Distributed deep learning for classifying multimodal, uncertain and rare data in ophthalmology – ADMIRE

Artificial intelligence in Health, and in Ophthalmology in particular, has been very successful over the past four years. The work carried out jointly by LaTIM (Inserm - Research Unit 1101) and Evolucare Technologies, as part of the RetinOpTIC project funded by FUI in 2015, resulted in 2018 in a software program that screens for diabetic retinopathy, a major ocular pathology, based on fundus photographs. The underlying algorithm, whose performance reaches that of a retinal expert, is being deployed in several clinical centers around the world, through the Evolucare Technologies cloud. The success of this solution is partly due to the large amount of data used for learning, namely 760,000 images from 100,000 diabetic patients in the OPHDIAT screening network in Ile-De-France. The common objective of LaTIM and Evolucare Technologies is to extend screening to all pathologies affecting the eye, or visible through the eye (cardiovascular pathologies, neurodegenerative diseases, etc.). In order for each of these conditions to be represented by a sufficient number of examples, the images must come from a much more diverse population than the diabetic population. This is only possible if the data come from several clinical centers.

Within the framework of the ADMIRE LabCom, we therefore propose that clinical centers using the Evolucare Technologies cloud participate, if they so wish, in the progressive enrichment of artificial intelligence (AI). In addition to a retrospective analysis of the data from each center, we propose that user feedback on AI diagnoses be used to refine learning over time. In a learning scenario on multi-center health data, it is no longer possible to export all the data to AI researchers. It is also not possible to conduct training in each clinical center, due to limited computing power. We therefore propose to divide learning between clinical centers, on the one hand, and a computer cluster, on the other hand: only abstract information (neural weights, gradients, etc.) will be transfered between the different parties. This will allow each clinical center to keep control over its patients' data and Evolucare Technologies to keep control over the developed AI. However, this raises security issues: in particular, it is necessary to ensure that data flowing between clients and the server, even abstract data, does not allow the extraction of sensitive information about patients or, conversely, about AI models in training. This also raises artificial intelligence challenges. This requires the ability to manage interpretations of varying quality, both from internationally recognized experts of a given pathology and from new AI users; in particular, initial learning should not be forgotten as AI is gradually refined. Finally, it must be possible to manage the variability of clinical centers in terms of collected imaging data (fundus photography, optical coherence tomography, etc.), contextual information and annotations.

Solving these various problems will pave the way for the screening of many diseases, and in particular for the screening of rare diseases, for which clinician' experience is necessarily weaker and the contribution of AI is particularly useful. The implementation of this distributed and secure AI platform will also enable LaTIM and Evolucare Technologies to jointly address new clinical problems in the coming years.