Related Experiment Video
Updated: May 6, 2026
07:45
Tear-Derived Exosomal miR-15a as New Diagnostic Tool for Diabetic Retinopathy
Published on: December 30, 2025
831
Auto Machine Learning for Diabetic Retinopathy Screening: A Head-to-Head Multiplatform Comparison Against Human
Tomasz Krzywicki1, Ceren Durmaz Engin2, Andrzej Grzybowski3
1From the Faculty of Mathematics and Computer Science (T.K.), University of Warmia and Mazury, 10-719 Olsztyn, Poland; Faculty of Electronics, Telecommunications and Informatics, Gdansk University of Technology, Poland.
American Journal of Ophthalmology
|May 4, 2026
Summary
Automated machine learning (AutoML) platforms show promise for diabetic retinopathy screening. Amazon SageMaker Canvas and AutoGluon performed best, highlighting the need for validation and threshold calibration in diverse clinical settings.
Area of Science:
- Ophthalmology
- Artificial Intelligence
- Medical Imaging
Background:
- Diabetic retinopathy (DR) is a leading cause of vision loss.
- Automated machine learning (AutoML) offers potential for efficient DR screening.
- Benchmarking AutoML platforms is crucial for clinical adoption.
Purpose of the Study:
- To evaluate and compare the performance of multiple AutoML platforms for DR screening using fundus photographs.
- To establish a unified framework for training and evaluating AutoML models.
- To use human consensus grading and an FDA-approved system (IDx-DR) as reference standards.
Main Methods:
- A retrospective diagnostic performance and benchmarking study was conducted.
- Image classifiers were trained on large public datasets (APTOS, DDR, EyePACS) and evaluated on an independent cohort (n=726).
- Platforms evaluated included Google Vertex AI, Amazon Rekognition, Amazon SageMaker Canvas, AutoGluon, AutoKeras, and Apple CreateML, assessing performance for any DR, referable DR (RDR), and sight-threatening DR (STDR).
Main Results:
- Amazon SageMaker Canvas and AutoGluon demonstrated strong performance, with AUCs up to 0.96 for STDR and 0.93-0.94 for RDR.
- Canvas offered balanced RDR performance (sensitivity 88.3%, specificity 85.5%), while AutoGluon prioritized sensitivity for any DR (95.9%).
- Google Vertex AI showed weaker, unstable performance; Amazon Rekognition and Canvas showed highest agreement with IDx-DR.
Conclusions:
- AutoML platforms can achieve clinically meaningful performance for diabetic retinopathy screening.
- Performance variations highlight the importance of platform selection and adaptability to clinical settings.
- External validation and careful threshold calibration are essential for reliable DR screening.
