Deep Learning to Distinguish ABCA4-Related Stargardt Disease from PRPH2-Related Pseudo-Stargardt Pattern Dystrophy

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Recessive Stargardt disease (STGD1) and pseudo-Stargardt pattern dystrophy (PSPD) present similar phenotypes, complicating clinical diagnosis.
• Both conditions are forms of inherited retinal diseases (IRDs).

Purpose of the Study:

• To evaluate a deep learning classifier's ability to differentiate STGD1 from PSPD using fundus autofluorescence (FAF) images.
• To compare the AI classifier's diagnostic performance against that of retinal specialists.

Main Methods:

• A convolutional neural network (CNN) was trained on 729 FAF images and fine-tuned using transfer learning on 370 FAF images for STGD1/PSPD classification.
• Retinal specialists evaluated the same dataset to compare diagnostic accuracy, sensitivity, and precision.

Main Results:

• The CNN achieved a test accuracy of 0.882, with an AUROC of 0.890, precision of 0.883, and sensitivity of 0.883.
• The AI classifier outperformed human experts, with average accuracies of 0.816 for specialists and 0.724 for fellows.

Conclusions:

• A pretrained CNN can effectively distinguish STGD1 from PSPD, demonstrating good accuracy even with limited data.
• This AI approach shows promise as a diagnostic aid for differentiating these visually impairing retinal conditions.