Auswahl der wissenschaftlichen Literatur zum Thema „Medical image anonymization“

Abstract Interdisciplinary exchange of medical datasets between clinicians and engineers is essential for clinical research. Due to the Data Protection Act, which preserves the rights of patients, full anonymization is necessary before any exchange can take place. Due to the continuous improvement of image quality of tomographic datasets, anonymization of patient-specific information is not sufficient. In this work, we present a prototype that allows to reliably obscure the facial features of patient data, thus enabling anonymization of neurological datasets in image space.

Clinical data sharing can facilitate data-driven scientific research, allowing a broader range of questions to be addressed and thereby leading to greater understanding and innovation. However, sharing biomedical data can put sensitive personal information at risk. This is usually addressed by data anonymization, which is a slow and expensive process. An alternative to anonymization is construction of a synthetic dataset that behaves similar to the real clinical data but preserves patient privacy. As part of a collaboration between Novartis and the Oxford Big Data Institute, a synthetic dataset was generated based on images from COSENTYX® (secukinumab) ankylosing spondylitis (AS) clinical studies. An auxiliary classifier Generative Adversarial Network (ac-GAN) was trained to generate synthetic magnetic resonance images (MRIs) of vertebral units (VUs), conditioned on the VU location (cervical, thoracic and lumbar). Here, we present a method for generating a synthetic dataset and conduct an in-depth analysis on its properties along three key metrics: image fidelity, sample diversity and dataset privacy.

Local differential privacy algorithms combined with deep generative models can enhance secure medical image sharing among researchers in the public domain without central administrators; however, these images were limited to the generation of low-resolution images, which are very insufficient for diagnosis by medical doctors. To enhance the performance of deep generative models so that they can generate high-resolution medical images, we propose a large-scale diffusion model that can, for the first time, unconditionally generate high-resolution (256×256×256) volumetric medical images (head magnetic resonance images). This diffusion model has 19 billion parameters, but to make it easy to train it, we temporally divided the model into 200 submodels, each of which has 95 million parameters. Moreover, on the basis of this new diffusion model, we propose another formulation of image anonymization with which the processed images can satisfy provable Gaussian local differential privacy and with which we can generate images semantically different from the original image but belonging to the same class. We believe that the formulation of this new diffusion model and the implementation of local differential privacy algorithms combined with the diffusion models can contribute to the secure sharing of practical images upstream of data processing.

Purpose: The purpose of this article is to offer an overview of the recent applications of cloud computing in medical imaging and their advances in the field. It reviews existing scientific and academic literature on cloud computing-based platforms and solutions for medical image analysis and storage of medical imaging data. Materials and methods: This article uses available scientific literature on the applications of cloud computing in medical imaging from PubMed, Google Scholar and ScienceDirect. Results: The review shows that interest and research in cloud computing applications in medical imaging has increased in recent years. This has led to new and more effective platforms and solutions for medical image analysis and storage of medical imaging data. Innovative applications of cloud computing in medical imaging try to address ethical and security concerns using authentication, encryption, anonymization and access controls. Conclusions: Cloud computing offers promising applications in medical imaging. Notwithstanding, further research is necessary to demonstrate their effectiveness, safety and security in a medical setting.

Background High-resolution medical images that include facial regions can be used to recognize the subject’s face when reconstructing 3-dimensional (3D)-rendered images from 2-dimensional (2D) sequential images, which might constitute a risk of infringement of personal information when sharing data. According to the Health Insurance Portability and Accountability Act (HIPAA) privacy rules, full-face photographic images and any comparable image are direct identifiers and considered as protected health information. Moreover, the General Data Protection Regulation (GDPR) categorizes facial images as biometric data and stipulates that special restrictions should be placed on the processing of biometric data. Objective This study aimed to develop software that can remove the header information from Digital Imaging and Communications in Medicine (DICOM) format files and facial features (eyes, nose, and ears) at the 2D sliced-image level to anonymize personal information in medical images. Methods A total of 240 cranial magnetic resonance (MR) images were used to train the deep learning model (144, 48, and 48 for the training, validation, and test sets, respectively, from the Alzheimer's Disease Neuroimaging Initiative [ADNI] database). To overcome the small sample size problem, we used a data augmentation technique to create 576 images per epoch. We used attention-gated U-net for the basic structure of our deep learning model. To validate the performance of the software, we adapted an external test set comprising 100 cranial MR images from the Open Access Series of Imaging Studies (OASIS) database. Results The facial features (eyes, nose, and ears) were successfully detected and anonymized in both test sets (48 from ADNI and 100 from OASIS). Each result was manually validated in both the 2D image plane and the 3D-rendered images. Furthermore, the ADNI test set was verified using Microsoft Azure's face recognition artificial intelligence service. By adding a user interface, we developed and distributed (via GitHub) software named “Deface program” for medical images as an open-source project. Conclusions We developed deep learning–based software for the anonymization of MR images that distorts the eyes, nose, and ears to prevent facial identification of the subject in reconstructed 3D images. It could be used to share medical big data for secondary research while making both data providers and recipients compliant with the relevant privacy regulations.

A huge percentage of people especially in developed countries spend a good chunk of their wealth in managing their health conditions. In order to adequately administer healthcare, governments and various organizations have embraced advanced technology for automating the health industry. In recent past, electronic health records have largely been managed by Enterprise Resource Planning and legacy systems. Big data framework steadily emerge as the underlying technology in healthcare, which offers solutions that limits capacity of others systems in terms of storage and reporting. Automation through cloud services supported by storage of structured and unstructured health data in heterogeneous environment has improved service delivery, efficiency, medication, diagnosis, reporting and storage in healthcare. The argument support the idea that big data healthcare still face information security concern, for instance patient image sharing, authentication of patient, botnet, correlation attacks, man-in-the-middle, Distributed Denial of Service (DDoS), blockchain payment gateway, time complexities of algorithms, despite numerous studies conducted by scholars in security management for big data in smart healthcare. Some security technique include digital image encryption, steganography, biometrics, rule-based policy, prescriptive analysis, blockchain contact tracing, cloud security, MapReduce, machine-learning algorithms, anonymizations among others. However, most of these security solutions and analysis performed on structured and semi-structured data as opposed to unstructured data. This may affect the output of medical reporting of patients’ condition particularly on wearable devices and other examinations such as computerized tomography (CT) Scans among others. A major concern is how to identify inherent security vulnerabilities in big healthcare, which generate images for transmission and storage. Therefore, this paper conducted a comparative survey of solutions that specifically safeguards structured and unstructured data using systems that run on big data frameworks. The literature highlights several security advancements in cryptography, machine learning, anonymization and protocols. Most of these security frameworks lacks implementation evidence. A number of studies did not provide comprehensive performance metrics (accuracy, error, recall, precision) of the models besides using a single algorithm without validated justification. Therefore, a critique on the contribution, performance and areas of improvements discussed and summarized in the paper.

Abstract The use of patient data for artificial intelligence (AI) research should be transparent, rigorous and accountable. In the UK, the General Data Protection Regulation, Data Protection Act 2018 and General Medical Council govern data handling and patients’ rights to privacy. We report on our multistep pseudonymization protocol for real-world skin lesion datasets, in preparation for research within a trusted research environment (TRE). Firstly, patients referred from primary care are triaged for community locality and imaging centre (CLIC) suitability. There, trained healthcare professionals capture lesion images (dermoscopic, macroscopic and regional) and patient information using a mobile application on trust-certified devices. Training is standardized across all CLIC sites, with specific anonymization training on removing in-frame clothing and jewellery, device positioning, and magnification to minimize identifiable features like eyes, nose and ears. Lesion datasets (paired images and clinical information) are subsequently transferred to an image management system (IMS) hosted on our trust-secured network. Within the IMS, images are manually inspected, and those with identifiable tattoos and piercings are excluded. All regional images are also excluded from transfer to the TRE. Before transfer to the TRE, images undergo a further round of review. Data fields are manually checked for identifiable patient information, patient names are removed, and dates of birth are rounded to 3-month granularity. The job ID, patient’s hospital number, date of clinical episode and responsible photographer are replaced with randomly generated project-specific identifiers. In an initial study period, 658 of 963 (68%) captured lesion datasets have undergone IMS manual inspection. Of these, 24 lesion datasets were excluded for identifiable features, 10 (41%) for more than one-third of the face being visible, 9 (38%) for full iris visibility, and 5 (21%) for tattoos. On breakdown by anatomical location these images were of the face (19, 80%), torso (2, 8%), limbs (2, 8%) and neck (1, 4%). The remaining 634 datasets (96%) were securely transferred to the TRE, where a further 5% were excluded due to potential identifiability. Although full anonymization is desirable, it is usually achieved by aggregating patient data. Pseudonymization, which allows for future reidentification in a secured fashion, strikes the balance between patient data privacy and clinical governance, while retaining a level of granularity sufficient for meaningful analysis. Currently, this protocol is manually intensive with room to partly automate. Use of common standardized protocols will strengthen the public trust in clinical AI.

Due to the rapid development of information technologies, informatization in the medical industry is essential. The main component of electronic health care is medical information systems designed for the accumulation, processing, analysis and transmis-sion of medical data. In the medical field, specialized software products are used to per-form diagnostic studies, process the results of laboratory tests, and make decisions at the stage of establishing a diagnosis. The use of mobile devices in medical information systems is developing. However, the degree of automation of processes in the provision of medical services and the protection of the personal and medical data of patients is still insufficient. The purpose of the research is to create a basic architecture of a software system that would simplify the process of developing software for automated input, processing, search and confidential patient access to their medical data in a medical information system based on multi-color barcoding of information using mobile devices. The architecture of the software system is proposed, in which, based on the princi-ples of distribution, anonymization, and data ownership, a patient can provide access to medical personnel to their medical data by reading a multi-color interference-resistant barcode from one smartphone (patient’s) by the camera of another smartphone (doctor’s). It is shown that in order to ensure the reliability of such transmission, it is neces-sary to use an interference-resistant barcode, which would ensure the integrity of the data in the conditions of possible distortion of the barcode image (change in lighting, scanning angle, trembling of the operator's hand, blurring or skewing of the image, etc.). The use of mobile devices for the barcode method of transmission and processing of data allows providing the protected electronic co-operating of a patient and a doctor both directly and remotely. It guarantees high reliability and confidentiality of the ex-change of data. The proposed technical solutions make it possible to improve the quality of medi-cal care and strengthen the protection of the patient's medical data.

Abstract Objectives: According to the World Health Organization (WHO), breast cancer is the most frequent malignant neoplasia and leading cause of cancer death among women worldwide. Low- and middle-income countries hold the worst survival rates mainly owing to a lack of access to appropriate diagnosis and treatment related resources. For proper early diagnosis, it is established that besides the physical structure itself (e.g., mammography units), there's a need for adequate interpretation of imaging and that might be a particularly major problem in low-income societies once there is a tendency of greater education setbacks. Mammography datasets can improve this resource-driven gap by enabling the development of artificial intelligence technologies (AI) which can make breast cancer diagnosis more accurate in a cost-effective and scalable way. We aim to create a new database of high quality digital mammography images suitable for AI development and education. Methods: Our mammography database was developed by means of retrospective selection of 100 exams performed by Hospital São Paulo - Federal University of São Paulo ranging from 2019 to 2023. The project is assumed to be safe, versatile, and usable, and required an extensive search for the appropriate tool. Ambra Health, an American company, has developed cloud-based software for medical image management and stood out as a viable alternative. Their platform meets international data security criteria, they also made the intended careful customization possible, in addition to the possibility of associating image and text attachments. The categories were created in accordance with the BI-RADS® descriptors, a wide range of clinical scenarios and additional materials available, and they served as the basis for the advanced search feature, which intuitively filters exams that meet the selected criteria simultaneously. The platform was integrated with an automatic anonymization system upon upload, ensuring data privacy. After submission, the exams are retained in a restricted area for anonymization verification, categorization, and attachment management, before being released to the end-user. So as to broaden geographic coverage, the descriptors were entered in American English, respecting the origin of the BI-RADS® lexicon, as for the website structure, automatic translation to the accessing browser standard language was selected. Results: Our website is active and available at http://mamografia.unifesp.br, with access granted upon a simple registration process. 941 mammography images from 100 anonymized cases, 62% of which include 3D images, can be filtered based on the combination of 113 clinical and imaging variables, as well as attachment availability. The language is adaptable to the user's native language, and categorized searches can be accessed directly from the browser or downloaded as customized datasets. Additionally, features such as saved searches or starred exams are also available. Conclusion: We have developed an online and free mammography database that is completely innovative by integrating various resources into a single platform. We provide high-resolution and 3D digital images that can be searched using an advanced search system. Moreover, we offer supplementary clinical information in various attachment formats, favoring a rich clinical correlation. In this way, we have achieved the ambivalence of our goal, which was to promote education and research. *"images speak louder than words" Database: https://mamografia.unifesp.br Tutorials: https://www.youtube.com/@Mamografiaunifesp e-mail: acesso@mamografia.unifesp.br password: acesso@mamografia12 (valid until dec/23) Citation Format: Natália Cordeiro, Gil Facina, Afonso Nazário, Vanessa Sanvido, Joaquim Araujo Neto, Morgana Silva, Simone Elias. Towards Precision Medicine in Breast Imaging: A Novel Open Mammography Database with Tailor-Made 3D Image Retrieval for Artificial Intelligence and Teaching [abstract]. In: Proceedings of the 2023 San Antonio Breast Cancer Symposium; 2023 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2024;84(9 Suppl):Abstract nr PO2-29-02.

Cette thèse aborde plusieurs problèmes de sécurité et de confidentialité lors du traitement d'images médicales dans des lacs de données. Ainsi, on explore la fuite potentielle de données lors de l'exportation de modèles d'intelligence artificielle, puis on développe une approche d'anonymisation d'images médicales qui protège la confidentialité des données. Le Chapitre2 présente une nouvelle attaque d'exfiltration de données, appelée Data Exfiltration by Compression (DEC), qui s'appuie sur les techniques de compression d'images. Cette attaque est effectuée lors de l'exportation d'un réseau de neurones entraîné au sein d'un lac de données distant et elle est applicable indépendamment de la tâche de traitement d'images considérée. En explorant à la fois les méthodes de compression sans perte et avec perte, ce chapitre montre comment l'attaque DEC peut être utilisée efficacement pour voler des images médicales et les reconstruire avec une grande fidélité, grâce à l'utilisation de deux ensembles de données CT et IRM publics. Ce chapitre explore également les contre-mesures qu'un propriétaire de données peut mettre en œuvre pour empêcher l'attaque. Il étudie d'abord l'ajout de bruit gaussien pour atténuer cette attaque, et explore comment les attaquants peuvent créer des attaques résilientes à cet ajout. Enfin, une stratégie alternative d'exportation est proposée, qui implique un réglage fin du modèle et une vérification du code. Le Chapitre 3 présente une méthode d'anonymisation d'images médicales par approche générative, une nouvelle approche pour trouver un compromis entre la préservation de la confidentialité des patients et l'utilité des images générées pour résoudre les tâches de traitement d'images. Cette méthode sépare le processus d'anonymisation en deux étapes : tout d'abord, il extrait les caractéristiques liées à l'identité des patients et à l'utilité des images médicales à l'aide d'encodeurs spécialement entrainés ; ensuite, il optimise le code latent pour atteindre le compromis souhaité entre l'anonymisation et l'utilité de l'image. Nous utilisons des encodeurs d'identité, d'utilité et un encodeur automatique génératif basé sur un réseau antagoniste pour créer des images synthétiques réalistes à partir de l'espace latent. Lors de l'optimisation, nous incorporons ces encodeurs dans de nouvelles fonctions de perte pour produire des images qui suppriment les caractéristiques liées à l'identité tout en conservant leur utilité pour résoudre un problème de classification. L'efficacité de cette approche est démontrée par des expériences sur l'ensemble de données de radiographie thoracique MIMIC-CXR, où les images générées permettent avec succès la détection de pathologies pulmonaires. Le Chapitre 4 s'appuie sur les travaux du Chapitre 3 en utilisant des réseaux antagonistes génératifs (GAN) pour créer une solution d'anonymisation plus robuste et évolutive. Le cadre est structuré en deux étapes distinctes : tout d'abord, nous développons un encodeur simplifié et un nouvel algorithme d'entraînement pour plonger chaque image dans un espace latent. Dans la deuxième étape, nous minimisons les fonctions de perte proposées dans le Chapitre 3 pour optimiser la représentation latente de chaque image. Cette méthode garantit que les images générées suppriment efficacement certaines caractéristiques identifiables tout en conservant des informations diagnostiques cruciales. Des expériences qualitatives et quantitatives sur l'ensemble de données MIMIC-CXR démontrent que notre approche produit des images anonymisées de haute qualité qui conservent les détails diagnostiques essentiels, ce qui les rend bien adaptées à la formation de modèles d'apprentissage automatique dans la classification des pathologies pulmonaires. Le chapitre de conclusion résume les contributions scientifiques de ce travail et aborde les problèmes et défis restants pour produire des données médicales sensibles, sécurisées et préservant leur confidentialité
This thesis aims to address some specific safety and privacy issues when dealing with sensitive medical images within data lakes. This is done by first exploring potential data leakage when exporting machine learning models and then by developing an anonymization approach that protects data privacy.Chapter 2 presents a novel data exfiltration attack, termed Data Exfiltration by Compression (DEC), which leverages image compression techniques to exploit vulnerabilities in the model exporting process. This attack is performed when exporting a trained network from a remote data lake, and is applicable independently of the considered image processing task. By exploring both lossless and lossy compression methods, this chapter demonstrates how DEC can effectively be used to steal medical images and reconstruct them with high fidelity, using two public CT and MR datasets. This chapter also explores mitigation measures that a data owner can implement to prevent the attack. It first investigates the application of differential privacy measures, such as Gaussian noise addition, to mitigate this attack, and explores how attackers can create attacks resilient to differential privacy. Finally, an alternative model export strategy is proposed which involves model fine-tuning and code verification.Chapter 3 introduces the Generative Medical Image Anonymization framework, a novel approach to balance the trade-off between preserving patient privacy while maintaining the utility of the generated images to solve downstream tasks. The framework separates the anonymization process into two key stages: first, it extracts identity and utility-related features from medical images using specially trained encoders; then, it optimizes the latent code to achieve the desired trade-off between anonymity and utility. We employ identity and utility encoders to verify patient identities and detect pathologies, and use a generative adversarial network-based auto-encoder to create realistic synthetic images from the latent space. During optimization, we incorporate these encoders into novel loss functions to produce images that remove identity-related features while maintaining their utility to solve a classification problem. The effectiveness of this approach is demonstrated through extensive experiments on the MIMIC-CXR chest X-ray dataset, where the generated images successfully support lung pathology detection.Chapter 4 builds upon the work from Chapter 4 by utilizing generative adversarial networks (GANs) to create a more robust and scalable anonymization solution. The framework is structured into two distinct stages: first, we develop a streamlined encoder and a novel training scheme to map images into a latent space. In the second stage, we minimize the dual-loss functions proposed in Chapter 3 to optimize the latent representation of each image. This method ensures that the generated images effectively remove some identifiable features while retaining crucial diagnostic information. Extensive qualitative and quantitative experiments on the MIMIC-CXR dataset demonstrate that our approach produces high-quality anonymized images that maintain essential diagnostic details, making them well-suited for training machine learning models in lung pathology classification.The conclusion chapter summarizes the scientific contributions of this work, and addresses remaining issues and challenges for producing secured and privacy preserving sensitive medical data

Several fundamental rights are inherently tied to human personality, including the right to one's name, the right to correspond through letters, the right to physical integrity, the moral rights of authors, and, notably for this discussion, the right to one's image, the right to honor, and the “right to confidentiality.” These rights may even hold constitutional significance, as the Constitutional Council asserts that they stem from the “protection of individual freedom.” Notably, laws that breach privacy in scenarios such as vehicle searches, tax investigations, or the publication of certain tax information have been ruled unconstitutional. In this chapter, the authors explore various techniques for safeguarding privacy, focusing specifically on fundamental techniques applicable in the medical sector, such as differential privacy, secure multiparty computation protocols, data anonymization, deidentification, and fingerprint privacy, among others.

In this study, StyleGAN2 was trained with panoramic radiographs, and original images were projected into the latent space of StyleGAN2. The resulting latent vectors were input into StyleGAN2, and corresponding images were generated to reconstruct the original images. The original and reconstructed images were evaluated by pediatric dentists and found to be similar. Our results suggest that StyleGAN2 could be applied to the anonymization and data compression of medical images.

A number of fundamental rights are inseparable from human personality. These are in particular the right to the name, the right to missive letters, the right to physical integrity, the moral right of the author and, in the field which interests us, the right to the image, the right to honor, or the “right to secrecy”. These rights even have constitutional value. The Constitutional Council considers that they proceed from “respect for individual freedom.” In particular, laws that violate the search of vehicles, tax searches, and the publication of certain elements of tax declarations, have been declared unconstitutional. In this chapter the authors will talk about the different techniques that can be used for the preservation of privacy. This chapter will illustrate an overview of the different fundamental techniques for the preservation of privacy in medical sector such as: differential privacy, secure multiparty computation protocol, data anonymization, deidentification, fingerprint privacy, and others.

Data clustering has been an integral and important part of data mining. It has wide applications in database anonymization, decision making, image processing and pattern recognition, medical diagnosis and geographical information systems, only to name a few. Data in real life scenario are having imprecision inherent in them. So, early crisp clustering techniques are very less efficient. Several imprecision based models have been proposed over the years like the fuzzy sets, rough sets, intuitionistic fuzzy sets and many of their generalized versions. Of late, it has been established that the hybrid models obtained as combination of these imprecise models are far more efficient than the individual ones. So, many clustering algorithms have been put forth using these hybrid models. The focus of this chapter is to discuss on some of the data clustering algorithms developed so far and their applications mainly in the area of decision making.