Related Experiment Video
Updated: Sep 22, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.0K
A lightweight deep learning model for automatic segmentation and analysis of ophthalmic images
Parmanand Sharma1,2, Takahiro Ninomiya3, Kazuko Omodaka3,4
1Department of Ophthalmology, Tohoku University Graduate School of Medicine, Sendai, Japan. sharma@oph.med.tohoku.ac.jp.
Scientific Reports
|May 20, 2022
Summary
A new lightweight deep learning model accurately detects glaucoma from retinal OCTA images, outperforming commercial software in precision and efficiency. This model is ideal for resource-limited devices, enhancing disease diagnosis in tele-screening and self-monitoring applications.
Area of Science:
- Ophthalmic imaging analysis
- Medical image segmentation
- Deep learning applications in healthcare
Background:
- Accurate detection and diagnosis of ophthalmic diseases rely on extracting information from medical images.
- Deep learning (DL) models are essential for automating this information extraction process.
Purpose of the Study:
- To develop a lightweight deep learning model for automated and precise feature segmentation in ophthalmic images.
- To evaluate the model's performance in detecting glaucoma from optical coherence tomography angiography (OCTA) images and compare it with existing methods.
Main Methods:
- Developed a lightweight DL model utilizing image dimensionality reduction and channel contraction for feature extraction and segmentation.
- Evaluated the model on OCTA images for glaucoma detection, performing Bland-Altman analysis and calculating Pearson's correlation coefficient.
- Compared the model's performance against commercial software and a popular biomedical image segmentation model (Unet).
Main Results:
- Achieved high performance in glaucoma detection from OCTA images with an area under the receiver-operator characteristic curve (AUC) of approximately 0.81.
- Demonstrated exceptionally low bias (~0.00185) and high correlation (p=0.9969) compared to manual segmentation, significantly outperforming commercial software.
- The model is 10 times lighter than Unet, exhibiting superior segmentation accuracy and reproducibility across 3670 OCTA images.
- Achieved a high Dice similarity coefficient, indicating broad applicability for precise image segmentation in various ophthalmic conditions and potentially other fields.
Conclusions:
- The developed lightweight DL model offers precise and automated feature segmentation for ophthalmic diseases, particularly effective for glaucoma detection in OCTA images.
- The model's efficiency and accuracy surpass current commercial software and established DL architectures like Unet, with potential for deployment on resource-limited devices.
- The channel narrowing concept can significantly reduce model parameters, enhancing diagnostic accuracy for self-monitoring and tele-screening applications.
