Automated machine learning model for fundus image classification by health-care professionals with no coding
Lucas Zago Ribeiro1, Luis Filipe Nakayama2,3, Fernando Korn Malerbi2
1Department of Ophthalmology and Visual Sciences, Federal University of São Paulo, São Paulo, SP, Brazil. lucaszagoribeiro@gmail.com.
Scientific Reports
|May 6, 2024
Summary
Code-free deep learning (CFDL) platforms show feasibility in ophthalmology for predicting outcomes from fundus images. Google Vertex outperformed Amazon Rekognition, demonstrating high accuracy for diabetic retinopathy and macular edema prediction.
Area of Science:
- Ophthalmology
- Artificial Intelligence
- Medical Imaging
Background:
- Deep learning models are increasingly used in medical image analysis.
- Code-free deep learning (CFDL) platforms offer accessibility for non-programmers.
- Assessing CFDL feasibility in ophthalmology is crucial for wider adoption.
Purpose of the Study:
- To evaluate the feasibility of CFDL platforms for predicting binary outcomes from fundus images.
- To compare the performance of Google Vertex and Amazon Rekognition CFDL platforms.
- To assess model performance on distinct ophthalmology datasets.
Main Methods:
- Utilized Messidor-2 and BRSET fundus image datasets.
- Developed deep learning models using Google Vertex and Amazon Rekognition CFDL platforms.
- Evaluated models using F1 score, AUC, precision, and recall without image preprocessing.
Main Results:
- CFDL models achieved high performance (AUC > 0.9) for diabetic retinopathy and macular edema.
- Google Vertex demonstrated superior performance over Amazon Rekognition.
- The BRSET dataset yielded the highest accuracy (AUC 0.994) with Google Vertex.
Conclusions:
- CFDL platforms are feasible for predicting binary outcomes from fundus images in ophthalmology.
- High accuracy was achieved in specific tasks, showcasing CFDL potential.
- CFDL platforms provide an accessible entry point for ophthalmologists into machine learning.
