Privacy preserving data governance for federated learning

Abstract

Systems and methods facilitate privacy-preserving data governance in a federated learning system by transmitting a portion of a potential data sample to a remote location. The portion is inspected for quality to exclude data samples that do not meet data governance criteria. The remote inspection focuses on auditing areas of interest, but remains private because the inspection cannot resolve any other identifiable object information, such as faces, body shapes, etc., because pixels or voxels outside of the portion are not included. The results of the inspection are sent back to the collaborator so that during a federated learning round, inappropriate data samples can be excluded.

Description

Technical Field

[0001] This embodiment involves federated learning. Background Technology

[0002] In recent years, deep learning has expanded from medical research to healthcare products due to advancements in computing hardware and the surge in healthcare data. Numerous entities, such as research centers, hospitals, and other healthcare providers, collect vast amounts of data daily. Data analytics can improve learning models and user experiences. The complex problem of training these models can be solved by distributed computing, leveraging the resources of participating devices available at the network edge: storage, computing power, cycle time, content, and bandwidth. In such distributed machine learning scenarios, datasets are transmitted to or stored among multiple edge devices. These devices solve distributed optimization problems to collaboratively learn the underlying model. For distributed computing, similar (or identical) datasets can be distributed across multiple devices, enabling them to solve problems in parallel. However, accessing large, diverse healthcare datasets remains a challenge due to regulatory concerns about sharing protected healthcare information.

[0003] Privacy and connectivity concerns can prevent data sharing between entities, thus hindering large-scale distributed approaches. For example, hospitals may be unwilling or not permitted to share medical data with other entities or unknown users. Federated learning is a distributed computing approach that enables entities to collaborate on machine learning projects without sharing sensitive data, such as patient records, financial data, or confidential information. The fundamental premise behind federated learning is that the model moves to satisfy the data, not the other way around. Therefore, the minimum data movement required across the federation is the model parameters and their updates. However, challenges remain in managing data flow, model training, and privacy issues. Summary of the Invention

[0004] Systems, methods, and computer-readable media for data curation in federated learning processes are provided.

[0005] In a first aspect, a method is provided for training a model using federated learning by multiple collaborators, the method comprising: receiving global model parameters from a parameter aggregation server by one of the collaborators; acquiring sample data by the collaborator; transmitting anonymized portions of the sample data to a policy management server configured to identify data samples that satisfy or do not satisfy one or more data selection criteria; receiving verification of the sample data from the policy management server by the collaborator; training the model using the verified sample data by the collaborator; and transmitting local model parameters of the model to the parameter aggregation server by the collaborator.

[0006] In a second aspect, a method is provided for training a model using federated learning by multiple collaborators, the method comprising: a collaborator receiving global model parameters from a parameter aggregation server; the collaborator acquiring sample data; the collaborator submitting anonymized portions of the sample data to a machine-trained model, the machine-trained model being configured to identify data samples that do not meet data selection criteria; the collaborator receiving validation of the sample data from the machine-trained model; the collaborator training the model using the validated sample data; and the collaborator transmitting local model parameters of the model to the parameter aggregation server.

[0007] In a third aspect, a system for federated learning is provided. The system includes multiple collaborators, a policy management server, and an aggregation server. Each of the multiple collaborators is configured to acquire sample data, anonymize the sample data, and transmit the anonymized sample data to the policy management server for validation. Each collaborator is further configured to train a local machine learning model using the validated sample data, update the local model weights of the local machine learning model, and send the updated local model weights to the aggregation server. The policy management server is configured to receive the anonymized sample data and identify sample data that meets or does not meet one or more data selection criteria. The aggregation server is configured to receive the updated model weights from the multiple collaborators, calculate the aggregated model weights, and transmit the aggregated model weights to the multiple collaborators to update the local machine learning model.

[0008] Any one or more of the foregoing aspects may be used alone or in combination. These and other aspects, features, and advantages will become apparent from the following detailed description of preferred embodiments, which should be read in conjunction with the accompanying drawings. The invention is defined by the following claims, and nothing in this section should be construed as limiting those claims. Other aspects and advantages of the invention will be discussed below in conjunction with preferred embodiments and may be claimed thereafter, independently or in combination. Attached Figure Description

[0009] Components and figures are not necessarily drawn to scale; rather, the focus is on illustrating the principles of the embodiments. Furthermore, in the figures, the same reference numerals are used throughout different views to designate corresponding parts.

[0010] Figure 1 An example federated learning system is described.

[0011] Figure 2 An example method for adaptive aggregation of model parameters according to an embodiment is described.

[0012] Figure 3An example workflow for adaptive aggregation of model parameters according to an embodiment is described.

[0013] Figure 4 An example aggregation server according to an embodiment is described. Detailed Implementation

[0014] The embodiments provide systems and methods for data governance for federated learning. The embodiments facilitate privacy-preserving data governance by transmitting a portion of a potential data sample to a remote location. The quality of this portion is reviewed to exclude data samples that do not meet data governance criteria. The remote inspection focuses on visually reviewing regions of interest, but maintains privacy because the inspection cannot resolve any other identifiable object information, such as faces, body shapes, etc., as voxels outside the specified portion are not included. The inspection results are sent to the round authors so that inappropriate data samples can be excluded during federated learning rounds.

[0015] Federated learning (FL) is a distributed approach for training models that involve multiple distributed devices / collaborators and at least one aggregation / central server. Each of the collaborators downloads the current model and computes an updated model on its own using local data (ala edge computing). The locally trained model is then sent from the collaborators back to the central server, where the model or parameters are aggregated. A single merged and improved global model is sent from the server back to the devices. Different schemes and settings exist (e.g., peer-to-peer), but the foundation of federated learning is that local data remains on the local site and is not shared by collaborators. Federated learning allows machine learning algorithms to gain experience from a wide range of datasets located in different locations. This approach enables multiple organizations to collaboratively develop models without needing to directly share secure data with each other. Over several training iterations, the shared model is exposed to significantly more data than any single organization could have internally. In other words, federated learning decentralizes machine learning by eliminating the need to centralize data in a single location. Instead, the model is trained iteratively multiple times in different locations.

[0016] In this embodiment, collaborators or remote locations include hospitals and medical centers. Utilizing federated learning, these sites can maintain full control and ownership of their patient data, as well as complete tracking of data access, thus limiting the risk of misuse by third parties. Existing medical data is often not fully utilized by machine learning because it resides in data silos or walled gardens, and privacy concerns restrict access. However, centralizing or publishing data presents not only regulatory, ethical, and legal challenges related to privacy and data protection, but also technological challenges. Anonymizing, controlling access to, and securely delivering healthcare data is a non-trivial task, and sometimes even an impossible one. Anonymized data from electronic health records may appear harmless and compliant, but just a few data elements can allow for patient re-identification. The same applies to genomic data and medical images, making them as unique as fingerprints. Therefore, unless the anonymization process compromises the data's fidelity, likely rendering it useless, the possibility of patient re-identification or information leakage cannot be ruled out. Thus, hospitals and medical centers can fully leverage federated learning.

[0017] Figure 1 An example of a federated learning system is described. As depicted, a federated learning system comprises three collaborating devices / remote devices, referred to herein as collaborators 131, and a central / aggregator server 121. In a typical federated learning system, depending on the application, there may be dozens, hundreds, thousands, or more devices / collaborators 131. Each collaborator 131 is configured to acquire local data and train a model locally on said local data over multiple epochs. To begin training, the aggregation server 121 sends global model parameters (e.g., model weights or vectors representing the weights) to all selected collaborators 131 for initialization. Each collaborator 131 uses the locally acquired data to train (initialized with the global model parameters) its local model and updates the model parameters for certain epochs. Collaborators 131 then send the updated model parameters back to the aggregation server 121. The aggregation server 121 aggregates the updated model parameters and then sends the global model parameters back to collaborators 131 for another round of training. The result is a system that allows training data to be securely retained in-situ, thereby improving security and privacy. For some models, such as medical applications, the data can now be used where previous privacy concerns and regulations might have prohibited its use. However, the distributed nature of federated learning also introduces potential drawbacks and complexities. Generally, the minimum data movement required across the federation is for the model parameters and their updates.

[0018] To achieve quality and generalizability in clinical practice, a sufficiently large data governance dataset is needed to train modern AI models. However, in the context of federated learning, the privacy-preserving objectives and decentralized nature of clinical data complicate data governance procedures in determining clinical-grade accuracy and hinder the process, even though it is secure and generalizable to unseen data. Data governance is a critical step in training deep networks on medical image data, as irrelevant data samples can degrade training performance. Currently, data governance is typically performed on centralized datasets or independently within local collaborating institutions. For example, these selection tasks are often accomplished by manually reviewing the entire dataset to be used. While data governance on centralized datasets does enable the use of standardized selection criteria, it may violate privacy protections, a key advantage of federated learning. Data governance via remote desktops is considered an alternative solution, but significant risks of compromising patient privacy remain, as sensitive visual information could still be leaked. On the other hand, independent data governance at local collaborating institutions cannot utilize standardized procedures and can lead to undesirable results due to inconsistencies in governance procedures and the involvement of examiners. This may also be impractical, as each site requires the efforts of an expert.

[0019] In federated learning, data governance can have a greater impact on training deep networks than in centralized learning. For collaborators with limited data during federated learning rounds, poor-quality data samples can cause more severe interference. Depending on the robustness of the federated learning model aggregation, this can significantly impact the overall federated learning progress and may lead to unstable federated learning rounds. The embodiments provide consistent data governance that can promote enhanced stability and generalizability of federated learning training without violating privacy protections, a fundamental requirement for large-scale clinical adoption. Data availability is a significant obstacle to the implementation of AI methods in clinical practice. By providing data governance procedures that do not contradict the purpose of federated learning, the embodiments help increase the availability of high-quality data.

[0020] The embodiments described herein provide distributed data processing while maintaining privacy and transport considerations. Model training occurs in a decentralized manner with multiple collaborators 131, where only local data is available to each collaborator 131. The multiple collaborators 131 do not share data. In the embodiments, aggregation of model parameters occurs on an aggregation server 121. While the remainder of this document describes the aggregation scheme, the embodiments are applicable to any federated learning scheme in which sample data is not shared between sites. For example, a federated learning system may use a peer-to-peer sharing strategy, where collaborators 131 exchange parameters (but not sample data) among themselves without using a centralized server. Verification / governance of each collaborator 131's local data is performed. Data governance promotes enhanced stability and generalizability of federated learning training without violating privacy protections. Verification / governance can be performed remotely or locally. When performed remotely, an anonymized portion of the sample data is transmitted to a verification server, which determines whether the data samples meet the model's data selection criteria. Similarly, when executed locally, the anonymized portion of the sample data is fed into a machine-trained classifier that determines whether the data samples meet the model's data selection criteria.

[0021] Figure 2 A method for data management for multiple collaborators 131 in a federated learning system that trains a model over multiple training epochs is described. The workflow describes one potential epoch among the multiple training epochs to be executed. Dozens, hundreds, thousands, or more epochs may be executed before the model is trained. For each epoch, collaborator 131 or a remote device trains a local model using local data. The local model parameters are then sent to an aggregation server 121, which is configured to aggregate parameters from multiple collaborators 131 into a single central model. Parameters for the single central model are then transferred from the aggregation server 121 to the epoch author 131 for use in the subsequent training epoch. As described in the following sections, [the following methods can be used]. Figure 1 , 3 These actions can be performed by any combination of the components shown in 4. The following actions can be performed by collaborator 131, aggregation server 121, cloud-based server, or a combination thereof. Additional, different, or fewer actions may be provided. The actions are performed in the order shown or another order. Data governance can occur at any point during one or more rounds of federated learning. The actions can also be repeated. Depending on the federated learning scheme, some actions may be skipped or adjusted. For example, in a peer-to-peer federated learning system, there is no aggregation server 121. Instead, collaborators 131 share parameters among themselves without a centralized controller.

[0022] Figure 3 An example of a policy control process is described. Figure 3In this process, collaborator 131 segments the entire body. Collaborator 131 masks the remaining anatomical structures, thus anonymizing the patient data. The sample data is transmitted to a policy management server 141, which is configured to perform data policy management on the masked data. The policy management data or verification is sent back to collaborators 131 that have performed local training in combination with other collaborators 131.

[0023] In action A110, collaborator 131, one of multiple collaborators 131, receives global model parameters from parameter aggregation server 121. Collaborator 131 may be located remotely from aggregation server 121. In one embodiment, collaborator 131 is a hospital site. Collaborator 131 is configured to train the model using locally acquired data, which, for privacy or security reasons, does not leave the respective collaborator 131's site. Each collaborator 131 acquires data, trains the model in one or more phases, and then transmits the model parameters to aggregation server 121 via a network. Aggregation server 121 may include one or more machines or servers. A hierarchy of aggregation servers may be used to receive model parameters, which may be further aggregated by additional servers. Aggregation server 121 may be configured to operate in the cloud or on multiple different machines. In one embodiment, aggregation server 121 and collaborator 131 are located remotely. Alternatively, aggregation server 121 and collaborator 131 may be located in the same location. Aggregation server 121 and collaborator 131 communicate using a network, which may include a wired network, a wireless network, or a combination thereof. The wireless network can be a cellular telephone network, LTE (Long Term Evolution), 4G LTE, wireless local area networks such as 802.11, 802.16, 802.20, WiMax (Global Interoperability for Microwave Access) networks, DSRC (also known as WAVE, ITS-G5, or 802.11p and future versions), 5G wireless networks, or wireless short-range networks. Furthermore, network 127 can be a public network such as the Internet, a private network such as an intranet, or a combination thereof, and can utilize various networking protocols currently available or developed in the future, including but not limited to networking protocols based on Transmission Control Protocol / Internet Protocol (TCP / IP).

[0024] The model can be any model trained using machine learning processes. Examples of medical applications include finding clinically similar patients, predicting hospitalizations, mortality rates, and ICU stays due to cardiac events. The model can also include applications in the field of medical imaging, such as for whole-brain segmentation and brain tumor segmentation. In one embodiment, the model can be used to identify disease-related biomarkers in the context of COVID-19.

[0025] Model parameters can be represented by one or more parameter vectors. A parameter vector can be a set of parameters from the model (e.g., a set) or a representation of that set. A parameter vector can be a randomly selected component of a parameter vector. The model may include thousands or millions of parameters. For bandwidth and timing considerations, compressing the parameter set into parameter vectors may be more efficient than transmitting and recompiling each parameter in that set. Parameter vectors can also be further compressed. In one embodiment, the input parameter vectors can also be compressed into sparse subspace vectors.

[0026] In one embodiment, a supervised learning process is used to train the model. Supervised learning processes can be used for predicting numerical values ​​(regression) and for classification purposes (predicting the appropriate class). Supervised learning processing can also include processing images, audio files, video, numerical data, and text, among other types of data. Classification examples include segmentation, object recognition, face recognition, credit risk assessment, speech recognition, and customer churn. Regression examples involve determining continuous numerical values ​​based on multiple (sometimes hundreds or thousands) input variables.

[0027] The model can include machine learning processes such as Support Vector Machines (SVM), boosting and bagged decision trees, k-nearest neighbors, Naive Bayes, discriminant analysis, logistic regression, and neural networks. In the example, a two-stage convolutional neural network including max-pooling layers is used. The two-stage convolutional neural network (CNN) uses linear rectified units for non-linearity and fully connected layers at the ends for image classification. In an embodiment, an adversarial training process can be used to train the model; for example, the model can include a generative adversarial network (GAN). For adversarial training methods, a generator network and a discriminator network are provided for training by a device. The generator network is trained to identify features of data in a domain A and transform data from domain A into data indistinguishable from data in domain B. During training, the discriminator network acts as a judge to score the likelihood that the transformed data from domain A is similar to data from domain B, for example, whether the data is fake data or real data from domain B. In an embodiment, the model is suitable for medical diagnosis, for example, a classification network configured to identify or diagnose diseases or problems upon input images or volumes.

[0028] In action A120, collaborator 131 acquires sample data. The training data for each collaborator 131 is not independent and identically distributed (non-independent and identically distributed). The data distributions of two different collaborators 131 are different and imbalanced (e.g., collaborators 131 have different orders of magnitude of acquired data). In the example, for image data, one device might have several gigabytes of medical imaging data, associated with images taken for multiple procedures on multiple patients, while another device has only a single image dataset. Although collaborator 131 with more data can provide more useful parameters, both datasets are useful for training the segmentation model. The data quality may also differ between devices. Some devices may include higher-quality sensors or more data storage devices, allowing for the capture of higher-quality data.

[0029] Collaborator 131 is configured to train or configure a local model using training data. In this embodiment, the training data is labeled. Labeled data is used for supervised learning. The model is trained by inputting known inputs and known outputs. Weights or parameters are adjusted until the model precisely matches the known inputs and outputs. In the example, to train a machine learning model to identify certain artifacts using acquired image data, images of artifacts with various configurations are needed as input variables. Labels, such as correct names, can be assigned manually or automatically. The correct set of input variables and the correct classification constitute the training dataset. For example, labels can be provided by requesting additional input from a user (requesting manual annotation), deriving from additional data (parsing text descriptions), or by incorporating additional data from other sources. Other methods for labeling data can be used; for example, cloud-based services can provide accurate but incomplete labels downloaded from the cloud to the device. Sample data can be, for example, medical imaging data, including images (2D) or volumes (3D) of patient imaging data.

[0030] In action A130, collaborator 131 transmits anonymized portions of the sample data to management server 141, which is configured to identify data samples that do not meet data selection criteria. As described above, the data samples can be, for example, imaging data of a patient. In one embodiment, collaborator 131 utilizes whole-body organ segmentation to extract the anatomical regions of interest (e.g., brain, lungs) requiring management and masks the remainder of the image, making it impossible to identify individual patients. The masked data includes the same resolution, orientation, and image quality (SNR, sharpness / contrast) in the anatomical regions of interest as the original volumetric data, allowing remote examiners to apply the same data management criteria.

[0031] In one embodiment, the anonymized portion is an anatomical region of volumetric / imaging data unrelated to the diagnosis. However, the anonymized portion can provide an indication of the overall quality of the whole-body segmentation. Therefore, the anonymized portion can serve as a proxy for relevant portions while still maintaining the privacy and security of the patient's diagnosis. In one example, when the diagnosis involves the patient's brain, the portion may include or be an unrelated organ, such as the lungs. As described below, if the lung segmentation is of low quality, the brain segmentation is also likely to be of low quality, and therefore the data sample can be set aside for further examination.

[0032] In a further embodiment, an anonymized portion of the sample data is fed into a machine-trained model configured to output a classification or indication of quality. The machine-trained policy management classification model can be trained centrally and then implemented at each site. Collaborators 131 locally submit their data samples to the machine-trained policy management classification model. Alternatively, a federated learning approach can be used to train the machine-trained policy management classification model. In such a case, the machine-trained policy management classification model is first trained on policy management data, which is then used to train the main model.

[0033] In action A140, collaborator 131 receives validation of the sample data from policy management server 141. Policy management server 141 may perform one or more validation tasks on the sample data to determine whether the sample data is sufficient for training the model. Since the sample data is only a part of the data, the patient remains anonymous. In one embodiment, the sample data may be reviewed by a remote examiner. The remote examiner may focus on visually reviewing the region of interest. The remote examiner cannot see any other identifiable subject information, such as face, body shape, etc., because pixels or voxels outside the segmentation mask are masked or not included. The masked data includes the same resolution, orientation, and image quality (SNR, sharpness / contrast) as the original volumetric data in the anatomical region of interest, so the remote examiner can apply the same data policy management criteria. The review results are sent to round author 131 so that inappropriate data samples can be excluded during federated learning rounds. In one embodiment, if a data sample does not meet the data policy management criteria, the data sample may be flagged for local review. The local review process can be performed by a clinician or operator and may include identifying problems with the data sample and correcting or fixing those problems, as well as resubmitting the data sample to the policy management server 141.

[0034] In one embodiment, the policy management server 141 can execute outlier detection methods to verify sample data. Outlier detection methods primarily identify low-quality data as outliers by exploring statistical properties, distance-based features, clustering-based features, and density-based features. An example of an outlier detection method includes the AVF (Attribute Value Frequency) algorithm. In outlier detection-based data cleaning methods, data samples are identified as outliers if they are far removed from the majority of all data. For outlier detection-based methods, only certain measures or characteristics of the sample data can be provided to the policy management server 141, such as statistical properties describing the sample data or regions of interest.

[0035] Another approach may include using a machine learning-based classifier. In one embodiment, an AI classification network may be implemented for each data management criterion, which can automatically identify data samples that do not meet the data selection criteria. The classification network may be trained on a centralized dataset or a local dataset, and in one embodiment, may be subsequently deployed to collaborator 131. A subset of data samples with low-confidence classification scores identified by the AI ​​method may be sent to a reviewer for human visual inspection using the methods described above.

[0036] In action A150, collaborator 131 trains the model using validated sample data. Unvalidated sample data may be set aside or manually reviewed on-site. Alternatively, poor-quality sample data may be used, but the low-quality indication can be sent to an aggregation server, which can discount the parameters from a particular collaborator 131. In one embodiment, gradient descent or stochastic gradient descent is used to train the model. Both techniques attempt to minimize the error function defined for the model. To train (minimize the error function), collaborator 131 first connects to the parameter server. Collaborator 131 can start with randomly initialized model parameters, or it can request initial model parameters from the parameter server. The initial parameters can also be derived from another pre-trained model instead of being randomly initialized. Initial parameters can be assigned to all subsequent collaborators 131. Alternatively, updated central parameters can be assigned if training has already begun. In one example, collaborators 131 may initially communicate with the parameter server at different times. The first collaborator 131 may communicate with the aggregation server 121 and be assigned randomly initialized model parameters. Similarly, the second collaborator 131 can communicate with the aggregation server 121 shortly thereafter and be assigned randomly initialized model parameters. At some point, collaborator 131 begins transmitting model parameters back to the aggregation server 121. As described in detail below, the aggregation server 121 updates the central model parameters and transmits the updated model parameters to the turn author 131. Any collaborator 131 that communicates with the parameter server for the first time after this time can be assigned central parameters instead of randomly initialized model parameters. In this way, new collaborators 131 can be added to the system at any time during the training process without interrupting the training process. Distributing the latest parameters to newly joined collaborators 131 can lead to faster learning in the early stages.

[0037] Gradient descent attempts to minimize the model's error function. Each collaborator 131 uses local training data to train a local model. Training the model involves adjusting the local model's internal weights or parameters until the local model can accurately predict the correct outcome given new input data points. The result of the training process is a model that includes one or more local parameters that minimize the error of the function given the local training data. These one or more local parameters can be represented as a parameter vector. Due to the limited local training data, the trained model may not be very accurate when predicting outcomes for unlabeled input data points. However, given initial parameters covering a wider range of data, the trained model can be trained to be more accurate. Better initial parameters can be obtained from the aggregation server 121.

[0038] In action A160, collaborator 131 transmits the model's local parameters to a parameter aggregation server. The aggregated model parameters are used by multiple collaborators 131 in a subsequent training round. Subsequent rounds are similar to those described in A110-A160. The difference in each iteration is a different starting point for one or more parameters in the model. The received central parameter vector may differ from the local parameter vector provided in A110. This process is repeated iteratively multiple times until the parameters converge or a predetermined number of iterations is reached. This process may be repeated hundreds or thousands of times. In one example, hundreds (e.g., 100 to 500) or thousands (e.g., 3000 to 5000) iterations may be performed. Depending on the complexity of the model and the type and amount of device and data, more or fewer iterations may be performed. If new data is added to the training data, the device can retrain the model and request new central parameters (and this process may be repeated fully or partially). The result of the training process is a model that, for example, is able to accurately predict a classification given an unlabeled input. This model is used with new data to generate, for example, a prediction or classification. In one example, for an image classification model, collaborator 131 uses a trained machine learning model to identify objects or features in newly acquired (unseen) imaging data.

[0039] In one embodiment, the machine-trained model is deployed to each collaborator 131 to manage data. The machine-trained model can be trained, stored, and updated on the management server. After being trained, the machine-trained model can be transferred to collaborator 131, who then locally submit their data samples to the model for management.

[0040] In one embodiment, data governance can be applied to training models used in medical applications. Deep learning holds the potential to create crucial tools for screening and diagnosing medical problems, such as the recent COVID-19 pandemic. However, access to large, diverse healthcare datasets remains a challenge due to regulatory concerns about sharing protected healthcare information. One issue representing a strong deterrent from multi-institutional / multi-country AI trials is the stringent regulatory and protection requirements for patient data. The U.S. Health Insurance Portability and Accountability Act (HIPAA) and the European General Data Protection Regulation (GDPR) both establish stringent rules regarding the storage and exchange of personally identifiable and health-related data, requiring authentication, authorization, traceability, and—in the case of GDPR—AI explainability, raising considerations about data disposal, ownership, and AI governance. Ethical, ethical, and scientific guidelines also stipulate respect for privacy, namely the ability to retain full control and confidentiality over people's personal information. Another reason is the lack of standardized electronic patient records. Managing electronic patient data is very expensive, and hospitals in impoverished areas may not be able to afford to participate in research that requires it, potentially perpetuating the aforementioned biases and equity issues. The decentralized nature of data complicates data governance, making it difficult to determine the completeness and quality of results. Using the data governance schemes described above can alleviate some of these problems and allow federated learning processes to generate accurate and useful models. Examples of model types that can be used with federated learning and leverage the data governance described above are provided below.

[0041] In one embodiment, adaptive aggregation is used to overcome previous problems in applying FL to COVID-19 diagnosis. A trained model uses chest computed tomography (CT) scans, which are more sensitive for COVID-19 diagnosis and are currently widely used for early screening of the disease. A segmentation network is used, which automatically quantifies abnormal CT patterns typically found in COVID-19 patients. A second trained model uses a classification network that automatically detects COVID-19 pathology and distinguishes it from other pneumonia, interstitial lung disease (ILD), and normal subjects on chest CT. In one embodiment, sample data (e.g., whole-body (or partial) segmentation) is masked to exclude everything except the lungs. The masked volumes are either validated remotely or by the machine-trained classification network. Once confirmed, the volume can be used by collaborator 131 to train the network for detecting and / or diagnosing COVID-19.

[0042] In one embodiment, the segmentation network comprises a U-Net-like architecture with 3D convolutional blocks containing 1x3x3 or 3x3x3 CNN kernels to handle anisotropic resolution. 3D input tensors are fed into 3D 1x3x3 convolutional layers, followed by batch normalization and leaky ReLU. The feature maps are then propagated through five DenseNet blocks. For the first two DenseNet blocks, the features are downsampled using 1x2x2 convolutions with a stride of 1x2x2. The anisotropic downsampling kernels are configured to preserve the inter-slice resolution of the input image volume. The last three DenseNet blocks comprise isotropic downsampling kernels with a stride of 2x2x2. The input to each decoder block is obtained by concatenating encoder output features with the same resolution and feature maps upsampled from the previous decoder. The upsampling kernels are constructed using transposed convolutional kernels with the same size and stride as the corresponding DenseNet block. The final network output is derived by projecting the feature maps onto two output channels and activating them with a softmax operation.

[0043] In one embodiment, a classification network is configured to perform COVID-19 pathological differentiation on chest CT data, extracting 2D axial features and 3D global features. The network includes a ResNet50 as a backbone axial feature extractor, which takes a series of in-plane CT slices as input and generates feature maps for the corresponding slices. Features extracted from all slices are then combined using a max-pooling operation. Global features are fed into a fully connected layer, which generates a COVID-19 prediction score for each case using a softmax operation.

[0044] For each model, during federated learning, data management is used to process non-independent, identically distributed data from each hospital site. Data management allows models to be trained efficiently without being contaminated by low-quality training epochs due to low-quality or insufficient data samples.

[0045] Figure 4 An example of a policy management server 141 is depicted. The policy management server 141 includes at least a memory 125, a processor 123, an interface 129, and a transceiver 127. The policy management server 141 can use the transceiver 127 to communicate with one or more collaborators 131 or sites to receive sample data and transmit validation data. The one or more collaborators 131 may include hospital sites or otherwise configured as centers for acquiring or storing patient medical data. For example, the one or more collaborators 131 may include medical imaging equipment and / or PACS systems configured to acquire or store medical imaging data for training models and generating model parameters.

[0046] Memory 125 may be a non-transitory computer-readable storage medium storing data representing instructions executable by processor 123 for time-varying readmission risk prediction. Instructions for implementing the processes, methods, and / or techniques discussed herein are provided on a non-transitory computer-readable storage medium or memory, such as cache, buffer, RAM, removable media, hard disk drive, or other computer-readable storage media. Non-transitory computer-readable storage media include various types of volatile and non-volatile storage media. Functions, actions, or tasks illustrated in the figures or described herein are performed in response to one or more sets of instructions stored in or on a computer-readable storage medium. These functions, actions, or tasks are independent of the instruction set, storage medium, processor, or processing strategy and may be performed by software, hardware, integrated circuits, firmware, microcode, etc., operating individually or in combination. Similarly, processing strategies may include multiprocessing, multitasking, parallel processing, etc. In one embodiment, instructions are stored on a removable media device for reading by a local or remote system. In other embodiments, instructions are stored at a remote location for transmission over a computer network or telephone line. In other embodiments, instructions are stored within a given computer, CPU, GPU, or system. The memory 125 can store a model or machine learning network.

[0047] Processor 123 is a general-purpose processor, central processing unit, control processor, graphics processing unit, digital signal processor, 3D rendering processor, image processor, application-specific integrated circuit, field-programmable gate array, digital circuit, analog circuit, combination thereof, or other device now known or hereafter developed for processing medical imaging data. Processor 123 is a single device or multiple devices operating in a serial, parallel, or discrete manner. Processor 123 may be the main processor of a computer such as a laptop or desktop computer, or it may be a processor used to handle some tasks in a larger system such as a server. Processor 123 is configured by instructions, design, hardware, and / or software to perform the actions discussed herein.

[0048] Processor 123 can be configured to receive sample data from one or more collaborators 131, examine the data, and return a determination of whether the sample data meets data management criteria set by an administrator, such as a model or federated learning process. In one embodiment, remote examination focuses on visually reviewing regions of interest (ROIs) of the data samples. The data reviewed by the processor may include only a portion of the data samples stored at collaborator 131. For example, collaborator 131 may mask a portion of an image or volume, making it impossible to identify an individual patient. Therefore, the examination performed by processor 123 may not reveal any identifiable object information, such as a patient's face, body shape, etc., because pixels or voxels outside the mask are obscured. The masked data includes the same resolution, orientation, and image quality (SNR, sharpness / contrast) as the original volumetric data in the anatomical region of interest, so the remote examiner can apply the same data management criteria. The examination results are sent to round author 131 so that inappropriate data samples can be excluded during federated learning rounds.

[0049] Processor 123 can be configured to implement a machine learning classification model, which is configured to identify data samples that satisfy data governance criteria. The machine learning classification model can be stored in memory 125. The machine learning classification model can be trained using a training dataset, which includes data samples that satisfy and do not satisfy the data governance criteria. Processor 123 can implement multiple models for different criteria. Supervised or unsupervised learning can be used to train one or more machine learning classification models. One or more machine learning classification models can include a neural network defined as multiple sequential feature units or layers. Sequence is used to indicate the general flow of output feature values ​​from one layer to the input to the next layer. Information from the next layer is fed into the next layer, and so on, until the final output. The layers can be feedforward only or can be bidirectional, including some feedback to the previous layer. Nodes in each layer or unit can be connected to all nodes or only a subset of nodes in the previous and / or subsequent layer or unit. Skip connections, such as from the output of one layer to the next layer in sequence and other layers, can be used. Deep architecture is defined as learning features at different levels of abstraction based on input data, rather than pre-programming features and attempting to associate them with attributes. The learned features reconstruct lower-level features (i.e., more abstract or compressed levels of features). Each node in a unit represents a feature. Different units are provided for learning different features. Various units or layers can be used, such as convolutions, pooling (e.g., max pooling), deconvolutions, fully connected layers, or other types of layers. Any number of nodes are provided within a unit or layer. For example, 100 nodes are provided. Subsequent units can have more, fewer, or the same number of nodes. Methods such as k-nearest neighbors are used based on the distribution of samples; unsupervised learning can also be used.

[0050] Different neural network configurations and workflows can be used in or applied to machine learning classification models, such as Convolutional Neural Networks (CNNs), Deep Belief Networks (DBNs), or other deep networks. CNNs learn feedforward mapping functions, while DBNs learn generative models of data. Furthermore, CNNs use shared weights across all local regions, while DBNs are fully connected networks (e.g., including different weights for all regions of the feature map). CNN training is fully discriminative through backpropagation. DBNs, on the other hand, employ layer-by-layer unsupervised training (e.g., pre-training), followed by discriminative refinement using backpropagation as needed. In one embodiment, the network arrangement used for training is a fully convolutional network (FCN). Other network arrangements can be used, such as 3D Very Deep Convolutional Networks (3D-VGGNet). VGGNet stacks many layers, containing narrow convolutional layers followed by max-pooling layers. A 3D Deep Residual Network (3D-ResNet) architecture can be used. ResNet uses residual blocks and skip connections to learn residual mappings.

[0051] Training data for machine learning classification models includes benchmark ground truth data or gold standard data obtained at each collaborator 131 or site. Benchmark ground truth data and gold standard data are data containing correct or reasonably accurate labels, which are validated manually or by some other accurate method. Training data can be obtained at any point before it is fed into the model. Each machine learning classification model can take training data (e.g., patient data) as input and output a prediction or classification, for example, based on one or more data criteria (such as quality, resolution, etc.) indicating that the data is acceptable. The predictions are compared with annotations from the training data. A loss function can be used to identify errors from the comparison. The loss function serves as a measure of how far the current set of predictions is from the corresponding true values. Some examples of loss functions that can be used include mean squared error, root mean square error, and cross-entropy loss. Mean squared error loss (or MSE for short) is calculated as the average of the squared differences between the predicted and actual values. Root mean square error is similarly calculated as the average of the root mean square errors between the predicted and actual values. For cross-entropy loss, each predicted probability is compared to the actual class output value (0 or 1), and a score is calculated that penalizes the probability based on its distance from the expected value. The penalty can be logarithmic, providing a small score (0.1 or 0.2) for small discrepancies and a large score (0.9 or 1.0) for large discrepancies. During training and repeated iterations, the network attempts to minimize the loss function, as a lower error between the actual and predicted values ​​indicates that the network is doing well in terms of learning. Different optimization algorithms can be used to minimize the loss function, among others, such as, for example, gradient descent, stochastic gradient descent, batch gradient descent, and mini-batch gradient descent. With the goal of minimizing the loss function, the input, output, comparison, and adjustment process is repeated for a predetermined number of iterations. Once adjusted and trained, the model is configured to be applied. In one embodiment, the trained model can be deployed to each of collaborators 131. Collaborators 131 can then apply the model to locally managed data using standardized management criteria.

[0052] The policy management server 141 may also include an interface device having a display for showing information to a user, such as a CRT (cathode ray tube) or LCD (liquid crystal display) monitor, and a keyboard and pointing device, such as a mouse or trackball, through which the user can provide input to the computer. Other types of devices may also be used to provide interaction with the user; for example, feedback provided to the user may be any form of sensory feedback, such as visual feedback, auditory feedback, or tactile feedback; and input from the user may be received in any form, including sound, speech, or tactile input. The interface may include one or more buttons, keypads, keyboards, mice, styluses, trackballs, rocker switches, touchpads, voice recognition circuitry, or other devices or components for inputting data to the policy management server 141 while reviewing sample data.

[0053] The various improvements described herein can be used together or individually. Although illustrative embodiments of the invention have been described herein with reference to the accompanying drawings, it should be understood that the invention is not limited to those precise embodiments, and various other changes and modifications can be made by those skilled in the art without departing from the scope or spirit of the invention.

Claims

1. A method for training a model using federated learning by multiple collaborators, the method comprising: The collaborators among the plurality of collaborators receive global model parameters from the parameter aggregation server; Sample data is obtained by the collaborators; The sample data is anonymized by the collaborator; The collaborator transmits anonymized sample data to a policy management server, which is configured to identify data samples that meet or do not meet one or more data selection criteria. The collaborator receives verification of the sample data from the policy management server; The collaborators train the model using validated sample data; and The collaborator transmits the local model parameters of the model to the parameter aggregation server.

2. The method according to claim 1, wherein the anonymized sample data includes a masked portion of the sample data.

3. The method of claim 2, wherein the masking portion comprises the same resolution, orientation, and SNR or sharpness / contrast as the sample data in the anatomical region of interest.

4. The method of claim 2, wherein one or more entities on the policy management server are unable to identify the subject information of the sample data because the voxels or pixels within the masked portion are blocked.

5. The method of claim 1, wherein the verification includes a classification confidence score of the sample data, and wherein the collaborator is configured to use only sample data with a classification confidence score exceeding a threshold.

6. The method of claim 1, wherein the policy management server includes a machine-trained classification network configured to classify anonymized portions of sample data as satisfying or not satisfying one or more data selection criteria.

7. The method of claim 1, wherein the model comprises an image classification model for performing medical diagnosis, and wherein the sample data comprises the patient's medical imaging data.

8. The method of claim 1, wherein identifying data samples that satisfy or do not satisfy one or more data selection criteria includes identifying outlier data samples.

9. The method of claim 1, wherein data samples that do not meet the one or more data selection criteria are marked for manual review by a clinician.

10. A method for training a model using federated learning by multiple collaborators, the method comprising: The collaborators among the plurality of collaborators receive global model parameters from the parameter aggregation server; Sample data is obtained by the collaborators; The sample data is anonymized by the collaborator; The collaborator submits anonymized sample data to a machine-trained policy classification model on the policy management server. The machine-trained policy classification model is configured to identify data samples that do not meet the data selection criteria. The collaborator receives verification of the sample data from the policy classification model trained by the machine; The collaborators train the model using validated sample data; and The collaborator transmits the local model parameters of the model to the parameter aggregation server.

11. The method of claim 10, wherein the anonymized sample data includes a masked portion of the sample data.

12. The method of claim 11, wherein the masking portion comprises the same resolution, orientation, and SNR or sharpness / contrast as the sample data in the anatomical region of interest.

13. The method of claim 10, wherein the verification includes a classification confidence score of the sample data, wherein the collaborator is configured to use only sample data with a classification confidence score exceeding a threshold.

14. The method of claim 10, wherein the machine training model comprises an image classification model for performing medical diagnosis, wherein the sample data comprises medical imaging data of a patient.

15. The method of claim 10, wherein identifying data samples that satisfy or do not satisfy one or more data selection criteria includes identifying outlier data samples.

16. The method of claim 15, wherein data samples that do not meet the one or more data selection criteria are labeled for manual review by a clinician.

17. A system for federated learning, the system comprising: Multiple collaborators, each of which is configured to acquire sample data, anonymize the sample data, and transmit the anonymized sample data to a policy management server for verification, and each collaborator is further configured to use the verified sample data to train a local machine learning model, update the local model weights of the local machine learning model, and send the updated local model weights to an aggregation server. The policy management server is configured to receive anonymized sample data and identify sample data that meets or does not meet one or more data selection criteria. and The aggregation server is configured to receive updated model weights from multiple collaborators, calculate aggregated model weights, and transmit the aggregated model weights to multiple collaborators to update the local machine learning model.

18. The system of claim 17, wherein each collaborator is configured to anonymize sample data by masking regions that do not include the region of interest, while maintaining the same resolution, orientation, and SNR or sharpness / contrast as the sample data.

19. The system of claim 17, wherein the policy management server includes a machine training model configured to identify sample data that meets or does not meet one or more data selection criteria.

20. The system of claim 17, wherein the local machine learning model includes an image classification model for performing medical diagnosis, wherein the sample data includes the patient's medical imaging data.

Images