Pith. sign in

REVIEW 4 major objections 7 minor 39 references

DICOM De-Identification via Hybrid AI and Rule-Based Framework for Scalable, Uncertainty-Aware Redaction

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A hybrid AI-and-rules pipeline can strip PHI from DICOM files with a 99.88 percent compliance pass rate, the authors report.

desk verdict A well-engineered system description whose headline 99.88% compliance claim is unverifiable as reported: the evaluation corpus is undescribed and the uncertainty threshold is tuned on the test set. read the letter →

arxiv 2507.23736 v1 pith:533M237M submitted 2025-07-31 stat.ML cs.LG

Kyle Naddeo , Nikolas Koutsoubis , Rahul Krish , Ghulam Rasool , Nidhal Bouaynaya , Tony OSullivan , Raj Krish This is my paper
classification stat.MLcs.LG
keywords DICOMde-identificationPHI/PIIredactionnamedentityrecognitionuncertaintyquantificationFasterR-CNNsyntheticclinicaldatamedicalimagingprivacyopticalcharacter
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 argues that DICOM de-identification can be automated to production grade by combining deterministic rules with AI models that report their own uncertainty, rather than relying on either approach alone. A two-tiered rule system strips obvious and inferred identifiers from metadata, a fine-tuned named-entity model catches PHI in clinical text, and an uncertainty-aware object detector localizes burned-in text in pixels before OCR and the same NER model decide what to redact. The headline evidence is an overall 99.88 percent pass rate across roughly 581,000 DICOM, HIPAA, and imaging-archive compliance checks, with 670 failures concentrated in a few archive subcategories. A sympathetic reader would care because, if the transfer holds, this is a path to releasing large medical imaging datasets for research without either exposing patients or forcing most work through manual review.

What carries the argument

The load-bearing object is an uncertainty-aware Faster R-CNN whose classification head is replaced by a variational density propagation layer: weights, inputs, and biases are treated as random variables, and the mean and variance are propagated through the network (with a first-order Taylor expansion at nonlinearities) so every detection comes with a variance estimate. The training loss combines bounding-box regression, a negative log-likelihood term in the predicted class distribution, and KL regularization of the weight posteriors; hyperparameters are selected by Bayesian optimization against a five-component objective that rewards detection precision, a negative slope of uncertainty under added noise, low variance of uncertainty on clean images, and a low false-negative rate at the chosen uncertainty threshold. On the text side, the framework uses a fine-tuned transformer NER model over rule-based 'recipe' tags plus Levenshtein-distance fuzzy matching, so near-miss spellings of a flagged name are also redacted. The uncertainty threshold itself is set at the smallest normalized variance that produced no false positive in testing, making the threshold a calibrated boundary between automatic redaction and human review.

What would settle it

Take a held-out set of real clinical DICOM files whose PHI has been manually enumerated, run the full pipeline, and count residual PHI plus the calibration of uncertainty on those real images. If the residual-PHI rate is materially above the 0.12 percent failure rate reported here, or if variance stops correlating with detection correctness on real data, the central claim of transferable, production-ready performance is falsified.

Watch

Extended reading notes

Core claim

The central claim is that the complete pipeline—rule-based metadata recipes, a transformer NER model fine-tuned on synthetic clinical text, and a Faster R-CNN detector with a Bayesian classification head—removes PHI/PII from DICOM metadata and pixels while preserving non-PHI text such as laterality markers and imaging parameters. The authors report a 99.88 percent overall pass rate on DICOM, HIPAA, and archive-specific compliance checks, and 0.779 mean average precision (at IoU 0.50:0.95), 0.997 at IoU 0.50, and 0.885 at IoU 0.75 for the pixel-text detector on an external real-image validation set. The uncertainty signal is not decorative: detections below a calibrated variance threshold are redacted automatically, while high-uncertainty detections are quarantined for human review, which is the mechanism that makes the claimed scalability credible.

Load-bearing premise

The paper's 99.88 percent pass rate depends on the synthetic training data—fake names, machine-generated clinical notes, and text stamped onto clean images—being close enough to real clinical PHI that the uncertainty thresholds and loss coefficients carry over to genuine DICOM files.

Editorial extensions

If this is right

  • If the 99.88 percent pass rate holds, hospitals and research consortia can release large imaging datasets automatically instead of hiring manual reviewers for every file.
  • Selective redaction preserves non-PHI text such as 'LEFT'/'RIGHT' and imaging parameters, so downstream research keeps more clinical utility than blanket pixel-masking tools.
  • The adjustable uncertainty threshold gives each institution a dial between maximum privacy and maximum throughput, with a documented audit trail of which detections were automatic versus human-reviewed.
  • The same two-stage design—detect text regions, OCR, then NER—could be retrained on synthetic data for other image formats and modalities, since the pipeline is not tied to DICOM-specific tag handling alone.

Reading between the lines

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

  • Editorial inference: the paper's own Section 3.2 notes that no uncertainty reactance could be measured on the external real-image validation set because the noise-augmented images were folded into training; that gap is exactly why the synthetic-to-real transfer question is the open risk.
  • Editorial inference: a practical extension would be a continuous recalibration loop in which human reviewers' accept/reject decisions on quarantined cases feed back into the uncertainty threshold, turning the static threshold into an adaptive one.
  • Editorial inference: the gap between mAP at IoU 0.50 (0.997) and mAP at IoU 0.50:0.95 (0.779) suggests bounding boxes are frequently imprecise; since OCR crops the detected region, imprecise boxes could clip PHI at the edges, making box-precision loss at least as important as classification accuracy for privacy.
  • Editorial inference: because the rule-based layer and NER threshold jointly produce near-zero failures on DICOM and HIPAA categories, the residual risk concentrates in archive-specific pixel categories; focusing future synthetic data generation on those edge cases may raise the overall pass rate beyond 99.88 percent.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper describes a hybrid DICOM de-identification framework that combines rule-based metadata scrubbing, a fine-tuned LUKE NER model, and an uncertainty-aware Faster R-CNN for pixel-level text detection, followed by OCR and NER-based redaction. Training data are synthetic: Faker for metadata, ChatGPT-4 for admission notes, and OpenCV-stamped text on TCIA images. The authors report strong detector performance on the MIDI validation set (mAP 0.779 at IoU 0.50:0.95) and a 99.88% overall pass rate across DICOM, HIPAA, and TCIA compliance checks (Table 3). The central claim is that the framework is production-ready and uncertainty-aware, with a threshold-based human-in-the-loop review for uncertain detections.

Significance. If the reported results are reliable, the framework addresses a real and important problem: scalable, privacy-preserving sharing of medical imaging data. The paper earns credit for using synthetic data to avoid training on real PHI, for propagating variance through the classification head of a Faster R-CNN, and for benchmarking the detector on the MIDI validation set, which is a concrete, externally defined test. The uncertainty-aware design and the explicit goal of preserving non-PHI text are also noteworthy. However, the main significance claim rests on Table 3, whose evaluation corpus is not described, and on an uncertainty threshold that is fitted on the test set. As written, the manuscript does not provide sufficient evidence to support the production-readiness claim.

major comments (4)
  1. [Section 3.3, Table 3] The evaluation corpus underlying Table 3 is not described. The table reports 581,265 'instances' across DICOM, HIPAA, and TCIA categories, but the paper never states how many DICOM files, studies, or patients were processed, what imaging modalities were included, what the data source was, how ground-truth labels were generated, or what constitutes a 'pass'/'fail' instance (per tag, per file, per validation rule). Without this information, the headline 99.88% pass rate cannot be interpreted or reproduced. The overall rate is also dominated by two subcategories, DICOM-IOD-1 (170,646 instances) and TCIA-REV (217,770 instances), which are not defined in the text.
  2. [Sections 2.1.3 and 4] There is a direct contradiction about real-world validation. Section 2.1.3 states that 'the models were validated using real-world medical DICOMS obtained from the Moffitt Cancer Center,' while Section 4 states that 'we have initiated a formal, IRB-approved retrospective clinical study at Moffitt Cancer Center to validate our framework using real patient imaging data.' If the Section 2.1.3 validation is complete, the study cannot be described as merely initiated; if the study has only been initiated, then Table 3 cannot rest on real-world Moffitt data and must rest on synthetic data, which would undermine the production-readiness claim. The manuscript must state clearly which data underlie Table 3 and must reconcile this contradiction.
  3. [Equations (25)-(26), Section 2.2.2] The uncertainty threshold u is selected as the largest value with zero false positives on the testing set (Eq. 25), and the false-negative rate FNRthresh is then computed on the same testing set at that threshold (Eq. 26). This constitutes fitting the threshold to the test data; the resulting FNR is an optimistically biased estimate of the operating point, not a predictive performance measure. A held-out calibration set or nested cross-validation is needed to report an unbiased FNR. The same concern applies to the six loss coefficients in Eq. (18), which are tuned via Bayesian optimization on the 2,000-image test split (Section 3.2) before that split is folded into training; after this procedure, no held-out data from the synthetic corpus remain to evaluate the final uncertainty-aware detector.
  4. [Section 3.2] The optimal loss coefficients are 'withheld as proprietary.' Since the loss in Eq. (18), the objective in Eq. (19), and the reported detector behavior all depend on these six coefficients, the experimental results cannot be reproduced or independently verified. Journals require disclosure of all hyperparameters and model configurations; the authors should release the full coefficient vector and the search ranges used in the Bayesian sweep, or present an argument why reproducibility is not required.
minor comments (7)
  1. [Section 2.2.2, before Eq. (17)] The sentence 'By removing constants and and adding coefficients' contains a duplicated 'and'; please correct this typo.
  2. [Section 2.2.2, near Eq. (25)] The text 'setting a practical threshold at the smallest normalized variance that yielded a false positive in testing' is inconsistent with Eq. (25), which defines the threshold as the largest variance with zero false positives; please correct the wording to match the equation.
  3. [Section 2.1.2, reference [39]] Reference [39] cites the i2b2 clinical data query tool at Mayo Clinic, but the text refers to sampling from the i2b2 dataset; please cite the original i2b2 data publication instead.
  4. [Throughout] There are several typos and wording issues: 'illistrated' (Section 2.2.2), 'especally' (Section 2.2.2), 'neglible' (after Eq. 10), and 'reactance' (used for the noise response in Figure 5). A careful proofread is recommended.
  5. [End matter] The paper contains duplicated Acknowledgments sections and three overlapping conflict-of-interest statements; these should be consolidated into a single Acknowledgments section and a single disclosure statement.
  6. [Table 3 caption] The caption should define the units of 'Fail', 'Pass', and 'Total' (e.g., per tag value, per file, per rule), and briefly describe the validation rules used for each category.
  7. [Section 3.2] The sentence 'we trained for 369 epochs' should state the early-stopping or model-selection criterion used to pick the final checkpoint, since the reported metrics come from that checkpoint.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: detector claims rest on an external MIDI validation benchmark, and the threshold-fitted FNR term is only an optimization objective, not a reported prediction.

full rationale

The paper's central performance claims do not reduce to their inputs by construction. The detector result in Table 2 (mAP 0.779 at IoU 0.50:0.95) is measured on the MIDI Challenge validation set, an external corpus not used to set the loss coefficients or the uncertainty threshold. The closest candidate for circularity is Equations 25-26: IoUthresh is chosen as the largest uncertainty threshold with FP(u)=0 on a custom test set, and FNRthresh is then computed at that threshold. However, FNRthresh appears only as one term in the Bayesian-optimization objective (Equation 19) for tuning internal loss coefficients; the paper never reports FNRthresh as a performance result or generalizes from it. Choosing a threshold on the evaluation set is a form of overfitting or threshold optimism, not a derivation that is equivalent to its input. The uncertainty formulation is adopted from the authors' prior work [27], but that is a legitimate method citation, and the downstream detector quality is independently checked on the MIDI validation set, so the self-citation is not load-bearing to the headline result. Separately, Table 3's 99.88% pass rate is not verifiable because the evaluation corpus is not described, and Section 2.1.3's statement that the model "was validated using real-world medical DICOMS obtained from the Moffitt Cancer Center" conflicts with Section 4's statement that a formal IRB-approved study "has been initiated." These are reproducibility and integrity concerns, not circularity, and under the hard rules they do not raise the circularity score.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a chain of assumptions about synthetic data faithfully representing real PHI, plus several fitted parameters (loss coefficients, uncertainty threshold, fuzzy threshold) that are either withheld or fit post hoc. The paper introduces no new physical or conceptual entities; its novelty is in the engineered combination.

free parameters (3)
  • Loss coefficients lambda in Equation 18 = withheld as proprietary
    Tuned by Bayesian optimization to maximize Psi (Eq 19) on the custom synthetic dataset; not disclosed, so the trained detector is not reproducible.
  • Uncertainty threshold u (normalized variance) = smallest normalized variance that yielded a false positive in testing
    Set post hoc on the test set (Section 2.2.2), making the reported FNR conditional on a test-fitted parameter.
  • FuzzyWuzzy Levenshtein similarity threshold = not specified
    The rule-based second tier uses a threshold on Levenshtein distance to flag near matches; the threshold value is not given, yet it affects which variants are redacted.
assumptions (4)
  • domain assumption Synthetic metadata and admission notes generated by Faker and ChatGPT-4 are statistically representative of real clinical DICOM PHI.
    The entire NER training and evaluation rest on this; Section 2.1.1-2.1.2. No evidence of distributional similarity is provided.
  • domain assumption Text embedded into clean TCIA images with OpenCV using random fonts and PHI strings approximates real burned-in text in medical images.
    Used to train the Faster R-CNN detector (Section 2.1.3); no real burned-in text is used in training, and validation on Moffitt data is only stated, not quantified.
  • domain assumption The first-order Taylor expansion and variance-only diagonal approximation in variational density propagation (Eq 10-13) are accurate enough for uncertainty quantification of the detection head.
    The approximation is taken from [27] and assumed valid for this detection task without sensitivity checks.
  • domain assumption Tesseract OCR errors are either absent or caught by the NER pipeline, so OCR noise does not break PHI identification.
    Pipeline assumes OCR-extracted text is faithful; no error analysis of the OCR stage is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DICOM De-Identification via Hybrid AI and Rule-Based Framework for Scalable, Uncertainty-Aware Redaction." pith.science (2026). https://pith.science/paper/533M237M

@misc{pith2026250723736,
  author       = {Pith},
  title        = {Pith review of: DICOM De-Identification via Hybrid AI and Rule-Based Framework for Scalable, Uncertainty-Aware Redaction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/533M237M}},
  note         = {Machine review of arXiv:2507.23736}
}
read the original abstract

Access to medical imaging and associated text data has the potential to drive major advances in healthcare research and patient outcomes. However, the presence of Protected Health Information (PHI) and Personally Identifiable Information (PII) in Digital Imaging and Communications in Medicine (DICOM) files presents a significant barrier to the ethical and secure sharing of imaging datasets. This paper presents a hybrid de-identification framework developed by Impact Business Information Solutions (IBIS) that combines rule-based and AI-driven techniques, and rigorous uncertainty quantification for comprehensive PHI/PII removal from both metadata and pixel data. Our approach begins with a two-tiered rule-based system targeting explicit and inferred metadata elements, further augmented by a large language model (LLM) fine-tuned for Named Entity Recognition (NER), and trained on a suite of synthetic datasets simulating realistic clinical PHI/PII. For pixel data, we employ an uncertainty-aware Faster R-CNN model to localize embedded text, extract candidate PHI via Optical Character Recognition (OCR), and apply the NER pipeline for final redaction. Crucially, uncertainty quantification provides confidence measures for AI-based detections to enhance automation reliability and enable informed human-in-the-loop verification to manage residual risks. This uncertainty-aware deidentification framework achieves robust performance across benchmark datasets and regulatory standards, including DICOM, HIPAA, and TCIA compliance metrics. By combining scalable automation, uncertainty quantification, and rigorous quality assurance, our solution addresses critical challenges in medical data de-identification and supports the secure, ethical, and trustworthy release of imaging data for research.

Figures

Figures reproduced from arXiv: 2507.23736 by the authors.

Figure 1
Figure 1. This diagram illustrates our general workflow for DICOM-based deidentification. A DICOM file consists [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Illustration of how Levenshtein Distance quantifies textual differences in DICOM tags. In Example 1, the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Examples of bounding-box overlap where green (solid line) boxes are the labels and blue (dashed line) are [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Examples of the effects of salt and pepper noise on model predictions (class prediction mean, M) and [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Examples illustrating the effect of loss coefficients on a model’s reactance to out-of-distribution data. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Relationship between uncertainty and IoU scores. The plot is divided into four quadrants: (Q1, Top-right) [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 37 canonical work pages

  1. [1]

    Beyond the HIPAA Privacy Rule: Enhancing Privacy, Improving Health Through Research.National Academies Press (US), 2009

    Institute of Medicine (US) Committee on Health Research and the Privacy of Health Information. Beyond the HIPAA Privacy Rule: Enhancing Privacy, Improving Health Through Research.National Academies Press (US), 2009

    work page 2009

  2. [2]

    Security and Privacy of Technologies in Health Informa- tion Systems: A Systematic Literature Review

    Shojaei, P.; Vlahu-Gjorgievska, E.; Chow, Y .-W. Security and Privacy of Technologies in Health Informa- tion Systems: A Systematic Literature Review. Computers, 13(41), 2024. https://doi.org/10.3390/ computers13020041

    work page 2024

  3. [3]

    B.; Kirby, J

    Freymann, J. B.; Kirby, J. S.; Perry, J. H.; Clunie, D. A.; Jaffe, C. C. Image data sharing for biomedical research—meeting HIPAA requirements for de-identification. J. Digit. Imaging, 25(1):14–24, 2012

    work page 2012

  4. [4]

    Y .; van Kernebeek, G.; Berendsen, B.; Oudkerk, M.; van Ooijen, P

    Aryanto, K. Y .; van Kernebeek, G.; Berendsen, B.; Oudkerk, M.; van Ooijen, P. M. Image de-identification methods for clinical research in the XDS environment. J. Med. Syst., 40(4):83, 2016

    work page 2016

  5. [5]

    Ethical considerations and bias mitigation in deep learning for medical imaging

    Kayode, S. Ethical considerations and bias mitigation in deep learning for medical imaging. Manuscript submit- ted for publication, 2024

    work page 2024

  6. [6]

    Methods for De-Identification of PHI

    Office for Civil Rights (OCR). Methods for De-Identification of PHI. HHS.gov, 22 Feb. 2023. https://www. hhs.gov/hipaa/for-professionals/privacy/special-topics/de-identification/index.html

    work page 2023

  7. [7]

    Clunie, D. A. et al. Report of the Medical Image De-Identification (MIDI) Task Group - Best Practices and Recommendations. arXiv pre-print, 2023. https://arxiv.org/abs/2303.10473

    work page Pith review arXiv 2023

  8. [8]

    The Limitations of De- Identification – Protecting Unit-Record Level Personal Information

    Office of the Victorian Information Commissioner (OVIC). The Limitations of De- Identification – Protecting Unit-Record Level Personal Information. Accessed 6 Aug. 2024. https://ovic.vic.gov.au/privacy/resources-for-organisations/ the-limitations-of-de-identification-protecting-unit-record-level-personal-information/

    work page 2024

Show all 39 references
  1. [9]

    Shahid, A. et al. A two-stage de-identification process for privacy-preserving medical image analysis.Healthcare (Basel), 10(5):755, 2022

    2022

  2. [10]

    Moore, S. M. et al. De-identification of medical images with retention of scientific research value.Radiographics, 35(3):727–735, 2015

    2015

  3. [11]

    Diaz, O. et al. Data preparation for artificial intelligence in medical imaging: A comprehensive guide to open- access platforms and tools. Phys. Med., 83:25–37, 2021

    2021

  4. [12]

    Murugadoss, K. et al. Building a best-in-class automated de-identification tool for electronic health records through ensemble learning. Patterns, 2(6):100255, 2021

    2021

  5. [13]

    hide in plain sight

    Chambon, P. J. et al. Automated de-identification of radiology reports combining transformer and “hide in plain sight” rule-based methods. J. Am. Med. Inform. Assoc., 30(2):318–328, 2023

    2023

  6. [14]

    Ahmed, M. I. et al. A systematic review of the barriers to the implementation of artificial intelligence in health- care. Cureus, 15(10):e46454, 2023

    2023

  7. [15]

    Abdar, M. et al. A review of uncertainty quantification in deep learning: Techniques, applications, and challenges. Inf. Fusion, 76:243–297, 2021

    2021

  8. [16]

    Challen, R. et al. Artificial intelligence, bias and clinical safety. BMJ Qual. Saf., 28(3):231–237, 2019

    2019

  9. [17]

    Seastedt, K. P. et al. Global healthcare fairness: We should be sharing more, not less, data. PLOS Digit. Health, 1(10):e0000102, 2022. 14 A PREPRINT - S EPTEMBER 24, 2025

    2022

  10. [18]

    Faker: Python library for generating fake data

    Faraglia, D. Faker: Python library for generating fake data. 2019. https://github.com/joke2k/faker

    2019

  11. [19]

    Clark, K. et al. The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Reposi- tory. J. Digit. Imaging, 26(6):1045–1057, 2013

    2013

  12. [20]

    Open Source Computer Vision Library

    Itseez. Open Source Computer Vision Library. 2015. https://github.com/itseez/opencv

    2015

  13. [21]

    fuzzywuzzy: Fuzzy String Matching in Python, 2014–2021

    SeatGeek Inc. fuzzywuzzy: Fuzzy String Matching in Python, 2014–2021. https://github.com/seatgeek/ fuzzywuzzy

    2014

  14. [22]

    Yamada, I. et al. LUKE: Deep Contextualized Entity Representations with Entity-aware Self-attention. 2020. https://arxiv.org/abs/2010.01057

    2020 arXiv

  15. [23]

    Choi, E. et al. Ultra-Fine Entity Typing. In Proceedings of the ACL, 2018

    2018

  16. [24]

    Zhang, Y . et al. Position-aware Attention and Supervised Data Improve Slot Filling. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 35–45, 2017

    2017

  17. [25]

    F.; De Meulder, F

    Tjong Kim Sang, E. F.; De Meulder, F. Introduction to the CoNLL-2003 Shared Task: Language-Independent Named Entity Recognition. 2003. https://arxiv.org/abs/cs/0306050

    2003 arXiv

  18. [26]

    Ren, S. et al. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks.IEEE Trans. Pattern Anal. Mach. Intell., 39(6):1137–1149, 2017

    2017

  19. [27]

    C.; Rasool, G.; Shterenberg, R.; Fathallah-Shaykh, H

    Dera, D.; Bouaynaya, N. C.; Rasool, G.; Shterenberg, R.; Fathallah-Shaykh, H. M. Premi-Um-CNN: Propagating Uncertainty Towards Robust Convolutional Neural Networks.IEEE Trans. Signal Process., 69:4669–4684, 2021

    2021

  20. [28]

    Tesseract: an open-source optical character recognition engine

    Kay, A. Tesseract: an open-source optical character recognition engine. Linux J., 2007(159):2, 2007

    2007

  21. [29]

    Rutherford, M. et al. A DICOM dataset for evaluation of medical image de-identification. Sci. Data, 8(183), 2021

    2021

  22. [30]

    Model inversion attacks that exploit confidence information and ba- sic countermeasures

    Fredrikson, M., Jha, S., & Ristenpart, T. Model inversion attacks that exploit confidence information and ba- sic countermeasures. Proceedings of the 22nd ACM SIGSAC Conference on Computer and Communications Security, 2015

    2015

  23. [31]

    Membership inference attacks against machine learning models

    Shokri, R., Stronati, M., Song, C., & Shmatikov, V . Membership inference attacks against machine learning models. Proceedings of the 38th IEEE Symposium on Security and Privacy, 2017

    2017

  24. [32]

    Karim, F. R. B., & Dera, D. Robust Bayesian Vision Transformer for Image Analysis and Classification. Pro- ceedings of the 2023 IEEE Western New York Image and Signal Processing Workshop (WNYISPW), Rochester, NY , USA, 2023, pp. 1–4, doi: 10.1109/WNYISPW60588.2023.10349631

    2023

  25. [33]

    DICOM Templates

    DICOM. “DICOM Templates.” DICOM, https://www.dicomstandard.org/resources/templates

  26. [34]

    ChatGPT-4,

    OpenAI, “ChatGPT-4,” OpenAI, 2024. [Online]. Available: https://openai.com/research/gpt-4. [Ac- cessed: 02-Mar-2025]

    2024

  27. [35]

    A de-identification pipeline for ultrasound medical images in DICOM format

    Eriksson Monteiro, Carlos Costa, and Jos ´e Lu´ıs Oliveira. A de-identification pipeline for ultrasound medical images in DICOM format. Journal of Medical Systems, 41(5):89, 2017

    2017

  28. [36]

    Mickley, Pouria Rouzrokh, Michael J

    Bardia Khosravi, John P. Mickley, Pouria Rouzrokh, Michael J. Taunton, A. Noelle Larson, Bradley J. Erickson, and Cody C. Wyles. Anonymizing Radiographs Using an Object Detection Deep Learning Algorithm.Radiology: Artificial Intelligence, 5(6):e230085, 2023

    2023

  29. [37]

    Transformers in medical image analysis

    Kelei He, Chen Gan, Zhuoyuan Li, Islem Rekik, Zihao Yin, Wen Ji, Yang Gao, Qian Wang, Junfeng Zhang, and Dinggang Shen. Transformers in medical image analysis. Intelligent Medicine, 3(1):17–39, 2023

    2023

  30. [38]

    Simon, Optimal State Estimation: Kalman, H-Infinity, and Nonlinear Approaches

    D. Simon, Optimal State Estimation: Kalman, H-Infinity, and Nonlinear Approaches. Hoboken, NJ, USA: Wiley- Interscience, 2006

    2006

  31. [39]

    i2b2 at Mayo Clinic: Self-service clinical data query tool

    Mayo Clinic. i2b2 at Mayo Clinic: Self-service clinical data query tool. Available at: https: //www.mayo.edu/research/centers-programs/center-clinical-translational-science/ research-resources/i2b2-self-service-clinical-data-query-tool , accessed 2024. 15

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.