Christian Doppler Laboratory for Artificial Intelligence in Retina – “CD AIR”
The Lab is focused on solving socio-technical hurdles for the deployment of AI in eye care as clinical-decision support systems. This interdisciplinary research team is composed of retinal specialists, computer scientists and software engineers, developing innovative image analysis methods for precision medicine in retinal disease.
The overarching goal of the Laboratory is to enable AI-driven clinical decision support systems (CDSSs) for the effective management of retinal diseases, representing the leading causes of blindness. The CDSS would naturally be rooted in rich multi-modal imaging offered by OCT device manufacturers, but would also exploit the data available in the associated EHR. Such modern AI based CDSS in retina is expected to:
- automate tedious tasks and provide objective examination summaries,
- allow detecting pathological changes of retina early and precisely, and
- improve the clinical workflow efficiency allowing retinal specialists to examine more patients more accurately. This would constitute a groundbreaking raise of the level of healthcare quality in ophthalmology.
The research will advance our understanding of the most devastating blinding diseases. It will widely expand the spectrum of accessible biomarkers in the current diagnostic and therapeutic area going far beyond the previously known features and established paradigms, open the horizon for novel subclinical biomarker detection providing a novel understanding of disease mechanisms and identification of novel therapeutic targets.
With the enormous advance of AI over the past decade, nowadays capable of a human-level pattern recognition performance, uptake of deep learning models in clinical routine has nevertheless still not happened. This stems from following very unique challenges that machine learning technology is facing when being deployed in a medical setting.
- High-dimensional Data
Medical data is dense rather than big, the so called “large p, small n” problem. There is often a vast amount of multi-modal detail available about a patient while the number of available patients is limited. Such a setting poses a considerable challenge for an effective machine learning.
- Data quality
Data is the fuel that powers AI, though much of real-world medical data is heterogeneous, noisy and incomplete or missing, acquired with different imaging devices and protocols. There is a large variability of expert opinions and ground truth labels are often not retrievable. Thus, machine learning methods have to handle data sparsity, and missing or incorrect values and labels.
- The last mile problem, aka the final percent
In a safety-critical environment like medicine, any wrong conclusion could have a catastrophic downstream effect. Thus, it is important that AI demonstrate robustness, and is able to handle also rare, corner-case situations, which is often ignored in performance evaluations. Currently, a clinician in the loop is needed to assure the credibility of the output.
- The distribution shift
It poses a fundamental limitation to machine learning abilities. As the imaging equipment changes, the models trained on retrospective data may not generalize to the newly acquired data anymore. Thus, algorithms that can withstand some extent of distribution shift are needed.
Most of current deep learning is supervised with a significant cost and manual effort required to annotate often at the pixel-level the large amount of data needed for learning. In addition, in cases of a distribution shift, supervised learning requires extensive relabelling to regain its performance. Methodologies behind unsupervised, self-supervised and weakly-supervised approaches need to be better exploited.
AI models need to demonstrate generalizability, interpretability and reproducibility to gain trust. This places a high bar on capabilities AI needs to demonstrate. Currently, most deep learning models are essentially black-boxes, where clinicians have limited understanding of how a model comes to its decision. Further effort into providing evidence for predictions by a model is needed for clinicians to properly weigh its output.
Use of AI in the clinic requires additional validation from socio-technical and clinical user experience aspects, to make sure the technology supports and not hinders the clinical workflows.
Despite the above challenges, we are in an exciting era at the interface of AI and retinal imaging, with ample opportunities to innovate and improve the current state of the art. This would allow to leverage the power and enourmous potential of deep learning to make a real impact in ophthalmology by further boosting retinal image analysis capabilities and enabling AI tools to find their way into the market and into the hands of the clinicians world-wide.
AI-based monitoring of retinal fluid in disease activity and under therapy
Schmidt-Erfurth U, Reiter GS, Riedl S, Seeböck P, Vogl WD, Blodi BA, Domalpally A, Fawzi A, Jia Y, Sarraf D, Bogunovic H
2021 Progress in Retinal and Eye Research
Christian Doppler Laboratory for Artificial Intelligence in Retina – “CD AIR” is supported by the Christian Doppler Research Association in collaboration with Heidelberg Engineering GmbH.
In Christian Doppler Laboratories, application-oriented basic research is pursued at a high level, and expert scientists cooperate with innovative companies. The Christian Doppler Research Association is an international best practice example for promoting this collaboration.
Christian Doppler Laboratories are financed jointly by the public purse and the participating companies. The most important public sponsor is the Federal Ministry for Digital and Economic Affairs (BMDW).
Contact and address
Christian Doppler Laboratory for Artificial Intelligence in Retina
Department of Ophthalmology and Optometry
Medical University of Vienna
1090 Vienna, Austria
Phone: +43 1 40400 73419
Reseponsible for contents:
Hrvoje Bogunovic, PhD
Department of Ophthalmology
Medical University of Vienna
Währinger Guertel 18-20