Related Experiment Video
Updated: Feb 19, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.6K
Multi-categorical deep learning neural network to classify retinal images: A pilot study employing small database.
Joon Yul Choi1, Tae Keun Yoo2, Jeong Gi Seo2
1Department of Electrical and Computer Engineering, Seoul National University, Seoul, South Korea.
Plos One
|November 3, 2017
Summary
Deep learning models for retinal disease detection show diminished performance with more categories. Transfer learning with ensemble classifiers improved accuracy for specific conditions, but large datasets are needed for clinical use.
Area of Science:
- Ophthalmology
- Medical Imaging Analysis
- Computer-Aided Diagnosis
Background:
- Deep learning (DL) offers a powerful approach for analyzing medical images, particularly fundus photographs for retinal disease detection.
- Automated detection systems using DL aim to assist in diagnosing various retinal conditions from fundus images.
Purpose of the Study:
- To apply deep learning convolutional neural network (CNN) using MatConvNet for automated detection of multiple retinal diseases from fundus photographs.
- To evaluate the impact of dataset categorization on DL model performance for retinal disease classification.
- To investigate the effectiveness of transfer learning and ensemble classifiers in improving multi-categorical retinal disease detection.
Main Methods:
- Utilized the STructured Analysis of the REtina (STARE) database, expanding it to 10 categories (normal retina and nine diseases).
- Employed a random forest transfer learning model based on the VGG-19 architecture.
- Tested ensemble classifiers, including clustering and voting approaches, in conjunction with transfer learning.
Main Results:
- DL model performance decreased significantly as the number of categories increased, achieving only 30.5% accuracy for 10 categories.
- Classifying three integrated categories (normal, diabetic retinopathy, dry age-related macular degeneration) yielded higher accuracy (72.8%).
- The best performance for 10-category classification using transfer learning with an ensemble classifier was 36.7% accuracy.
Conclusions:
- The study highlights the ineffectiveness of current DL techniques for multi-class retinal disease detection in clinics due to small dataset sizes.
- Transfer learning combined with ensemble classifiers shows potential for improving multi-categorical retinal disease detection performance.
- Further research with large-scale hospital datasets is crucial to validate the clinical applicability of these algorithms.
