An OCT retinal image classification model based on improved ResNet-34 network.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal diseases are a primary cause of vision loss, necessitating early detection.
• Manual diagnosis of retinal conditions from Optical Coherence Tomography (OCT) images is complex and time-consuming.
• Automated classification methods are crucial for efficient and accurate diagnosis.

Purpose of the Study:

• To develop an automated classification system for diabetic macular oedema (DME), choroidal neovascularisation (CNV), vitreous warts (Drusen), and normal retinal images using OCT.
• To enhance the feature extraction capabilities for improved diagnostic accuracy.
• To provide a reliable technical solution for early clinical diagnosis of retinal diseases.

Main Methods:

• A novel CBAM-AMP network (CANet) was designed, integrating the Convolutional Block Attention Module (CBAM) into a ResNet-34 architecture.
• Automatic Mixed Precision (AMP) and transfer learning were employed to accelerate training and improve model generalization.
• Data preprocessing techniques including median filtering, normalization, and data augmentation were utilized to optimize image quality and address class imbalance.

Main Results:

• The CANet model achieved a total classification accuracy of 0.9890 on the OCT-2017 dataset.
• Area Under the Curve (AUC) values reached 1 for all categories, with a recall of 1 for CNV.
• Ablation experiments confirmed the significant contributions of CBAM (0.9% accuracy increase), AMP (1.6% increase), and transfer learning (9.4% increase), demonstrating robustness to noisy data.

Conclusions:

• The integrated CANet model significantly enhances OCT image classification performance.
• The proposed method offers an efficient and robust solution for automated retinal disease diagnosis.
• This technology has the potential to assist clinicians in the early and accurate identification of sight-threatening retinal conditions.