A method for processing sensitive data secured by a trusted third party and a set of sensitive data processing tools adapted for implementing such a method.

Abstract

The present invention relates to a method for processing sensitive data, particularly biomedical images, securely, automatically, and reproducibly on a cloud computing infrastructure. The invention also discloses the device for implementing this method. The invention relies in particular on cloud computing, cryptography, biomedical imaging, pseudonymization, anonymization, and advanced signal and image processing technologies. The invention also covers a use case for such a method through the secure implementation of image processing technologies (a business application) applied to biomedical images. In one embodiment, these images are obtained from magnetic resonance imaging (MRI), specifically for applying advanced processing with the business application to map the apparent transverse relaxation rate (R2*) and perform quantitative susceptibility imaging (QSM).Figure for the abbreviation: figure 1.

Description

Title of the invention: Method for processing sensitive data secured by a trusted third party and set of sensitive data processing equipment adapted for implementing such a method. Technical field

[0001] The invention relates to the field of processing sensitive data, particularly remotely.

[0002] The invention relates to this purpose to a method for processing sensitive data secured by a trusted third party and a set of sensitive data processing adapted for implementing such a method. Prior art

[0003] The context is that of securing the processing of sensitive data, particularly of the "biomedical images" type, and especially in the cloud within the framework of large-scale evaluations, for example for scientific research, clinical research, or diagnostic purposes. The problems that the invention aims to partially solve are as follows:

[0004] - The digitization of health data is increasing. This is a guarantee for the This allows individuals to receive medical monitoring based on complete and accurate information. It offers the possibility of streamlining healthcare spending. It is also a resource for scientific research. Whether in the creation of health data platforms for scientific research purposes, or directly within the framework of biomedical research, the data is intended to comply with the FAIR principles: findable, accessible, interoperable, and reusable.

[0005] - the GDPR, General Data Protection Regulation (DCP) requires a privacy impact assessment when processing sensitive personal data, necessitating the implementation of technical measures to limit the impact on individuals, such as pseudonymization (reversible) or anonymization (irreversible). These measures involve the use of encryption techniques.

[0006] - the development of quantum computers makes it possible to anticipate obsolescence In the coming decades, encryption methods will be state-of-the-art. The first quantum computing services are already accessible online. In this context, a leak of encrypted data can have a long-term impact that must be considered. So-called "post-quantum" encryption techniques are anticipated as disruptive solutions. They are considered more difficult to break and enabling calculations on encrypted data, these recent methods have good characteristics to meet the need for enhanced confidentiality in the long term, while allowing processing by a third party that does not offer all the necessary security guarantees.

[0007] - Existing processes and devices put in place to pseudonymize and / or Anonymizing digital patient data (DPD) in biomedical image processing can be improved. The DICOM format is the most commonly used standard. It contains descriptive identifying metadata as well as image data. Image data can be considered biometric data. The most obvious example concerns brain images, for which surface rendering allows for indirect identification. The shape of cerebral sulci, the volume of brain structures, and other anatomical characteristics are specific to the individual and can, through cross-referencing, enable identification. Large databases including brain images are currently being developed. These sometimes combine other sensitive information (genetic, medical history, etc.). The protection of this sensitive data, and in particular the images, needs to be strengthened.

[0008] - Major incidents resulting in massive leaks of patient data are have become frequent. The vulnerability of information systems, particularly in hospitals, is now being brought to the public's attention through numerous examples leading to the circulation of sensitive information collected illegally or as a result of negligence.

[0009] - Cloud computing is a strategic growth path with challenges Cybersecurity is becoming increasingly important. The creation of spaces for sharing and pooling data, particularly in healthcare, is developing, requiring a high degree of innovation in information security as well as in the services offered. This presents an opportunity to reduce the cost and time of evaluating new biomedical image processing solutions, for example, for diagnostic support, thereby accelerating the market launch of innovative medical devices.

[0010] There is therefore a need for advanced technologies to address a problem insufficiently covered by the prior art in order to provide guarantees on the processing of sensitive data.

[0011] Furthermore, an increasing number of advanced algorithms are being developed for biomedical image processing, with significant recent developments in artificial intelligence-based approaches. Some of these algorithms are used occasionally by users to test or compare their effectiveness. For example, in biomedical research, it is often necessary to compare the results of an algorithm under development with the results of existing algorithms. state-of-the-art algorithms with similar objectives exist. Access to these state-of-the-art algorithms is sometimes complex and time-consuming.

[0012] Indeed, in practice there are, for example, several ways of proceeding: • either the source code, object code, or executables are available or made available, • either the data to be processed is transmitted so that the actors able to apply a particular algorithm can do so on the data of a third party.

[0013] Each of these methods has disadvantages: • If source code, object code, or executables are made available, it is necessary to master their compilation and use, and there is no guarantee that the various computer components will be identical (libraries or the underlying operating system, for example). This potentially compromises the reproducibility of the results. Furthermore, it often requires installation on a dedicated workstation, with similar underlying reproducibility issues, and involves the maintenance of such physical systems, with the associated costs, expertise, and time commitment. • Data exchange between multiple actors is not always possible and is subject to confidentiality and anonymization restrictions governing the transfer and use of personal data, health data, or research data. The location of processing may be mandated by regulations, which restricts, slows down, or even, in some cases, makes it impossible to apply an algorithm to data.

[0014] Source code, object code, and executables performing the processing, on the one hand, and personal data, health data, and biomedical image data, on the other, are generally two categories of sensitive data that must be combined for the processing to be applied. One of the problems that the present invention seeks to solve is enabling these two categories of sensitive data to be combined.

[0015] It is common practice and known to those skilled in the art to use cloud computing technologies to perform specific image processing. This processing can, for example, be carried out on dedicated virtual servers. This type of solution eliminates the need for hardware installation and configuration, and therefore allows for more reproducible data processing by standardizing the configurations of the computer servers on which the processing is performed. However, the security requirements for such systems are significant and depend, for example, on the purpose and duration of the processing, and whether the data Data is retained, and for how long, for example. The use of cloud computing technology also optimizes the use of IT resources, allowing them to be allocated on demand only when needed. Finally, the location of processing can also be adapted. This flexibility in server location and on-demand deployment offers several advantages well-known to those in the field: reducing carbon footprint and meeting regulatory and security requirements.

[0016] However, the prior art in the field of cloud computing does not adequately meet the security and traceability requirements for processing sensitive data, particularly biomedical image data. The problem that the invention aims to solve is to provide a trusted third-party solution with a high level of security and guarantees. Description of the invention

[0017] The invention relates to this purpose a method for processing sensitive data securely, the sensitive data including a first part of data to be processed and a second part of data called informative, or metadata, the second part of data comprising a first portion of data called identifying data which are relating to a subject from which the first part of data originates and possibly a second portion of data called useful for processing the first part of data to be processed,

[0018] The treatment process comprising the following steps: A. Receipt of a request to process sensitive data by an initial IT unit, B. configuration by the first computer unit of a second computer unit capable of processing sensitive data, C. provision of sensitive data to a third-party IT unit, D. generation by the third computer unit of a signature relating to at least part of the sensitive data, E. extraction, by the third computer unit and from the sensitive data, of the first part of the data and the possible second portion of the data, A. concatenation of the first part of the data, the possible second part of the data, and the signature into a data container to form a first data container, B. encrypted transmission of the data container to the second computer unit, C. After retrieval and possible decryption, processing by the second computer unit of the first part of the data, this processing possibly taking into account the information present in the second portion of data, D. concatenation by the second computing unit of the processing results with at least the signature in a second data container, E. Transmission by the second computer unit of the second computer container in encrypted form to the third computer unit, F. verification by the third computer unit of the correspondence of the signature with the sensitive data and, association of the result of the processing with the first portion of data.

[0019] By computer unit, it is understood that a computer unit, such as a computer as such, a set of servers, a virtual machine running on one or more servers, or an application container running on one or more servers, independent of any other computer unit.

[0020] Thus, in a conceivable application of the invention, the first, second and third processing units can be provided in the form of software containers implemented on one or more servers. Brief description of the drawings

[0021] The present invention will be better understood upon reading the description of exemplary embodiments, given purely by way of illustration and in no way limiting, with reference to the accompanying drawings in which:

[0022] [Fig.1] illustrates a treatment assembly according to the invention.

[0023] [Fig.2] illustrates an example of sensitive data as processed within the framework of the invention.

[0024] Identical, similar or equivalent parts of the different figures bear the same numerical references so as to facilitate the transition from one figure to another.

[0025] The different parts represented in the figures are not necessarily shown on a uniform scale, in order to make the figures more legible.

[0026] The different possibilities (variants and embodiments) should be understood as not being mutually exclusive and can be combined with each other.

[0027] Detailed description of particular embodiments

[0028] The invention describes a method for deploying secure services, particularly in the cloud, for automatically processing biomedical images using Automated Specific Processing (ASP). The schematic diagram of one embodiment is shown [Fig. 1].

[0029] In this figure: 1. The customer accesses the service 2. A secure private server is deployed implementing access control principles, encryption in transit and at rest, and according to the client's specifications (regions of the world where the data is processed). 3. The client can transfer their data directly from their information system via a secure communication. 4. The server implements access control, quality control, where appropriate pseudonymization and anonymization, specific data encryption and sizes the resources necessary for TSA. 5. Resources are requested 6. Resources are made available for processing 7. The data is sent to the computing server 8. The TSA is performed by the dedicated computing server 9. The result of the processing sent 10. The interface server formats the results and re-identifies them as needed for consistency in the client's information system. 11. The result is delivered.

[0030] In one embodiment, the TSA is performed directly on the specific server. In another embodiment, the TSA is performed by the interface server.

[0031] - In one embodiment, one or more application containers is / are used to encapsulate the software elements necessary for the implementation of the TSA. In another embodiment, the software elements necessary for the implementation of the TSA are directly installed on the computing or interface server.

[0032] The method and device are characterized in that the technical safety measures include:

[0033] - Encryption at rest, for example by implementing encryption in data containers.

[0034] - Encryption of exchanges by secure communication of the TLS / SSL type by Certificate generation by a trusted third party, and / or SSH via the generation of secure key pairs by a trusted third party, and / or via the transfer of an encrypted container. The trusted third party providing the key and certificate is, for example, a service installed on the supervisor server or on another dedicated secure server with specific access control.

[0035] - In one embodiment, to enhance security, communications are encrypted using hybrid encryption protocols including classical and post-quantum encryption.

[0036] - The level of asymmetric encryption is adapted to the criticality of the data, by example with 4096-bit RSA encryption.

[0037] - In one embodiment of the invention, the generation of encryption keys This method can be based on random numbers generated by a quantum computer. Maximizing entropy is common practice when generating encryption keys, whether for SSH or AES encryption. To increase entropy, seed files containing a series of random bits can be generated. In one embodiment of the invention, a seed file is obtained using a quantum processor with at least one qubit. For a key of size Nk, a quantum processor with Nq qubits (Nq > 1) is initialized in a state where the Nq qubits are in equiprobable superposition states. A measurement is then performed, randomly assigning a value of 0 or 1 to each qubit. This process is then repeated N times, where N is an integer greater than or equal to Nk / Nq.

[0038] - Access control is performed. An intrusion detection system is installed. and monitored. A firewall filters incoming and outgoing connections and the ports on which they are made by defining a list of allowed IP addresses. In one embodiment, the default port used for SSH communication is changed to enhance security. In one embodiment, the servers are equipped with an intrusion detection system. In one embodiment, the list of users accessing remote resources is limited to unprivileged users.

[0039] The invention is based on the analysis of the Automated Specific Image Processing subsequently performed:

[0040] - The minimization of DICOM image metadata is based on the analysis of Information strictly necessary for TSA is retained. The input DICOM data is converted to perform this operation. Unnecessary metadata is discarded. Directly identifying metadata is discarded. Identifying information in the dataset is replaced with a signature or token. Digital metadata required for TSA is retained. In one embodiment, and if this measure is suitable for reducing the risk of indirect identification, the digital metadata can be encrypted using a homomorphic encryption algorithm that allows direct processing of the ciphertext.

[0041] - In one embodiment, if the TSA allows it, the image data will be be encrypted using a homomorphic or even completely homomorphic encryption method.

[0042] - In one embodiment, the specific processing is implemented on one or multiple application containers.

[0043] - In one embodiment, the resulting image and / or digital data The data processed is encrypted using a homomorphic or fully homomorphic encryption method. In the following example, we consider information that cannot be disclosed individually, even to a trusted third party, but requires aggregating data from multiple sources. The trusted third party performs this aggregation according to instructions provided by a client. From the perspective of the information to be processed, this could be the age of a study participant, the dose of medication that participant receives, or any other variable given as a number on a finite scale (between Vmin and Vmax). In one embodiment of the invention, this variable is converted into an integer between 0 and Nv using a linear operation that maps the interval [Vmin, Vmax] to [0, Nv] for all participants. This converted variable is then encrypted using fully homomorphic encryption.The numerical variable is then transmitted to the trusted third party by the various sources. The trusted third party calculates the average of the numerical variables and provides the result. Because this result is encrypted, the trusted third party cannot interpret it. The aggregated, encrypted result can then be deciphered by the client, without them having had access to the information from the various sources.

[0044] - In one embodiment, the homomorphic encryption method used is based on the problem of factoring polynomials with integer coefficients for which two polynomials describing spinors and representing successive rotations are used to encrypt the data.

[0045] As shown in [Fig. 2], biomedical images output from biomedical imaging devices are mostly in DICOM (Digital Image Communication in Medicine) format. They generically contain the image data itself, as well as metadata including identifying information and technical metadata relating to the image data. Only a fraction of this metadata is strictly necessary to perform a particular processing operation. Indeed, a number of algorithms require only a few parameters, as will be seen later through examples. The image data itself may also contain the same type of classification (identifying, technical), which may overlap, and only a portion of which is strictly necessary for the intended processing.

[0046] TSA of the R2* brain mapping type and / or QSM for the characterization of brain lesions:

[0047] The TSAs performed on the images can be varied. To illustrate, in one embodiment of the invention, they involve the reconstruction of relaxation time mapping R2* and / or magnetic susceptibility mapping QSM. The TSA This can also be followed by automated analysis that calculates statistics (such as mean, standard deviation) in regions of interest derived from image segmentation, and then generates a summary report. This summary report can then be compared to statistics obtained from populations of healthy subjects.

[0048] Magnetic susceptibility has well-known effects in MRI. It causes a distortion of the magnetic field and shortens the effective lifetime of the nuclear magnetic resonance signal. The extraction of synthetic quantitative parameters, namely R2* and QSM, is achievable using specific image processing techniques on the machine.

[0049] The acquisition is specific in order to be as sensitive as possible to the susceptibility of the imaging system on which the experiment is performed: • Typically, the imaging system operates at 1.5 Tesla, 3 Tesla, or 7 Tesla in humans, and with higher magnetic field values, up to 17.6 Tesla, in small animals to date. The nominal field of the imaging device is generally denoted by Bo. Susceptibility effects tend to increase with the nominal magnetic field Bo, meaning this metadata must be retained for processing that seeks to analyze susceptibility effects. • To generate an MRI signal, means are needed to excite the magnetization. These consist of a tuned radio frequency field emission device called an antenna or array of radio frequency transmitting antennas and capable of delivering an oscillating magnetic emission field denoted B1+. • To detect a signal, means of receiving the signal are required. These consist of a tuned radio frequency field receiving device called an antenna or array of radio frequency receiving antennas and a receiving oscillating magnetic field denoted B1-. • To convert the signal, sampling methods are required, generally using an analog-to-digital converter. To reconstruct and process the raw signal, specialized computer and processing resources are needed. • Imaging sequences correspond to the way the different elements of T-MRI are controlled. The static field B0 and the radiofrequency field B1+, as well as the reception and processing methods, have already been mentioned. Another major element, particularly for signal localization, consists of what are commonly called imaging gradients, which involve varying the value of the component along three axes (direction aligned with the field B0, as well as in two directions in a plane transverse to B0). according to Bo of the magnetic field. These methods are very classic and known to those skilled in the art. • To be sensitive to susceptibility effects, one can, for example, use a three-dimensional gradient echo imaging sequence with multiple echo times (TE). From this sequence, DICOM images can be reconstructed representing the value at each volume element (voxel) of the covered space, the average signal amplitude in these voxels as a function of TE, as well as an average phase map. For example, the parameters of the following imaging protocol are important for optimizing, analyzing, and addressing susceptibility effects: • The value of the Bo field • The type of imaging system used (Manufacturer, brand, model) • The name of the sequence and any variants or versions thereof • If a particular subsampling was performed, and which one • The type of image reconstruction used • The echo time (TE) value for each image • The acquisition bandwidth (BW) value for each picture • The echo encoding direction for each image • The orientation of the acquisition volume relative to Bo • The three-dimensional field of view covered and the voxel size • To analyze the multiple images using parameters that synthesize the information from the different echoes, parametric models are used. It is common to extract the apparent transverse relaxation time T2* from this type of acquisition, or equivalently its inverse R2*=l / T2*. In this case, it is assumed that the signal varies with the echo proportionally to an exponential decay of the type: I 0050 ! S œd (TE) = x eX p( - $ xTE)(D

[0051] Starting from this seemingly simple model, in reality a multitude of algorithms can be used to extract R2* for all voxels. To mention only the most common method, one can start with amplitude-only images (without phase) and adjust the reconstructed signals to the above model for each voxel using a least-squares minimization algorithm. For other algorithms based solely on the amplitude image, the reader can refer, for example, to this publication1. We can already see that the simple extraction of the parameter R2* leads to variability in algorithms, which is added to the variability mentioned above in their implementation on a given environment. Applying the FAIR principles in this case, with a view to traceability of the processing performed (here corresponding to the reconstruction of data derived from the image data initially provided by the imaging system), requires keeping track of more than just the algorithm used, but also of the computing environment on which this algorithm was applied. • A second parametric model synthesizing the information contained in images resulting from multi-echo acquisition has been described and is known as quantitative susceptibility imaging (QSM). The basic principles and physical models have been described for a long time and will not be repeated here.2 What is relevant to emphasize, however, is the complexity of the processing involved, which significantly increases the problem of reproducibility and traceability of such complex processes. To give an idea of ​​this complexity, the main steps of this processing are listed below (steps that follow one another, describing a pipeline): • Verification and loading of the acquisition parameters necessary for reconstruction, at a minimum the value of Bo or the system frequency, the observed kernel ('H by default in general), the orientation of the volume relative to Bo, the absolute or relative size of the voxels, the absolute or relative spacing between the different echo times, possibly the acquisition bandwidth and the direction of the reading gradient (because some sequence options allow acquisitions of multiple echoes with a reading gradient always in the same direction or alternately in one direction for even echoes, and in the other direction for odd echoes, an element well known to the person of art3). • Fitting a magnetic field map from phase information. Several processing methods are possible for this step, and the methods depend on the type of data available. In the most frequent case of images available for regularly spaced echo times (corresponding to an identical interval between the different echo times, preferably with reading gradients in the same direction), it is possible, for example, to synthesize a complex signal from the amplitude and phase measurements, then perform a Fourier transform heavily completed with zeros, and then determine the index of the spectral maximum. and to convert this index into a frequency. Other processing methods are possible, for example by combining a phase time unfolding followed by a least-squares weighted adjustment. Susceptibility effects are often very significant near the interfaces between low-density regions (such as those containing gases, the nose, sinuses, ear canals, digestive system, and pulmonary airways, for example) and tissues composed primarily of water. The resulting field distortions become predominant compared to internal tissue variations (differences of two to three orders of magnitude can be observed), which has led to the development of techniques for filtering these effects. Here too, numerous filtering techniques have been proposed, ranging from simple high-pass filtering (the reference method) to more sophisticated approaches based on the expected physical properties of these effects, particularly harmonic characteristics (solutions to Laplace's equations), leading to various filtering processes4 to extract only the so-called 'internal' effects. In this filtering phase, it is often necessary to define a volume of interest within which the internal field will be estimated. This requires defining and implementing an image segmentation process, for example, in the case of brain imaging to extract the brain. Next, these internal effects are fitted to a physical model to reconstruct a magnetic susceptibility map. Since field data alone is insufficient to provide a unique solution, it is almost always necessary to add other information, in line with what are called Bayesian approaches.2 In particular, it is possible to rely on information that is not based solely on field deformation, but also on anatomical information, whether this information is a priori spatially uniform or results from the processing of localized information derived from the signal amplitude, or from prior segmentation (see, for example,5). In these constrained reconstructions, regularization parameters are often introduced that give relative weight to the measurements compared to the priors, and here again the methods of The selection of these parameters is multifaceted, with, for example, a technique known to the art practitioner called "L-curve".

[0052] The example of R2* and QSM parametric imaging clearly illustrates the problem associated with the complexity of this type of processing: traceability of the processes is necessary for their reproduction, and guarantees must be provided for the data used in the processing. Currently, there is no technical solution that allows easy access to reproducible processes usable industrially on a large scale. This is precisely what the invention proposes to address through the given embodiment of R2* and QSM parametric imaging.

Claims

1. Demands A method for processing sensitive data securely, sensitive data including a first part of the data to be processed and a second part of the data called informative, or metadata, the second part of the data comprising a first portion of so-called identifying data which relates to a subject from whom the first part of the data originates and possibly a second portion of so-called useful data for processing the first part of the data to be processed, The treatment process includes the following steps: A. Receipt of a request to process sensitive data by an initial IT unit, B. configuration by the first computer unit of a second computer unit capable of processing sensitive data, C. provision of sensitive data to a third-party IT unit, D. generation by the third computer unit of a signature relating to at least part of the sensitive data, E. extraction, by the third computer unit and from the sensitive data, of the first part of the data and the possible second portion of the data, A. concatenation of the first part of the data, the possible second part of the data, and the signature into a data container to form a first data container, B. encrypted transmission of the data container to the second computer unit, C. After retrieval and possible decryption, processing by the second computer unit of the first part of the data, this processing possibly taking into account the information present in the second portion of the data, D. concatenation by the second computer unit of the results of the processing with at least the signature in a second data container, E. transmission by the second computer unit of the second computer container in encrypted form to the third computer unit, F. verification by the third computer unit of the correspondence of the signature with the sensitive data and, association of the result of the processing with the first portion of data.

2. Processing method according to claim 1 wherein prior to the concatenation step G, an anonymization step is provided for the first first part of data and / or the second part, said step consisting of modifying or deleting elements of said first part of data and / or the second part which are subject-specific and which are not useful for the processing.

3. A processing method according to claim 1 or 2, wherein during step H, of encrypted transmission of the data container, at least one of the first data portion and the second data portion is at least partially encrypted by a homomorphic encryption with respect to the processing, and wherein during step I, of processing by the second computer unit, said at least one of the first data portion and the second data portion is not previously decrypted,

4. Processing method according to claim 1 to 3, wherein at least one of the transmission steps H. and K., the transmitted container is encrypted by a TLS / SSL transfer protocol by generation of certificates by a trusted third party, and / or SSH by generation of secure key pairs by a trusted third party, and / or via the transfer of an encrypted container.

5. Processing method according to any one of claims 1 to 5, wherein the sensitive data is in DICOM format and includes at least one imaging data, said at least one imaging being obtained by magnetic resonance imaging.

6. A treatment method according to any one of claims 1 to 5 further comprising the following step: A. After step K. of transmitting the result, resetting or deleting the second computer unit by the first computer unit.

7. A processing method according to any one of claims 1 to 6, wherein the first computer unit is configured to keep a log of each step implemented by each of the second and third computer units.

8. Processing method according to any one of claims 1 to 8 wherein at least one of the second and third processing units is provided by an application container.

9. A set of data processing systems for the secure processing of sensitive data, the sensitive data including a first set of data to be processed and a second set of data referred to as informational data, or metadata, the second set of data comprising a first portion of data referred to as identifying data relating to a subject from whom the first set of data originates and possibly a second portion of data referred to as useful for processing the first set of data to be processed, said processing system comprising: - a first computer unit capable of receiving a request to process sensitive data and to be configured, a second computer unit capable of performing the processing of sensitive data, - a third computer unit capable of receiving sensitive data from a third computer unit, said third computer unit being configured to: • extract, from the sensitive data,the first part of the data and the possible second part of the data, • generate a signature relating to at least part of the sensitive data • concatenate the first part of the data, the possible second part of the data and the signature into a data container in order to form a first data container, • transmit, in encrypted form, a first data container to the second computer unit, - the second computer unit capable of: • After retrieval and possible decryption, process the first part of the data, this processing possibly taking into account the information present in the second part of the data. • Concatenate the processing results with at least the signature into a second container, • transmit data from the second container in encrypted form to the third computer unit, the third computer unit being further configured to verify the correspondence between the signature and the sensitive data and associate the result of the processing with the first portion of data

Images