Related Experiment Video
Updated: Oct 5, 2025
14:08
Automated Midline Shift and Intracranial Pressure Estimation based on Brain CT Images
Published on: April 13, 2013
42.9K
Asymmetry between right and left fundus images identified using convolutional neural networks
Tae Seen Kang1, Bum Jun Kim1, Ki Yup Nam2
1Department of Ophthalmology, Gyeongsang National University Changwon Hospital, #11 Samjeongja-ro, Seongsan-gu, Changwon, 51472, Republic of Korea.
Scientific Reports
|January 28, 2022
Summary
Convolutional neural networks (CNNs) can accurately distinguish between right and left fundus images, identifying the macula as a key feature. However, data augmentation using image flipping may introduce bias in machine learning models.
Area of Science:
- Ophthalmology
- Computer Vision
- Medical Imaging
Background:
- Fundus image analysis is crucial for diagnosing ocular diseases.
- Distinguishing between right and left eye images is essential for accurate AI-driven diagnostics.
- Previous studies often use image flipping for data augmentation, potentially introducing bias.
Purpose of the Study:
- To evaluate the capability of convolutional neural networks (CNNs) in discriminating between right and left fundus images.
- To identify discriminative regions within fundus images utilized by CNNs.
- To assess the impact of data augmentation techniques on CNN performance.
Main Methods:
- Trained multiple CNN architectures (DenseNet121, ResNet50, VGG19) on a large dataset of fundus images (98,038 images).
- Utilized the Ocular Disease Intelligent Recognition dataset for augmentation.
- Employed class activation mapping to visualize discriminative image regions.
Main Results:
- CNNs achieved high accuracy (>99.3% and >91.1%) in differentiating right and left fundus images.
- Model accuracy varied with CNN architecture depth and complexity.
- Class activation mapping pinpointed the macula as the primary discriminative region.
- DenseNet121 showed no significant discrimination for left-eye-only images (55.1%).
Conclusions:
- CNNs are highly effective in distinguishing lateral eye images from fundus photographs.
- The macula is a critical anatomical landmark for lateral eye discrimination by CNNs.
- Caution is advised when using horizontal flipping for data augmentation in fundus image datasets to prevent bias.
