RFMiD: Retinal Image Analysis for multi-Disease Detection challenge.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Existing deep learning models for retinal disease diagnosis focus on common pathologies like diabetic retinopathy, glaucoma, and age-related macular degeneration.
• Computer-aided diagnosis systems often overlook rare but sight-threatening conditions, limiting their clinical adoption by ophthalmologists.
• There is a need for automated systems capable of detecting a wider spectrum of retinal diseases, including both frequent and rare conditions.

Purpose of the Study:

• To advance the state-of-the-art in automatic ocular disease classification by including rare pathologies alongside common ones.
• To organize a grand challenge, "Retinal Image Analysis for multi-Disease Detection" (ISBI 2021), to foster research in this area.
• To introduce a new dataset, the "Retinal Fundus Multi-disease Image Dataset" (RFMiD), for training and evaluating multi-disease detection models.

Main Methods:

• The challenge featured two sub-challenges: binary disease screening and 28-class multi-label disease classification.
• A new dataset, RFMiD, comprising fundus images for multi-disease detection was utilized.
• Top-performing solutions involved a combination of data preprocessing, data augmentation, pre-trained models, and model ensembling.

Main Results:

• The challenge received 74 submissions, indicating significant interest from the scientific community.
• The top methodologies demonstrated the effectiveness of combining various AI techniques for multi-disease detection.
• The results pave the way for more generalizable retinal screening models that account for both common and rare diseases.

Conclusions:

• The developed multi-disease detection approach, incorporating rare pathologies, enables the creation of more comprehensive retinal screening tools.
• This work moves beyond single-disease detection systems, addressing a critical gap in current automated ophthalmology tools.
• The success of the challenge and the performance of the top methods highlight the potential of AI in improving the scope and accuracy of retinal disease diagnosis.