Developments in Artificial Intelligence for Ophthalmology: Federated Learning

Abstract

Technological advances in machine learning and increasing global connectivity have accelerated and encouraged global collaborative training in artificial intelligence (AI) health research.1 Multisite collaboration allows for access to larger and more heterogenous data sets which lead to more robust and externally generalizable AI machine learning models.2,3 AI will revolutionize the future practice of ophthalmology.4–8 Quite a number of specific algorithms have been developed and are now available for aiding diagnosing and monitoring various eye diseases.9–13 However, a major concern with collaborative machine learning in health care is data privacy concerns which limit data sharing and the clinical implementation of what is technically possible.14–16 The great incentive of collaborative training to achieve better deep learning models and the rising need for data security has led to increasing focus on privacy-preserving approaches such as federated learning (FL), generative adversarial networks, and cryptographic methods such as concurrent use of blockchain technology with AI. In 2016, Google first introduced FL,17 a distributed machine learning framework that allows collaborative machine learning while preserving data privacy. Over the years, FL has shown to produce AI models comparable to AI models built by centralized training18 and is increasingly seen as an attractive alternative to traditional centralized training.1,18 Specific to health research where data is often not independent and identically distributed, FL has shown to remain robust.18 FL also allows for the training of AI models amongst various data types including imaging—magnetic resonance imaging for brain tumor segmentation,19 chest x-ray for coronavirus disease 2019 clinical outcome prediction,20 electronic medical records for predicting hospitalization,21 retinal fundus photographs,22 histology slides—in cancer diagnosis,23 genomics,24 and even Internet of Medical Things.25 In a FL framework,26 initial weights of the global model parameters are broadcast by a centralized server to selected participating sites. Each participating site then trains a local model (shares the same model architecture with the global model) using its local data and sends updated model parameters to the centralized server. The server then aggregates them to update the global model weight. In this setting, original data never leaves the local device or site. The updated global model is then broadcast to a new set of participating sites for another round of local training. The process is repeated until the global model converges. The key advantage of FL is the generation of a higher quality model from a larger training data set obtained from multiple sources, beyond what could have been achieved with the data of a single device or site. Limiting the transfer of raw data also allows for high levels of data privacy and minimizes data aggregation costs. FEDERATED LEARNING IN OPHTHALMOLOGY Despite the relative infancy of FL, there is an increasing number of studies demonstrating the use of FL in the training of medical AI models.19–21 Within ophthalmology, a simulated FL model has been shown to achieve similar performance to local learning (training on internal single data set) and centralized learning (where data is pooled) in microvasculature segmentation and diabetic retinopathy classification using optical coherence tomography angiography images.27 The acquisition of optical coherence tomography angiography images from different machines shows that FL remained robust even in heterogenous data types, which is often the case in not independent and identically distributed medical data sets. Two studies further demonstrated the real-world clinical success of FL in ophthalmology via the evaluation of retinopathy of prematurity (ROP).22,28 The first study22 demonstrated the use of FL in diagnostics by developing a deep learning model for fundus image-based classification of ROP and showed similar performance to centralized learning.22 The second study28 utilized the FL framework to evaluate interinstitution ROP diagnostic performance via a FL-derived vascular severity score. Demographic factors and clinically labeled fundus images were used to train an FL model to generate a vascular severity score. The generated vascular severity score for each category of ROP severity was then compared across institutions to assess for interinstitution variability without the need to transfer confidential patient data.28 While the number of studies evaluating the use of FL in ophthalmology is limited, these early works demonstrate the feasibility of FL in the building of AI models for diagnostics and classification of ophthalmic diseases. In addition, it shows the potential variable benefits of FL beyond diagnostics such as disease classification and reporting standardization. CURRENT AND FUTURE DEVELOPMENTS IN FEDERATED MACHINE LEARNING While FL provides an option for better privacy preservation and collaborative training, there are still limitations to the FL framework that future research may address. First, the robustness of the FL model is dependent on the quality of data contributed by each participating site. In FL, as original data is kept private at local sites and not directly examined by the entire federation, there may be unintentional inclusion of poor quality data that may compromise model performance. Despite this, several approaches to circumvent this have been proposed such as data cleaning at each local site with pretraining data normalization methods. In addition, the inclusion of heterogenous input data is overall advantageous in creating a generalizable global model and FL has shown to produce more robust models than models trained on local site data alone.18 Second, as with all deep learning models, explainability analyses are important for acceptance in real-world settings. While explainability analyses such as heat map generation via ablation analysis and saliency analysis are possible, further studies are required to validate these in the FL framework to increase clinician and patient acceptance. Third, sites with large and good quality data sets may feel little incentive to participate in FL as local model training may yield acceptable results. Hence the shift from technical proof of concepts to actual clinical application of federated AI will require intellectual property considerations and the design of fair incentive mechanisms to attract participants to join the federation. Fourth, while original data remains private, the transfer of model updates in FL is still susceptible to privacy attacks. Attackers may use model parameters to reconstruct original data sets (reconstruction attack) or infer if specific data records were used in model training (inference attack). Several approaches to further enhance privacy in FL have been proposed such as the use of cryptographic methods and differential privacy. Fifth, to join the federation, local participating sites require the infrastructure and computational capabilities for model training and access to the communication network for model update sharing. This may limit the participation of less technologically developed sites. Efforts to address the digital divide across the world will allow for widespread implementation of FL. With the explosion of big data and increased global digitalization, privacy-preserving technologies such as FL will play an increasingly important role. FL is a promising approach that supports global health collaboration. Further research is necessary to address the current barriers to clinical application and in doing so, will render FL a key strategy in preserving privacy in AI health research.

Ähnliche Arbeiten