Self-Supervised Feature Learning via Exploiting Multi-Modal Data for Retinal Disease Diagnosis

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Automatic diagnosis of retinal diseases from fundus images aids clinical decisions but requires extensive annotated data.
• Unsupervised/self-supervised learning methods reduce annotation needs but often use single imaging modalities.
• Multi-modal imaging, like combining fundus images with fundus fluorescein angiography (FFA), can enhance diagnostic accuracy for vitreoretinal diseases.

Purpose of the Study:

• To develop a novel self-supervised feature learning method that effectively utilizes multi-modal retinal data for disease diagnosis.
• To address the limitation of current self-supervised methods by incorporating multi-modal information.
• To learn both modality-invariant and patient-similarity features for improved diagnostic performance.

Main Methods:

• Synthesized corresponding fundus fluorescein angiography (FFA) modality from available data.
• Formulated a patient feature-based softmax embedding objective to learn shared semantic information across modalities.
• Developed a mechanism for the neural network to capture modality-invariant and patient-similarity features.

Main Results:

• The proposed method demonstrated superior performance compared to existing self-supervised feature learning techniques.
• The method achieved diagnostic accuracy comparable to supervised learning baselines.
• Evaluated on two public benchmark datasets for retinal disease diagnosis, confirming its effectiveness.

Conclusions:

• The novel self-supervised multi-modal approach significantly enhances retinal disease diagnosis.
• This method offers a promising alternative to supervised learning, reducing reliance on large annotated datasets.
• The approach effectively leverages cross-modal information for more robust and accurate automated diagnostics.