Related Experiment Video
Updated: Jun 12, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.7K
Performance analysis of pretrained convolutional neural network models for ophthalmological disease classification
1Department of Biostatistics, Faculty of Medicine, Izmir Katip Celebi University, Izmir, Turkey.
Arquivos Brasileiros De Oftalmologia
|September 19, 2024
Summary
Convolutional neural networks (CNNs) show promise in classifying eye diseases from fundus images. ResNet50 demonstrated superior performance in detecting various ophthalmological conditions.
Area of Science:
- Ophthalmology
- Artificial Intelligence
- Medical Imaging
Background:
- Diabetic retinopathy, glaucoma, and age-related macular degeneration are leading causes of vision loss.
- Early detection and classification of these diseases are crucial for effective treatment.
- Fundus image analysis using deep learning offers a potential solution for large-scale screening.
Purpose of the Study:
- To evaluate the classification performance of pretrained convolutional neural network (CNN) architectures.
- To assess the ability of VGG16, Inceptionv3, and ResNet50 models to detect eight different eye diseases from fundus images.
Main Methods:
- Utilized a public ocular disease intelligent recognition database with 10,000 fundus images.
- Implemented and compared three pretrained CNN models: VGG16, Inceptionv3, and ResNet50.
- Augmented training data and split the dataset into 70% training, 10% validation, and 20% testing sets.
Main Results:
- ResNet50 achieved 97.1% accuracy, excelling in cataract classification (AUC=0.964, score=0.903).
- VGG16 achieved 96.2% accuracy with high specificity (99.2%) but lower sensitivity (56.9%).
- Specific models showed varying strengths in classifying different diseases, with ResNet50, Inceptionv3, and VGG16 showing particular utility for certain conditions.
Conclusions:
- Pretrained CNN architectures effectively identify ophthalmological diseases from fundus images.
- ResNet50 is suitable for classifying glaucoma, cataract, hypertension, and myopia.
- Inceptionv3 and VGG16 show promise for other specific disease classifications, including diabetic retinopathy.
