Fundus image analysis of retinitis pigmentosa using artificial intelligence

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Retinitis pigmentosa (RP) is an inherited retinal disease causing progressive vision loss.
• Early diagnosis of RP is crucial for management but often delayed due to subtle early symptoms.
• Color fundus images, common in checkups, are underutilized for RP detection.

Purpose of the Study:

• To evaluate the efficacy of finetuning deep learning models for identifying retinitis pigmentosa (RP) from color fundus images.
• To assess the diagnostic performance of pretrained convolutional neural network (CNN) models adapted for RP detection.
• To explore the interpretability of deep learning models in identifying key fundus regions relevant to RP.

Main Methods:

• Transfer learning was employed to finetune pretrained CNN models (VGG16, Resnet50, InceptionV3).
• A dataset of 321 color fundus images from 201 Japanese subjects (107 RP patients, 94 controls) was utilized.
• Gradient-weighted Class Activation Mapping (Grad-CAM) was used to visualize model attention.

Main Results:

• The InceptionV3 model achieved the highest diagnostic accuracy of 96.97% for RP detection.
• Model performance closely matched the average diagnostic accuracy of ophthalmologists.
• Grad-CAM analysis indicated the model focused on clinically relevant retinal areas, suggesting interpretability.

Conclusions:

• Finetuned deep learning models demonstrate high accuracy in detecting retinitis pigmentosa from fundus images.
• AI-powered tools can potentially serve as supportive systems for ophthalmologists in early RP screening.
• The interpretability of the models supports their clinical utility in identifying disease-specific features.