Disentangling representations of retinal images with generative models.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal fundus images are vital for early eye disease detection.
• Technical variations in fundus images, like camera type, can hinder AI reliability.
• AI models may learn superficial patterns instead of causal factors in image data.

Purpose of the Study:

• To develop a population model for retinal fundus images.
• To disentangle patient attributes from camera effects in fundus images.
• To enable controllable and realistic generation of retinal fundus images.

Main Methods:

• Proposed a population model for retinal fundus images.
• Introduced a disentanglement loss based on distance correlation.
• Utilized qualitative and quantitative analyses to validate the model.

Main Results:

• The model successfully disentangles patient attributes from camera effects.
• Learned information is encoded in disentangled subspaces.
• Enabled controllable and realistic image generation based on learned subspaces.

Conclusions:

• The proposed disentanglement loss is effective for retinal image analysis.
• The model facilitates more reliable AI applications in ophthalmology.
• Achieved controllable and realistic generation of retinal fundus images.