ARTORG Media Releases
19 September 2019
Advanced AI boosts clinical analysis of eye images
A fast and reliable machine learning tool, developed by the ARTORG Center, University of Bern, and the Department of Ophthalmology, Inselspital brings Artificial Intelligence (AI) closer to clinical use in Ophthalmology. The novel method published in Nature Scientific Reports on September 19, 2019 presents a tool that reliably extracts meaning from extensive image data. Based on a convolutional neural network (CNN) the tool is able to provide results within seconds, thus supporting the doctor with comprehensive image analysis during a consultation with the patient.
Modern medical imaging devices allow ophthalmologists to monitor chronic eye conditions in detail. Ophthalmologists mostly choose Optical Coherence Tomography (OCT), an imaging tool that generates 3D images of the eye at extremely high resolution. But without AI support the large amount of images and information exceeds the capacity of an individual expert. The challenge of this study was, to provide AI-tools, capable of analyzing a large amount of data at very high speed to facilitate the use of all available information from image analysis during patient consultations.
The research team from Artificial Intelligence in Medical Imaging (AIMI) laboratory at the ARTORG Center for Biomedical Engineering Research, University of Bern, and the Department of Ophthalmology at Inselspital, Bern University Hospital now presents a machine learning method capable of identifying a wide range of biomarkers from OCT-scans of the retina virtually providing clinically relevant data support instantaneously.
Artificial Intelligence spots biomarkers for each disease type
“In our approach, the AI classifies patient OCT scans on the basis of disease-typical biomarkers”, explains Prof. Dr. Raphael Sznitman, group Head of the ARTORG ́s AIMI lab. Biomarkers are landmarks and features in OCT scans that can indicate a disease or can be used to show worsening or improvement after treatment. “What sets our results apart is that our AI algorithm provides a rich biomarker characterization, able to classify scans on the basis of well understood and known indications from the clinical community. Here, we manage to identify these biomarkers autonomously, without the cost and effort of having a trained human eye specialist previously mark the structures, the technology needs to focus on.”
3D imaging monitors sight-threatening macular diseases
The most frequent eye diseases worldwide are linked to degenerative eye conditions that deteriorate the macula (part of the rear part of the eye or retina), ultimately leading to loss of sight. Prof. Dr. med. Sebastian Wolf, Chairman and Head of the Department of Ophthalmology at Inselspital, Bern University Hospital, as a clinician uses OCT-scans for the therapy of chronic retinal conditions, such as age-related macular degeneration (AMD) or diabetic macular edema (DME). “As patient numbers are growing, we need to develop automated AI tools in the clinical setting to assist doctors in analyzing the abundant data of OCT scans. Having accurate, comprehensive information from the analysis of a patient’s OTC at hand during the consultation, is key to improve management of such diseases in the future. The tool presented in this paper is an important step in achieving the goal of better care for the patient.”
Machine learning makes the abundance of images exploitable
To assist eye doctors in clinical routine and research, computer programs can automatically extract, summarize and present the most important information from the growing number of routinely generated OCT scans. “This automated analysis can provide a cost effective and reliable tool for doctors to having to go through every image manually”, says Thomas Kurmann PhD student at ARTORG AIMI lab. “Our results so far are showing, that our Artificial Intelligence can consistently classify the most common disease types automatically with great precision, and identify a wide range of biomarkers typically found in pathological eye scans.”
Part of a large collaborative effort between University of Bern, the Inselspital and RetinAI Medical AG, the findings have the potential to be transformative for ophthalmologists. The next steps for this technology will be to understand how it can be incorporated into the clinic to improve the management of chronic eye conditions.
Prestigious grant to use Artificial Intelligence for improved glucose control
The ARTORG Center for Biomedical Engineering Research of the University of Bern is the recipient of a grant from JDRF, the leading global funder for type 1 diabetes research. Thanks to the grant, a team led by Stavroula Mougiakakou will investigate a large, real-world dataset to develop advanced algorithms for automated insulin delivery that are capable of predicting dangerously low or high blood sugar levels. The goal is to optimize and personalize insulin treatment.
People with diabetes have a need to control their blood sugar levels to a normal range at all times. Today, scientifically validated automated insulin delivery (AID) systems are available as tools to enhance self-management. These systems empower people with diabetes to more successfully control their condition to prevent hypoglycemic (low-glucose) and hyperglycemic (high-glucose) events. However, these tools still have some shortcomings, because the algorithms used in these systems do not react adequately to variables influencing the blood sugar fluctuations in individuals, such as food intake or physical activity.
Bernese research group succeeds with advanced algorithms
The laboratory Artificial Intelligence in Health and Nutrition of the ARTORG Center proposes to use big data and deep and reinforcement learning technologies (machine learning tools) to improve the prediction accuracy of AID algorithms. The Artificial Intelligence (AI) algorithms will be trained to foresee dangerously low or high blood sugar levels in real life situations. “If we can predict future blood glucose levels, we can provide early warnings and thus improve each patient’s safety”, explains Prof. Dr. Stavroula Mougiakakou, who leads the laboratory.
Prof. Mougiakakou ́s team is one of only eight laboratories to receive the prestigious research grants awarded through a request for applications (RFA) by the US-based diabetes research foundation JDRF. The grant of about 144,000 USD most importantly provides access to big data, containing diabetes-specific patients’ information from thousands of glucose monitors and insulin pumps. The de-identified data were collected by Tidepool — a nonprofit organization committed to making diabetes data more accessible, actionable and meaningful for people with diabetes, clinicians, and researchers — through the Tidepool Big Data Donation Project.
“We are honored and proud that JDRF recognizes the potential of and our expertise in applications of AI in diabetes,” says Stavroula Mougiakakou, principal investigator of the project. “This grantgives us a unique opportunity to access big diabetes-related data and use it synergetically with advanced AI algorithms to uncover patterns and trends that bring us closer to more precise and personalized insulin treatment.” Prof. Mougiakakou has introduced the use of AI in insulin treatment optimization in the late 1990s.
Machine learning to gain diabetes insights from big data
The data JDRF and Tidepool will provide access to has been de-identified and combined in meaningful ways for the use of clinicians and researchers. “This data is a big step forward for ourresearch”, says Qingnan Sun, PhD student at the ARTORG laboratory working on the JDRF funded project. “The data access will help us to refine the algorithms that are used in AID systems, making it possible to warn a person at least half an hour before they develop hypo- or hyperglycemia.”
Personalizing blood sugar predictions
“First the AI algorithms will analyze glucose data to detect for each person, how age, bodily fitness, insulin treatment, number of years with the disease, as well as daily routines influence his or her glucose control,” explains Prof. Mougiakakou. “Subsequently the model uses these findings to predict hypo- or hyperglycemic events early enough so that the person with diabetes can react and prevent their onset. It is important to mention that the model will continue to learn individual’spattern and habits while in use.”