Deep learning-based classification of retinal atrophy using fundus autofluorescence imaging.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Retinal atrophy can stem from various causes, including age-related macular degeneration (AMD) and inherited retinal diseases (IRDs).
• Differentiating between these etiologies using fundus autofluorescence (FAF) imaging is crucial for accurate diagnosis and treatment.
• Current diagnostic methods may face challenges in distinguishing between geographic atrophy (GA) secondary to AMD and atrophy from IRDs like Stargardt disease (STGD1) or Pseudo-Stargardt Pattern Dystrophy (PSPD).

Purpose of the Study:

• To develop and evaluate a deep learning model for the automatic classification of retinal atrophy based on its etiology using FAF images.
• To differentiate between geographic atrophy (GA) associated with AMD and atrophy secondary to IRDs.

Main Methods:

• A multi-layer deep convolutional neural network (CNN) was trained using 314 FAF images (110 GA, 204 IRDs).
• The model was trained and validated on 251 images, with subsequent testing on 63 untrained images.
• Two approaches were used: initial training/validation/testing and 10-fold cross-validation to assess performance.

Main Results:

• The first approach achieved a peak accuracy of 0.92 and an AUC-ROC of 0.981.
• Mean accuracy in the first approach was 87.30 ± 2.96%.
• The second approach (10-fold cross-validation) yielded a mean accuracy of 0.79 ± 0.06.

Conclusions:

• A deep learning algorithm can effectively classify retinal atrophy according to its etiology from FAF images.
• The developed model demonstrates good accuracy and AUC-ROC values for differentiating GA from IRDs.
• This AI-driven approach shows promise for improving the differential diagnosis of retinal atrophy in clinical practice.