Laceration assessment: advanced segmentation and classification framework for retinal disease categorization in...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal disorders threaten vision, exacerbated by factors like aging and diabetes.
• Optical coherence tomography (OCT) detects early abnormalities but requires segmentation for precise quantification.
• Current methods face challenges in simultaneously classifying diseases and segmenting retinal features.

Purpose of the Study:

• To develop an innovative deep learning framework for simultaneous OCT image classification and retinal laceration segmentation.
• To enhance the diagnostic capabilities for various retinal diseases using OCT imaging.
• To improve the quantification of retinal abnormalities through accurate image segmentation.

Main Methods:

• A parallel mask-guided convolutional neural network (PM-CNN) was developed for OCT B-scan classification.
• A grade activation map (GAM) from PM-CNN guided a V-Net network (GAM V-Net) for segmentation.
• The framework was trained and validated on a combined dataset of four public and one real-time dataset, covering 11 retinal disease categories.

Main Results:

• The dual framework achieved a classification accuracy of 99.10±0.10% for retinal diseases.
• The segmentation performance yielded a Dice coefficient of 78.33±0.15% for retinal lacerations.
• The model demonstrated strong generalizability by effectively segmenting retinal fluids and identifying lacerations on unseen data.

Conclusions:

• The proposed deep learning framework offers a powerful tool for simultaneous diagnosis and quantification of retinal diseases from OCT scans.
• This integrated approach improves diagnostic efficiency and accuracy in ophthalmology.
• The framework's generalizability highlights its potential for clinical application in diverse retinal conditions.