Multi-modal classification of retinal disease based on convolutional neural network.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal diseases like age-related macular degeneration and diabetic retinopathy cause irreversible blindness if not diagnosed and treated promptly.
• Optical Coherence Tomography (OCT) and OCT Angiography (OCTA) offer complementary retinal views, enhancing diagnostic accuracy when integrated.
• Current diagnostic methods can be limited by human error, cost, and the need for specialized expertise.

Purpose of the Study:

• To develop and validate a multi-modal deep learning classification model for automated retinal disease diagnosis.
• To improve the accuracy and efficiency of retinal disease classification by integrating OCT and OCTA imaging modalities.
• To address challenges of limited training data and class imbalance in deep learning models for medical image analysis.

Main Methods:

• A novel two-branch multi-modal classification model was designed, integrating OCT and OCTA images using intermediate fusion.
• Preprocessing techniques included bright line cropping to remove black edges and preserve lesion features.
• Data augmentation and loose matching methods were employed to overcome insufficient data, coupled with a two-step training strategy.

Main Results:

• The multi-modal model achieved high performance metrics on an external test set: 0.9667 average accuracy, 0.9418 precision, 0.8569 recall, 0.9422 specificity, and 0.8921 F1-Score.
• The proposed model demonstrated superior accuracy compared to single-modal models and early/late fusion multi-modal approaches.
• The model effectively handles limited and imbalanced training datasets, showing robust performance.

Conclusions:

• The integrated OCT and OCTA multi-modal model offers a more accurate and efficient approach to diagnosing retinal diseases.
• This AI-driven solution can reduce human error, lower screening costs, and enable more uniform and effective mass screening for retinal conditions.
• The developed model provides a viable solution for enhancing deep learning performance in medical imaging, particularly with limited or imbalanced data.