Towards Selection of Large Multimodal Models as Engines for Burned-in Protected Health Information Detection in Medical Images

Guillermo Jimenez Perez; Matthias Lenga; Pedro Osorio; Tuan Truong

arxiv: 2511.02014 · v2 · pith:TD5TFFFKnew · submitted 2025-11-03 · 💻 cs.CV

Towards Selection of Large Multimodal Models as Engines for Burned-in Protected Health Information Detection in Medical Images

Tuan Truong , Guillermo Jimenez Perez , Pedro Osorio , Matthias Lenga This is my paper

Pith reviewed 2026-05-22 13:21 UTC · model grok-4.3

classification 💻 cs.CV

keywords PHI detectionLarge Multimodal Modelsburned-in textmedical imaging OCRpatient privacyLMM benchmarkingtext extraction

0 comments

The pith

Large multimodal models extract text from medical images more accurately than traditional OCR yet this does not always improve protected health information detection.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper benchmarks three large multimodal models for spotting protected health information burned into medical images. These models reach lower word and character error rates during text extraction than tools such as EasyOCR. Better text extraction does not reliably raise the overall accuracy of PHI detection across all images. Performance gains concentrate on cases with complex imprint patterns. The work supplies guidance on model choice for different conditions and outlines a modular approach for deployment.

Core claim

LMMs exhibit superior OCR efficacy with WER of 0.03-0.05 and CER of 0.02-0.03 compared to conventional models like EasyOCR. However, this OCR improvement does not consistently correlate with enhanced overall PHI detection accuracy. The strongest performance gains are observed on test cases with complex imprint patterns. In scenarios where text regions are well readable with sufficient contrast and strong LMMs are used for text analysis after OCR, different pipeline configurations yield similar results. Empirically grounded recommendations for LMM selection tailored to operational constraints and a deployment strategy that leverages scalable and modular infrastructure are proposed.

What carries the argument

Two pipeline configurations for LMMs: text analysis alone versus OCR integrated with semantic analysis, tested on GPT-4o, Gemini 2.5 Flash, and Qwen 2.5 7B for burned-in PHI detection.

If this is right

LMMs deliver stronger text extraction especially for complex imprint patterns in medical images.
Pipeline choice can be matched to text readability and model strength for similar final results in clear cases.
Modular infrastructure enables scalable PHI detection without full system redesign.
Strong LMMs used for analysis after OCR produce comparable outcomes to other setups when contrast is sufficient.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

LMM-based detection could reduce the need for manual PHI review in high-volume radiology settings.
The selective gains suggest prioritizing LMM use on difficult images rather than applying them uniformly.
Open models such as Qwen may support cost-effective scaling in resource-limited environments.
Broader validation across real hospital data streams would clarify operational reliability.

Load-bearing premise

The chosen test cases, including those with complex imprint patterns, represent real-world medical imaging data under varying contrast and readability conditions.

What would settle it

Applying the same LMM pipelines to a larger collection of medical images from multiple modalities and scanners with diverse imprint qualities would reveal whether the OCR advantages and selective PHI gains persist.

read the original abstract

The detection of Protected Health Information (PHI) in medical imaging is critical for safeguarding patient privacy and ensuring compliance with regulatory frameworks. Traditional detection methodologies predominantly utilize Optical Character Recognition (OCR) models in conjunction with named entity recognition. However, recent advancements in Large Multimodal Model (LMM) present new opportunities for enhanced text extraction and semantic analysis. In this study, we systematically benchmark three prominent closed and open-sourced LMMs, namely GPT-4o, Gemini 2.5 Flash, and Qwen 2.5 7B, utilizing two distinct pipeline configurations: one dedicated to text analysis alone and another integrating both OCR and semantic analysis. Our results indicate that LMM exhibits superior OCR efficacy (WER: 0.03-0.05, CER: 0.02-0.03) compared to conventional models like EasyOCR. However, this improvement in OCR performance does not consistently correlate with enhanced overall PHI detection accuracy. The strongest performance gains are observed on test cases with complex imprint patterns. In scenarios where text regions are well readable with sufficient contrast, and strong LMMs are employed for text analysis after OCR, different pipeline configurations yield similar results. Furthermore, we provide empirically grounded recommendations for LMM selection tailored to specific operational constraints and propose a deployment strategy that leverages scalable and modular infrastructure.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper benchmarks three LMMs for burned-in PHI detection and reports better OCR on complex cases, but the test set details are too thin to support the selection recommendations.

read the letter

The core of this work is a side-by-side test of GPT-4o, Gemini 2.5 Flash, and Qwen 2.5 7B on extracting burned-in PHI from medical images. They compare two pipelines, one focused on text analysis after OCR and one that folds in semantic understanding, and they show the LMMs hitting WER of 0.03-0.05 and CER of 0.02-0.03, which beats EasyOCR. The biggest reported lift shows up on the harder imprint patterns, while the pipelines converge when the text is already clean and high-contrast. They close with some practical notes on picking a model under different cost or latency limits and a modular deployment sketch. That part is straightforward and could save someone time when they need to choose an engine for a privacy pipeline.

Referee Report

1 major / 1 minor

Summary. The manuscript benchmarks three large multimodal models (GPT-4o, Gemini 2.5 Flash, Qwen 2.5 7B) in two pipeline configurations (text analysis alone versus OCR plus semantic analysis) for detecting burned-in protected health information in medical images. It reports that LMMs achieve superior OCR performance (WER 0.03-0.05, CER 0.02-0.03) relative to EasyOCR, yet this does not consistently improve overall PHI detection accuracy; the largest gains occur on test cases with complex imprint patterns. The paper supplies empirically grounded recommendations for LMM selection under different operational constraints together with a modular deployment strategy.

Significance. If the empirical results hold under broader conditions, the work would supply practical guidance for choosing between LMM-based and conventional OCR pipelines in medical-image privacy workflows, especially for difficult imprint cases. The direct reporting of concrete error metrics and the explicit comparison of pipeline configurations constitute a useful, falsifiable contribution to the literature on automated PHI redaction.

major comments (1)

Abstract and results sections: the claim that LMMs deliver the strongest PHI-detection gains specifically on complex imprint patterns rests on an uncharacterized test set. No quantitative information is supplied on total image count, sampling method, imaging modalities, contrast/font degradation distributions, or the criteria used to label cases as 'complex,' making it impossible to judge whether the reported differential performance generalizes beyond the chosen examples.

minor comments (1)

The abstract is dense; a single sentence stating the approximate size of the evaluation set and the range of modalities examined would immediately contextualize the reported WER/CER figures.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive and detailed review. We have revised the manuscript to address the concern regarding insufficient characterization of the test set, which we agree is necessary to support claims about differential performance on complex imprint patterns.

read point-by-point responses

Referee: Abstract and results sections: the claim that LMMs deliver the strongest PHI-detection gains specifically on complex imprint patterns rests on an uncharacterized test set. No quantitative information is supplied on total image count, sampling method, imaging modalities, contrast/font degradation distributions, or the criteria used to label cases as 'complex,' making it impossible to judge whether the reported differential performance generalizes beyond the chosen examples.

Authors: We agree that the original manuscript provided insufficient quantitative details on the test set to allow readers to evaluate generalizability of the reported gains on complex patterns. In the revised version we have added a new subsection under Methods titled 'Test Set Characterization' that specifies the total number of images, the sampling procedure (stratified random sampling from a de-identified clinical archive), the distribution across imaging modalities, measured distributions of contrast and font degradation, and the explicit criteria used to designate cases as 'complex' (images exhibiting at least two of: overlapping imprints, local contrast below a defined threshold, or non-standard fonts). We have also updated the abstract and results sections to reference these dataset properties. These changes directly respond to the referee's request for information needed to assess generalizability. revision: yes

Circularity Check

0 steps flagged

No significant circularity in empirical benchmarking study

full rationale

The paper is a purely empirical benchmarking study that reports direct performance measurements (WER, CER, PHI detection accuracy) from evaluating specific LMMs and OCR pipelines on chosen test cases. There are no derivations, equations, fitted parameters renamed as predictions, self-citation chains, or ansatzes that reduce any claimed result to its own inputs by construction. All reported outcomes are presented as observed experimental results rather than derived quantities, satisfying the criteria for a self-contained analysis against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claims rest on empirical observations rather than theoretical derivations; the main unstated premise is that the chosen test images adequately represent clinical variability in PHI imprinting.

axioms (1)

domain assumption The test cases with complex imprint patterns and varying readability are representative of real-world medical imaging scenarios requiring PHI detection.
Invoked to support generalization of performance gains and pipeline recommendations.

pith-pipeline@v0.9.0 · 5780 in / 1250 out tokens · 34507 ms · 2026-05-22T13:21:41.751733+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages · 4 internal anchors

[1]

Towards Selection of Large Multimodal Models as Engines for Burned-in Protected Health Information Detection in Medical Images

INTRODUCTION In recent years, the detection of Protected Health Informa- tion (PHI) has become increasingly important, particularly within the realm of medical data management. Traditional pipelines for PHI detection typically consist of Optical Char- acter Recognition (OCR) followed by analysis steps [1, 2, 3, 4, 5]. However, the advent of Large Multimod...

work page internal anchor Pith review Pith/arXiv arXiv 2025
[2]

The text local- ization module identifies text areas within images, while the text extraction module serves as an OCR engine, converting pixel-level text into machine-encoded text

METHODS In this study, we employ the PHI pipeline put forward by [7], which comprises three integral components: text localization, text extraction, and text analysis (Figure 1). The text local- ization module identifies text areas within images, while the text extraction module serves as an OCR engine, converting pixel-level text into machine-encoded tex...

work page
[3]

<0> Patient name: John Doe </0>

EXPERIMENTS 3.1. Data We use two datasetsRadPHI-testandMIDIwhich were introduced in [7] and are publicly available 1. These datasets cover a range of imaging modalities, imprint variations, and text complexities to support robust benchmarking. RadPHI- test (Figure 2a) includes 1000 radiological images equally distributed across four modalities with synthe...

work page doi:10.5281/zenodo.17201610 2024
[4]

RadPHI-test Table 3 presents a summary of the performance and latency metrics for three benchmark LMMs evaluated on the RadPHI- test dataset

RESULTS 4.1. RadPHI-test Table 3 presents a summary of the performance and latency metrics for three benchmark LMMs evaluated on the RadPHI- test dataset. Among the models, GPT-4o achieves the highest overall precision and recall, both exceeding 0.99, followed by Gemini 2.5 Flash and Qwen 2.5-VL 7B. The performance of the models is generally higher when e...

work page arXiv
[5]

LMM for Text Extraction The evaluation of OCR capabilities indicates that LMMs gen- erally outperform traditional OCR models (Table 1), although trade-offs depend on the dataset

DISCUSSION 5.1. LMM for Text Extraction The evaluation of OCR capabilities indicates that LMMs gen- erally outperform traditional OCR models (Table 1), although trade-offs depend on the dataset. On the RadPHI-test dataset, featuring complex and partially obscured imprints, the LMM- based approach (Setup B) shows slight improvements but in- curs a 20% to 4...

work page
[6]

The paper emphasizes the nuanced trade-offs between accuracy, latency, and pri- vacy, providing practical recommendations for deployment in healthcare settings

CONCLUSION In this study, we analyze the strengths and limitations of state-of-the-art models like GPT-4o, Gemini 2.5 Flash, and Qwen2.5-VL 7B under two different configurations of the LMM-based PHI detection pipeline. The paper emphasizes the nuanced trade-offs between accuracy, latency, and pri- vacy, providing practical recommendations for deployment i...

work page
[7]

Eth- ical approval was not required, as confirmed by the license attached to the sub-datasets

COMPLIANCE WITH ETHICAL STANDARDS This research study was conducted retrospectively using hu- man subject data derived from subsets of open-access data, which was introduced and published in the work of [7]. Eth- ical approval was not required, as confirmed by the license attached to the sub-datasets

work page
[8]

ACKNOWLEDGMENTS The authors would like to thank the Bayer team of the AI Innovation Platform for providing computing infrastructure and technical support

work page
[9]

Identification and classification of DICOM files with burned-in text content,

Petr Vcelak, Martin Kryl, Michal Kratochvil, and Jana Kleckova, “Identification and classification of DICOM files with burned-in text content,”International Jour- nal of Medical Informatics, vol. 126, pp. 128–137, June 2019

work page 2019
[10]

A Method for Efficient De- identification of DICOM Metadata and Burned-in Pixel Text,

Jacob A. Macdonald, Katelyn R. Morgan, Brandon Konkel, Kulsoom Abdullah, Mark Martin, Cory En- nis, Joseph Y . Lo, Marissa Stroo, Denise C. Snyder, and Mustafa R. Bashir, “A Method for Efficient De- identification of DICOM Metadata and Burned-in Pixel Text,”Journal of Imaging Informatics in Medicine, vol. 37, no. 5, pp. 1–7, Oct. 2024

work page 2024
[11]

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 Medi- cal Systems, vol. 41, no. 5, pp. 89, Apr. 2017

work page 2017
[12]

De-Identification of Medical Imaging Data: A Comprehensive Tool for En- suring Patient Privacy,

Moritz Rempe, Lukas Heine, Constantin Seibold, Fabian H¨orst, and Jens Kleesiek, “De-Identification of Medical Imaging Data: A Comprehensive Tool for En- suring Patient Privacy,” Oct. 2024, arXiv:2410.12402

work page arXiv 2024
[13]

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

Kyle Naddeo, Nikolas Koutsoubis, Rahul Krish, Ghu- lam Rasool, Nidhal Bouaynaya, Tony OSullivan, and Raj Krish, “DICOM De-Identification via Hybrid AI and Rule-Based Framework for Scalable, Uncertainty- Aware Redaction,” July 2025

work page 2025
[14]

Evaluation of Vision- Language Models for Detection and Deidentification of Medical Images with Burned-In Protected Health Infor- mation,

Taehee Lee, Hyungjin Kim, Seong Ho Park, Seonhye Chae, and Soon Ho Yoon, “Evaluation of Vision- Language Models for Detection and Deidentification of Medical Images with Burned-In Protected Health Infor- mation,”Radiology, vol. 315, no. 3, pp. e243664, June 2025

work page 2025
[15]

Exploring AI-Based Sys- tem Design for Pixel-Level Protected Health Informa- tion Detection in Medical Images,

Tuan Truong, Ivo M. Baltruschat, Mark Klemens, Grit Werner, and Matthias Lenga, “Exploring AI-Based Sys- tem Design for Pixel-Level Protected Health Informa- tion Detection in Medical Images,”Journal of Imaging Informatics in Medicine, July 2025

work page 2025
[16]

David Clunie et al., “Summary of the National Cancer Institute 2023 Virtual Workshop on Medical Image De- identification-Part 2: Pathology Whole Slide Image De- identification, De-facing, the Role of AI in Image De- identification, and the NCI MIDI Datasets and Pipeline,” Journal of Imaging Informatics in Medicine, July 2024

work page 2023
[17]

GPT-4o System Card

OpenAI et al., “GPT-4o System Card,” Oct. 2024, arXiv:2410.21276

work page internal anchor Pith review Pith/arXiv arXiv 2024
[18]

Gemini 2.5: Pushing the Frontier with Advanced Reasoning,

Google DeepMind and Google Research, “Gemini 2.5: Pushing the Frontier with Advanced Reasoning,” 2025

work page 2025
[19]

Qwen2.5-VL Technical Report

Alibaba Cloud / QwenLM, “Qwen2.5-VL Technical Re- port,” 2025, arXiv:2502.13923

work page internal anchor Pith review Pith/arXiv arXiv 2025
[20]

Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling,

Zhe Chen et al., “Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling,” Sept. 2025

work page 2025
[21]

TrOCR: Transformer-based Optical Charac- ter Recognition with Pre-trained Models,

Minghao Li, Tengchao Lv, Jingye Chen, Lei Cui, Yi- juan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, and Furu Wei, “TrOCR: Transformer-based Optical Charac- ter Recognition with Pre-trained Models,” 2021, eprint: 2109.10282

work page arXiv 2021
[22]

General OCR Theory: Towards OCR-2.0 via a Unified End-to-end Model

Haoran Wei et al., “General OCR Theory: To- wards OCR-2.0 via a Unified End-to-end Model,” 2024, eprint: 2409.01704

work page internal anchor Pith review Pith/arXiv arXiv 2024
[23]

JaidedAI/EasyOCR: Ready-to-use OCR with 80+ sup- ported languages and all popular writing scripts includ- ing Latin, Chinese, Arabic, Devanagari, Cyrillic and etc.,

“JaidedAI/EasyOCR: Ready-to-use OCR with 80+ sup- ported languages and all popular writing scripts includ- ing Latin, Chinese, Arabic, Devanagari, Cyrillic and etc.,”

work page
[24]

Medical Image De- Identification Resources: Synthetic DICOM Data and Tools for Validation,

Michael W. Rutherford et al., “Medical Image De- Identification Resources: Synthetic DICOM Data and Tools for Validation,” Aug. 2025, 1956 citations (Se- mantic Scholar/arXiv) [2025-10-17] arXiv:2508.01889 [cs]

work page arXiv 2025
[25]

Ultralytics YOLOv11,

Ultralytics, “Ultralytics YOLOv11,” 2024, https://docs.ultralytics.com/models/yolo11/ (accessed Oct. 15, 2025)

work page 2024
[26]

AI Innovation Platform,

Bayer AG, “AI Innovation Platform,” 2025, https://app.innovationplatform.ai/#/ (accessed Oct. 17, 2025)

work page 2025
[27]

Cloud Run: Build apps or web- sites quickly on a fully managed platform,

Google Cloud, “Cloud Run: Build apps or web- sites quickly on a fully managed platform,” 2025, https://cloud.google.com/run (accessed Oct. 17, 2025)

work page 2025
[28]

Cloud SQL: Cloud SQL for MySQL, PostgreSQL, and SQL Server,

Google Cloud, “Cloud SQL: Cloud SQL for MySQL, PostgreSQL, and SQL Server,” https://cloud.google.com/sql (accessed Oct. 17, 2025)

work page 2025
[29]

Vertex AI: Get batch inferences from a custom trained model,

Google Cloud, “Vertex AI: Get batch inferences from a custom trained model,” https://cloud.google.com/vertex- ai/docs/predictions/get-batch-predictions (accessed Oct. 17, 2025)

work page 2025

[1] [1]

Towards Selection of Large Multimodal Models as Engines for Burned-in Protected Health Information Detection in Medical Images

INTRODUCTION In recent years, the detection of Protected Health Informa- tion (PHI) has become increasingly important, particularly within the realm of medical data management. Traditional pipelines for PHI detection typically consist of Optical Char- acter Recognition (OCR) followed by analysis steps [1, 2, 3, 4, 5]. However, the advent of Large Multimod...

work page internal anchor Pith review Pith/arXiv arXiv 2025

[2] [2]

The text local- ization module identifies text areas within images, while the text extraction module serves as an OCR engine, converting pixel-level text into machine-encoded text

METHODS In this study, we employ the PHI pipeline put forward by [7], which comprises three integral components: text localization, text extraction, and text analysis (Figure 1). The text local- ization module identifies text areas within images, while the text extraction module serves as an OCR engine, converting pixel-level text into machine-encoded tex...

work page

[3] [3]

<0> Patient name: John Doe </0>

EXPERIMENTS 3.1. Data We use two datasetsRadPHI-testandMIDIwhich were introduced in [7] and are publicly available 1. These datasets cover a range of imaging modalities, imprint variations, and text complexities to support robust benchmarking. RadPHI- test (Figure 2a) includes 1000 radiological images equally distributed across four modalities with synthe...

work page doi:10.5281/zenodo.17201610 2024

[4] [4]

RadPHI-test Table 3 presents a summary of the performance and latency metrics for three benchmark LMMs evaluated on the RadPHI- test dataset

RESULTS 4.1. RadPHI-test Table 3 presents a summary of the performance and latency metrics for three benchmark LMMs evaluated on the RadPHI- test dataset. Among the models, GPT-4o achieves the highest overall precision and recall, both exceeding 0.99, followed by Gemini 2.5 Flash and Qwen 2.5-VL 7B. The performance of the models is generally higher when e...

work page arXiv

[5] [5]

LMM for Text Extraction The evaluation of OCR capabilities indicates that LMMs gen- erally outperform traditional OCR models (Table 1), although trade-offs depend on the dataset

DISCUSSION 5.1. LMM for Text Extraction The evaluation of OCR capabilities indicates that LMMs gen- erally outperform traditional OCR models (Table 1), although trade-offs depend on the dataset. On the RadPHI-test dataset, featuring complex and partially obscured imprints, the LMM- based approach (Setup B) shows slight improvements but in- curs a 20% to 4...

work page

[6] [6]

The paper emphasizes the nuanced trade-offs between accuracy, latency, and pri- vacy, providing practical recommendations for deployment in healthcare settings

CONCLUSION In this study, we analyze the strengths and limitations of state-of-the-art models like GPT-4o, Gemini 2.5 Flash, and Qwen2.5-VL 7B under two different configurations of the LMM-based PHI detection pipeline. The paper emphasizes the nuanced trade-offs between accuracy, latency, and pri- vacy, providing practical recommendations for deployment i...

work page

[7] [7]

Eth- ical approval was not required, as confirmed by the license attached to the sub-datasets

COMPLIANCE WITH ETHICAL STANDARDS This research study was conducted retrospectively using hu- man subject data derived from subsets of open-access data, which was introduced and published in the work of [7]. Eth- ical approval was not required, as confirmed by the license attached to the sub-datasets

work page

[8] [8]

ACKNOWLEDGMENTS The authors would like to thank the Bayer team of the AI Innovation Platform for providing computing infrastructure and technical support

work page

[9] [9]

Identification and classification of DICOM files with burned-in text content,

Petr Vcelak, Martin Kryl, Michal Kratochvil, and Jana Kleckova, “Identification and classification of DICOM files with burned-in text content,”International Jour- nal of Medical Informatics, vol. 126, pp. 128–137, June 2019

work page 2019

[10] [10]

A Method for Efficient De- identification of DICOM Metadata and Burned-in Pixel Text,

Jacob A. Macdonald, Katelyn R. Morgan, Brandon Konkel, Kulsoom Abdullah, Mark Martin, Cory En- nis, Joseph Y . Lo, Marissa Stroo, Denise C. Snyder, and Mustafa R. Bashir, “A Method for Efficient De- identification of DICOM Metadata and Burned-in Pixel Text,”Journal of Imaging Informatics in Medicine, vol. 37, no. 5, pp. 1–7, Oct. 2024

work page 2024

[11] [11]

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 Medi- cal Systems, vol. 41, no. 5, pp. 89, Apr. 2017

work page 2017

[12] [12]

De-Identification of Medical Imaging Data: A Comprehensive Tool for En- suring Patient Privacy,

Moritz Rempe, Lukas Heine, Constantin Seibold, Fabian H¨orst, and Jens Kleesiek, “De-Identification of Medical Imaging Data: A Comprehensive Tool for En- suring Patient Privacy,” Oct. 2024, arXiv:2410.12402

work page arXiv 2024

[13] [13]

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

Kyle Naddeo, Nikolas Koutsoubis, Rahul Krish, Ghu- lam Rasool, Nidhal Bouaynaya, Tony OSullivan, and Raj Krish, “DICOM De-Identification via Hybrid AI and Rule-Based Framework for Scalable, Uncertainty- Aware Redaction,” July 2025

work page 2025

[14] [14]

Evaluation of Vision- Language Models for Detection and Deidentification of Medical Images with Burned-In Protected Health Infor- mation,

Taehee Lee, Hyungjin Kim, Seong Ho Park, Seonhye Chae, and Soon Ho Yoon, “Evaluation of Vision- Language Models for Detection and Deidentification of Medical Images with Burned-In Protected Health Infor- mation,”Radiology, vol. 315, no. 3, pp. e243664, June 2025

work page 2025

[15] [15]

Exploring AI-Based Sys- tem Design for Pixel-Level Protected Health Informa- tion Detection in Medical Images,

Tuan Truong, Ivo M. Baltruschat, Mark Klemens, Grit Werner, and Matthias Lenga, “Exploring AI-Based Sys- tem Design for Pixel-Level Protected Health Informa- tion Detection in Medical Images,”Journal of Imaging Informatics in Medicine, July 2025

work page 2025

[16] [16]

David Clunie et al., “Summary of the National Cancer Institute 2023 Virtual Workshop on Medical Image De- identification-Part 2: Pathology Whole Slide Image De- identification, De-facing, the Role of AI in Image De- identification, and the NCI MIDI Datasets and Pipeline,” Journal of Imaging Informatics in Medicine, July 2024

work page 2023

[17] [17]

GPT-4o System Card

OpenAI et al., “GPT-4o System Card,” Oct. 2024, arXiv:2410.21276

work page internal anchor Pith review Pith/arXiv arXiv 2024

[18] [18]

Gemini 2.5: Pushing the Frontier with Advanced Reasoning,

Google DeepMind and Google Research, “Gemini 2.5: Pushing the Frontier with Advanced Reasoning,” 2025

work page 2025

[19] [19]

Qwen2.5-VL Technical Report

Alibaba Cloud / QwenLM, “Qwen2.5-VL Technical Re- port,” 2025, arXiv:2502.13923

work page internal anchor Pith review Pith/arXiv arXiv 2025

[20] [20]

Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling,

Zhe Chen et al., “Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling,” Sept. 2025

work page 2025

[21] [21]

TrOCR: Transformer-based Optical Charac- ter Recognition with Pre-trained Models,

Minghao Li, Tengchao Lv, Jingye Chen, Lei Cui, Yi- juan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, and Furu Wei, “TrOCR: Transformer-based Optical Charac- ter Recognition with Pre-trained Models,” 2021, eprint: 2109.10282

work page arXiv 2021

[22] [22]

General OCR Theory: Towards OCR-2.0 via a Unified End-to-end Model

Haoran Wei et al., “General OCR Theory: To- wards OCR-2.0 via a Unified End-to-end Model,” 2024, eprint: 2409.01704

work page internal anchor Pith review Pith/arXiv arXiv 2024

[23] [23]

JaidedAI/EasyOCR: Ready-to-use OCR with 80+ sup- ported languages and all popular writing scripts includ- ing Latin, Chinese, Arabic, Devanagari, Cyrillic and etc.,

“JaidedAI/EasyOCR: Ready-to-use OCR with 80+ sup- ported languages and all popular writing scripts includ- ing Latin, Chinese, Arabic, Devanagari, Cyrillic and etc.,”

work page

[24] [24]

Medical Image De- Identification Resources: Synthetic DICOM Data and Tools for Validation,

Michael W. Rutherford et al., “Medical Image De- Identification Resources: Synthetic DICOM Data and Tools for Validation,” Aug. 2025, 1956 citations (Se- mantic Scholar/arXiv) [2025-10-17] arXiv:2508.01889 [cs]

work page arXiv 2025

[25] [25]

Ultralytics YOLOv11,

Ultralytics, “Ultralytics YOLOv11,” 2024, https://docs.ultralytics.com/models/yolo11/ (accessed Oct. 15, 2025)

work page 2024

[26] [26]

AI Innovation Platform,

Bayer AG, “AI Innovation Platform,” 2025, https://app.innovationplatform.ai/#/ (accessed Oct. 17, 2025)

work page 2025

[27] [27]

Cloud Run: Build apps or web- sites quickly on a fully managed platform,

Google Cloud, “Cloud Run: Build apps or web- sites quickly on a fully managed platform,” 2025, https://cloud.google.com/run (accessed Oct. 17, 2025)

work page 2025

[28] [28]

Cloud SQL: Cloud SQL for MySQL, PostgreSQL, and SQL Server,

Google Cloud, “Cloud SQL: Cloud SQL for MySQL, PostgreSQL, and SQL Server,” https://cloud.google.com/sql (accessed Oct. 17, 2025)

work page 2025

[29] [29]

Vertex AI: Get batch inferences from a custom trained model,

Google Cloud, “Vertex AI: Get batch inferences from a custom trained model,” https://cloud.google.com/vertex- ai/docs/predictions/get-batch-predictions (accessed Oct. 17, 2025)

work page 2025