitoshi Nagatsuka, Keisuke Kawasaki, Yoshihiko Furuki, and Toshifumi Ozaki. Deep learning for osteoporosis classifi- cation using hip radiographs and patient clin- ical covariates.Scientific Reports, 10(1):1534,

    work page

  8. [8]

    Convolutional neural networks for automated fracture detection and localiza- tion on wrist radiographs.Radiology: Ar- tificial Intelligence, 1(1):e180001, 2019

    Yi Liang Thian, Yang Li, Pooja Jagmohan, Daron Sia, Vincent Ern Yao Chan, and Ryan Thin Lim Tan. Convolutional neural networks for automated fracture detection and localiza- tion on wrist radiographs.Radiology: Ar- tificial Intelligence, 1(1):e180001, 2019. doi: 10.1148/ryai.2019180001

    work page doi:10.1148/ryai.2019180001 2019

Show all 50 references
  1. [9]

    Analysis of ai applications in os- teoporosis and fracture risk assessment: A systematic review.Frontiers in Medicine, 10:1198954, 2023

    Liao Liu, Ren Wei Liu, Andrew Makmur, Janelle Yi Ting Phua, Wilson Ong, Tan Yi Liang, Dominic Fong Kuan Ting, Jiong Hao Tan, and James Thomas Patrick Decourcy Hallinan. Analysis of ai applications in os- teoporosis and fracture risk assessment: A systematic review.Frontiers in...

    2023 doi

  2. [10]

    Deep learning for prediction of bone mineral den- sity from hand radiographs.Radiology: Ar- tificial Intelligence, 4(3):e210229, 2022

    Yi Liang Thian, Patricia Yoong, Wenrui Fang, Moe Moe Tin Win, Hui Zhang, Dominic Fong Kuan Ting, and James Hallinan. Deep learning for prediction of bone mineral den- sity from hand radiographs.Radiology: Ar- tificial Intelligence, 4(3):e210229, 2022. doi: 10.1148/ryai.210229

    2022 doi

  3. [11]

    Surapaneni, Richard Friedman, and M

    Nithya Pinnamaneni, Lakshmi Gullapudi, MichaelS.Harris, MerhawiMamo, KrishnaM. Surapaneni, Richard Friedman, and M. Ed- ward Oates. The use of convolutional neural networks for the classification of bone density from pelvic radiographs.Journal of Computer Assisted Tomography, ...

    2021 doi

  4. [12]

    King-Smith, Danyal Kamil, Unaiza Khalid, Yuli Luo, Leon Rybak, Yilan Zhou, Ashish Saha, Shobhit Teki, Moro O

    Ryan Lindsey, Anshul Gupta, Scott Foocharoen, Simrita Dasgupta, Sophia L. King-Smith, Danyal Kamil, Unaiza Khalid, Yuli Luo, Leon Rybak, Yilan Zhou, Ashish Saha, Shobhit Teki, Moro O. Salifu, and Samy I. McFarlane. A deep learning ap- proach to detect osteoporosis from chest r...

    2023 doi

  5. [13]

    Hand x-ray- based deep learning models for prediction of osteoporosis: a comparative study using the osteohand dataset.EBioMedicine, 88:104416,

    Xinong Wang, Jie Zhu, Xiangdong Cheng, Baosheng Guo, Pei Su, Wenzhou Xia, Haoyan He, Xuesong Wan, and Yu Jiang. Hand x-ray- based deep learning models for prediction of osteoporosis: a comparative study using the osteohand dataset.EBioMedicine, 88:104416,

  6. [14]

    Angelopoulos and Stephen Bates

    Anastasios N. Angelopoulos and Stephen Bates. A gentle introduction to conformal pre- diction and distribution-free uncertainty quan- tification.arXiv preprint arXiv:2107.07511, 2021

    2021 arXiv

  7. [15]

    Secondopinionneeded: commu- nicating uncertainty in medical machine learn- ing.npj Digital Medicine, 4(1):4, 2021

    Benjamin Kompa, Jasper Snoek, and An- drewLBeam. Secondopinionneeded: commu- nicating uncertainty in medical machine learn- ing.npj Digital Medicine, 4(1):4, 2021. doi: 10.1038/s41746-020-00367-3

    2021 doi

  8. [16]

    Harrison, and Shun Miao

    Ke Yan, Dakai Jin, Jinzheng Cai, Zizhao Zhang, Guozyu Tong, Weiyi Xie, Jing Xiao, Le Lu, Jingjing Lu, Adam P. Harrison, and Shun Miao. Opportunistic screening of osteo- porosis using plain film chest x-ray, 2021

    2021

  9. [17]

    Deep learning for bone mineral density and t-score prediction from chest x-rays: A multicenter study.Biomedicines, 10(9):2323, 2022

    Yoichi Sato, Norio Yamamoto, Naoya Ina- gaki, Yusuke Iesaki, Takamune Asamoto, To- mohiro Suzuki, and Shunsuke Takahara. Deep learning for bone mineral density and t-score prediction from chest x-rays: A multicenter study.Biomedicines, 10(9):2323, 2022. doi: 10.3390/biomedicin...

    2022 doi

  10. [18]

    Syme, Alexan- der Bilbily, Saba Zuberi, Maksims Volkovs, Tomi Poutanen, and Mark D

    Keyvan Golestan, Catriona A. Syme, Alexan- der Bilbily, Saba Zuberi, Maksims Volkovs, Tomi Poutanen, and Mark D. Cicero. Approxi- matingfemoralneckbonemineraldensityfrom hand, knee, and pelvis x-rays using deep learn- ing.Journal of Medical Artificial Intelligence, 6:7, June 2...

    2023 doi

  11. [19]

    Automated bone mineral density predic- tion and fracture risk assessment using plain radiographs via deep learning.Nature Com- munications, 12:5472, 2021

    Chen-I Hsieh, Kang Zheng, Chihung Lin, Ling Mei, Le Lu, Weijian Li, Fang-Ping Chen, Yirui Wang, Xiaoyun Zhou, Fakai Wang, Guotong Xie, Jing Xiao, Shun Miao, and Chang-Fu Kuo. Automated bone mineral density predic- tion and fracture risk assessment using plain radiographs via d...

    2021

  12. [20]

    Prediction of osteoporosis from simple hip radiography using deep learning algorithm

    Ryoungwoo Jang, Jae Ho Choi, Namkug Kim, Jae Suk Chang, Pil Whan Yoon, and Chul-Ho Kim. Prediction of osteoporosis from simple hip radiography using deep learning algorithm. Scientific Reports, 11(1):19997, 2021. doi: 10. 1038/s41598-021-99549-6

    2021

  13. [21]

    Deep learning for osteoporosis clas- sification using hip radiographs and patient clinical covariates.Biomolecules, 10(11):1534,

    Norio Yamamoto, Shintaro Sukegawa, Akira Kitamura, Ryosuke Goto, Tomoyuki Noda, Keisuke Nakano, Kiyofumi Takabatake, Ho- taka Kawai, Hitoshi Nagatsuka, Keisuke Kawasaki, Yoshihiko Furuki, and Toshifumi Ozaki. Deep learning for osteoporosis clas- sification using hip radiograph...

  14. [22]

    Ar- tificial intelligence in dental panoramic radio- graphs for osteoporosis screening: A system- atic review.Diagnostics, 14(2):276, 2024

    Alexis Gaudin, Fabien Bornert, Wacław Ba- con, Damien Offner, and François Clauss. Ar- tificial intelligence in dental panoramic radio- graphs for osteoporosis screening: A system- atic review.Diagnostics, 14(2):276, 2024. doi: 10.3390/diagnostics14020276

    2024 doi

  15. [23]

    Development of an artificial intelli- gence system to classify bone quality and pre- dict fracture risk in dental panoramic radio- graphs.Diagnostics, 11(12):2219, 2021

    Norio Yamamoto, Shintaro Sukegawa, Akira Kitamura, Ryosuke Goto, Tomoyuki Noda, Keisuke Nakano, Kiyofumi Takabatake, Ho- taka Kawai, Hitoshi Nagatsuka, Keisuke Kawasaki, Yoshihiko Furuki, and Toshifumi Ozaki. Development of an artificial intelli- gence system to classify bone ...

    2021 doi

  16. [24]

    Osteoporosis prediction from hand and wrist x-rays using image segmentation and self-supervised learning, 2023

    Jeong-Hwan Kim, Chan-Ho Kim, Chan-Hyun Park, Jung-Woo Chang, Hyeong-Cheol Shin, Min-Ho Lee, Chul-Ho Lee, Sang-Hyun Kim, Young-Jin Kim, Jae-Hyun Lee, and Hyun- Jin Kim. Osteoporosis prediction from hand and wrist x-rays using image segmentation and self-supervised learning, 2023

    2023

  17. [25]

    L. Mao, Z. Xia, L. Pan, J. Chen, X. Liu, Z. Li, Z. Yan, G. Lin, H. Wen, and B. Liu. Deep learning for screening primary osteopenia and osteoporosis using spine radiographs and pa- tient clinical covariates in a chinese popula- tion.Frontiers in Endocrinology, 13:971877,

  18. [26]

    doi: 10.3390/biom10111534. 10

    doi

  19. [27]

    The ex- ploration of feature extraction and machine learning for predicting bone density from sim- ple spine x-ray images in a korean population

    Sangwoo Lee, E K Choe, Hye Young Kang, J W Yoon, and Hyung Seok Kim. The ex- ploration of feature extraction and machine learning for predicting bone density from sim- ple spine x-ray images in a korean population. Skeletal Radiology, 49(4):613–618, 2020. doi: 10.1007/s00256-0...

    2020 doi

  20. [28]

    Deep learning for automatic assessment of vertebral fractures on spine x-ray images: the vfss-10k dataset and a benchmark model.European Radiol- ogy, 31(9):7082–7093, 2021

    Yan-Ran Wang, Jing Li, Jian-Wu Wei, Hua Jiang, and Xiao-Feng Jiang. Deep learning for automatic assessment of vertebral fractures on spine x-ray images: the vfss-10k dataset and a benchmark model.European Radiol- ogy, 31(9):7082–7093, 2021. doi: 10.1007/ s00330-021-07788-y

    2021

  21. [29]

    Application of artificial intelligence methods on osteoporosis classification with radiographs—a systematic review.Bioengineering, 11(5):484, 2024

    Ren Wei Liu, Wilson Ong, Andrew Mak- mur, Naresh Kumar, Xi Zhen Low, Ge Shu- liang, Tan Yi Liang, Dominic Fong Kuan Ting, Jiong Hao Tan, and James Thomas Patrick Decourcy Hallinan. Application of artificial intelligence methods on osteoporosis classification with radiographs—a...

    2024 doi

  22. [30]

    Machine learning value in the diagnosis of vertebral fractures: A system- atic review and meta-analysis.European Jour- nal of Radiology, page 111598, 2024

    Meng-Chen Liu, Yan Wang, Yang-Yang Liu, Jia-Ming Liu, Xin-Ya Wang, Jian-Xiong Wu, and Lei Chu. Machine learning value in the diagnosis of vertebral fractures: A system- atic review and meta-analysis.European Jour- nal of Radiology, page 111598, 2024. doi: 10.1016/j.ejrad.2024....

    2024

  23. [31]

    Deep learning in the radiologic diagnosis of osteoporosis: a literature review

    Yu He, Jiaxi Lin, Shiqi Zhu, Jinzhou Zhu, and Zhonghua Xu. Deep learning in the radiologic diagnosis of osteoporosis: a literature review. Journal of International Medical Research, 52 (4):03000605241244754, 2024. doi: 10.1177/ 03000605241244754

    2024

  24. [32]

    Deeplearningoflumbarspinex-ray for osteopenia and osteoporosis screening: A multicenter retrospective cohort study.Bone, 140:115561, 2020

    Bin Zhang, Keyan Yu, Zhenyuan Ning, Ke Wang, Yuhao Dong, Xian Liu, Shuxue Liu, Jian Wang, Cuiling Zhu, Qinqin Yu, Yuwen Duan, Siying Lv, Xintao Zhang, Yanjun Chen, Xiaojia Wang, Jie Shen, Jia Peng, Qiuying Chen, Yu Zhang, Xiaodong Zhang, and Shuix- ingZhang. Deeplearningoflumb...

    2020 doi

  25. [33]

    A tutorial on conformal prediction.Journal of Machine Learning Research, 9:371–421, 2008

    Glenn Shafer and Vladimir Vovk. A tutorial on conformal prediction.Journal of Machine Learning Research, 9:371–421, 2008

    2008

  26. [34]

    Conformalized quantile re- gression.Advances in Neural Information Pro- cessing Systems, 32:3543–3553, 2019

    Yaniv Romano, Evan Patterson, and Em- manuel Candes. Conformalized quantile re- gression.Advances in Neural Information Pro- cessing Systems, 32:3543–3553, 2019

    2019

  27. [35]

    Efficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc V Le. Efficientnet: Rethinking model scaling for convolutional neural networks. InProceedings of the 36th International Conference on Machine Learn- ing (ICML), pages 6105–6114, 2019

    2019

  28. [36]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  29. [37]

    Test- time data augmentation for estimation of het- eroscedastic aleatoric uncertainty in deep neu- ral networks.Medical Imaging with Deep Learning Conference, 2018

    Murat Serdar Ayhan and Philipp Berens. Test- time data augmentation for estimation of het- eroscedastic aleatoric uncertainty in deep neu- ral networks.Medical Imaging with Deep Learning Conference, 2018. URLhttps:// openreview.net/forum?id=rJe6syrtvH

    2018

  30. [38]

    Springer, New York, NY,

    Vladimir Vovk, Alexander Gammerman, and Glenn Shafer.Algorithmic Learning in a Random World. Springer, New York, NY,

  31. [39]

    A survey of un- certainty in deep neural networks.CoRR, abs/2107.03342, 2021

    Jakob Gawlikowski, Cedric Rommel Nana Tassi, Mohsin Ali, Jongseok Lee, Matthias Humt, JianxiangFeng, AnnaKruspe, Rudolph Triebel, Peter Jung, Ribana Roscher, Muham- mad Shahzad, Wentong Yang, Richard Bam- ler, and Xiao Xiang Zhu. A survey of un- certainty in deep neural networ...

    2021 arXiv

  32. [40]

    An uncertainty-aware multiple-sampling strategy for image-to-image translation.IEEE Sig- nal Processing Letters, 29:867–871, 2022

    Lijun Yang, Yufeng Qi, Jiancheng Zhang, Xingjian Wang, and Lin Yang. An uncertainty-aware multiple-sampling strategy for image-to-image translation.IEEE Sig- nal Processing Letters, 29:867–871, 2022. doi: 10.1109/LSP.2022.3166648

    2022

  33. [41]

    Deep neural networks for automatic detection of osteoporotic vertebral fractures on ct scans.Computers in Biol- ogy and Medicine, 123:103899, 2020

    Naofumi Tomita, Yvonne Y Cheung, and Saeed Hassanpour. Deep neural networks for automatic detection of osteoporotic vertebral fractures on ct scans.Computers in Biol- ogy and Medicine, 123:103899, 2020. doi: 10.1016/j.compbiomed.2020.103899

    2020

  34. [42]

    Predictionofosteoporosisusingmachinelearn- ing based on routine blood test.Journal of Medical Internet Research, 23(3):e25082, 2021

    Kang Cheol Kim, Hyun Cheol Cho, Tae Young Jang, Young Joong Kim, and Kyung Hoi Koo. Predictionofosteoporosisusingmachinelearn- ing based on routine blood test.Journal of Medical Internet Research, 23(3):e25082, 2021. doi: 10.2196/25082. 12

    2021 doi

  35. [45]

    Automatic brain tu- mor segmentation based on cascaded convolu- 11 tional neural networks with uncertainty esti- mation.Medical Image Analysis, 67:101816,

    Guotai Wang, Wenqi Li, Tom Vercauteren, and Sebastien Ourselin. Automatic brain tu- mor segmentation based on cascaded convolu- 11 tional neural networks with uncertainty esti- mation.Medical Image Analysis, 67:101816,

  36. [46]

    doi: 10.1016/j.media.2020.101816

    2020

  37. [2005]

    doi: 10.1007/ b106715

    ISBN 978-0-387-25061-1. doi: 10.1007/ b106715

  38. [2016]

    doi: 10.1001/jama.2016.17216

    2016

  39. [2017]

    doi: 10.1038/nature21056

    doi

  40. [2020]

    doi: 10.1038/s41598-020-77469-1

    doi

  41. [2022]

    doi: 10.3389/fendo.2022.971877

    2022

  42. [2023]

    doi: 10.1016/j.ebiom.2023.104416

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.