Multi-resolution visual Mamba with multi-directional selective mechanism for retinal disease detection.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Retinal diseases impact quality of life and incur high medical costs.
• Optical coherence tomography (OCT) is crucial for diagnosing retinal conditions.
• Current deep learning models like CNNs struggle with global context, while Transformers face computational challenges with long-range dependencies.

Purpose of the Study:

• To introduce a novel deep learning model, the Multi-Resolution Visual Mamba (MRVM), for enhanced OCT image classification.
• To address the limitations of existing models in capturing both local and global features in retinal images.
• To improve the accuracy and efficiency of automated retinal disease diagnosis.

Main Methods:

• The MRVM model integrates convolutional layers for local feature extraction and a "retinal Mamba" component for global dependency capture.
• A multi-directional selection mechanism (MSM) is employed within the retinal Mamba to enhance feature extraction across various orientations.
• The model processes multi-scale global features to improve classification performance.

Main Results:

• The MRVM model achieved superior accuracy in differentiating retinal images with various lesions.
• Achieved overall accuracies of 98.98% and 96.21% on two public datasets.
• Demonstrated enhanced detection accuracy compared to traditional deep learning methods.

Conclusions:

• The MRVM model offers a powerful new approach for the accurate identification of retinal diseases using OCT imaging.
• This method provides a novel perspective for developing robust AI algorithms in medical image-assisted diagnosis.
• The MRVM model has the potential to significantly improve diagnostic capabilities in ophthalmology.