DIAMOND: Device-Independent diAbetic Macular edema ONset preDiction

Diabetic macular edema (DME) is a major complication of diabetes. Characterized by retinal thickening in the macula and often accompanied by hard exudate deposition, DME is a prevalent cause of vision loss among diabetic patients. The challenge's focus is on center-involved diabetic macular edema (ci-DME), a critical form of DME responsible for significant vision impairment. The presence of DME is generally assessed through 3D optical coherence tomography (OCT) imaging. A recent study has shown that the presence of ci-DME can also be assessed using 2D color fundus photography (CFP) [1].

The DIAMOND challenge seeks to revolutionize the approach to diagnosing and treating ci-DME by integrating AI and deep learning with ultra-wide-field color fundus photography (UWF-CFP), a recent evolution of CFP. More challenging than assessing the presence of ci-DME, the goal is to develop and evaluate models that can predict if a patient will develop ci-DME within a year, using UWF-CFP images alone. Because it offers a much wider field of view, we hypothesize UWF-CFP is more likely to capture early signs of DME than standard CFP. Success in this challenge could significantly improve early detection and treatment planning, reduce vision loss incidents, and exemplify AI's efficacy in healthcare.

For training, the DIAMOND Challenge uses data collected in 14 French Hospitals in the framework of the EVIRED project (https://evired.org/). EVIRED not only focuses on predicting the development of ci-DME but also broadly aims at forecasting the onset of diabetic retinopathy (DR) complications in general. Initial experiments with a simple baseline algorithm (ResNet-50) on EVIRED data suggest the feasibility of ci-DME onset prediction using UWF-CFP (area under the ROC curve = 0.73). For performance evaluation, DIAMOND will also use independent data from the LAZOUNI Ophthalmology Clinic in Tlemcen, Algeria.

By leveraging diverse datasets, including data from French and Algerian populations and images captured using different UWF-CFP devices, the challenge underscores its commitment to developing solutions that are universally applicable, transcending geographic and demographic boundaries. This generality is critical in ensuring that the predictive models developed are robust and effective across different population groups, enhancing their clinical utility on a global scale.

Moreover, the DIAMOND Challenge introduces a significant methodological challenge that sets it apart from typical predictive modeling competitions. Participants in this challenge will not have access to the training, validation, or test datasets and, consequently, will not have the opportunity to train their models directly. Instead, they are required to submit their code, in a Docker container, which the organizing committee will then run on a specialized cloud-based cluster. This unique approach simulates a real-world scenario where data accessibility is often restricted due to privacy regulations, ethical considerations, or logistical issues.

[1] Varadarajan AV, Bavishi P, Ruamviboonsuk P, et al. Predicting optical coherence tomography-derived diabetic macular edema grades from fundus photographs using deep learning. Nat Commun. 2020;11(1):130. doi:10.1038/s41467-019-13922-8