Multicolor image classification using the multimodal information bottleneck network (MMIB-Net) for detecting diabetic...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) detection relies on fundus images.
• Multicolor (MC) imaging captures multiple cSLO fundus views per patient.
• Limited research explores deep learning for MC image classification, especially using multimodal features.

Purpose of the Study:

• To propose a novel deep learning model for DR detection using MC images.
• To leverage multimodal information bottleneck theory for feature extraction and classification.
• To enhance the accuracy of DR diagnosis through advanced imaging analysis.

Main Methods:

• Developed a multimodal information bottleneck network (MMIB-Net).
• MMIB-Net extracts and represents features from multiple MC image modalities simultaneously.
• Classification is performed using combined representations from all modalities.

Main Results:

• The proposed MMIB-Net achieves accurate classification of MC images for DR detection.
• Comparative experiments confirm the effectiveness of multimodality and information bottleneck.
• The method demonstrates superior performance compared to existing approaches.

Conclusions:

• The MMIB-Net is a novel and effective approach for DR identification using MC images.
• Multimodal analysis combined with information bottleneck theory significantly improves classification performance.
• This represents the first application of a multimodal information bottleneck convolutional neural network for DR detection in MC images.