Intraocular Cytokine Level Prediction from Fundus Images and Optical Coherence Tomography.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• The link between retinal imaging and intraocular cytokine levels is understudied.
• No previous research has compared deep learning models based on fundus photography and OCT for cytokine prediction.

Purpose of the Study:

• To predict intraocular cytokine concentrations using deep learning with color fundus photographs (CFP) and optical coherence tomography (OCT) retinal images.
• To compare the predictive performance of CFP-based versus OCT-based deep learning models.

Main Methods:

• A deep learning pipeline using convolutional neural networks (ResNet18 via AutoGluon) was developed for feature extraction and regression.
• Four prediction approaches were evaluated: CFP alone, CFP with clinical data, OCT alone, and OCT with clinical data.
• Performance was assessed using the mean coefficient of determination (R²).

Main Results:

• Overall prediction performance was poor, with mean R² values below zero for all tested approaches.
• Image-only models outperformed models incorporating demographic/clinical data.
• No significant performance difference was found between CFP-based and OCT-based models.

Conclusions:

• Deep learning models using CFP or OCT images show limited ability to predict intraocular cytokine concentrations.
• Incorporating demographic or clinical data did not improve predictive performance and, in some cases, worsened it.
• Further research is needed to improve the accuracy of predicting cytokine profiles from retinal images.