Automated Detection of Retinal Detachment Using Deep Learning-Based Segmentation on Ocular Ultrasonography Images.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal detachment (RD) diagnosis relies on skilled interpretation of ocular ultrasonography (USG).
• Automated detection methods can enhance diagnostic accuracy and accessibility.
• Deep learning offers powerful tools for image analysis in healthcare.

Purpose of the Study:

• To develop an automated pipeline for detecting retinal detachment (RD) using deep learning-based segmentation on B-scan ocular ultrasonography (USG) images.
• To create a computational pipeline integrating segmentation and classification for improved RD detection.
• To evaluate the pipeline's performance against existing deep learning models.

Main Methods:

• A computational pipeline was developed using an encoder-decoder segmentation network and a machine learning classifier.
• The pipeline was trained and validated on 279 B-scan ocular USG images from 204 patients.
• Performance metrics including precision, recall, and F-scores were calculated for segmentation and RD detection.

Main Results:

• The automated pipeline achieved a 96.3% F-score for RD detection, outperforming end-to-end models (ResNet-50, MobileNetV3).
• Validation on an independent test set showed the pipeline's superior performance (96.5% F-score) compared to classification models (62.1%, 84.9%).
• The segmentation model demonstrated high F-scores for various ocular structures (retina/choroid: 84.7%, sclera: 78.3%, optic nerve sheath: 88.2%).

Conclusions:

• The proposed automated segmentation and classification method enhances RD detection in ocular USG images.
• This approach offers potential clinical benefits, particularly in resource-limited settings.
• The novel deep/machine learning pipeline shows promise for improving diagnostic accuracy and accessibility in ocular USG.