PADAr: physician-oriented artificial intelligence-facilitating diagnosis aid for retinal diseases

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Digital imaging offers promising avenues for early retinopathy detection and monitoring disease progression.
• Tracking retinopathic abnormalities over time is crucial for assessing disease risk.

Purpose of the Study:

• To develop an innovative, physician-oriented artificial intelligence (AI) system for diagnosing and monitoring multiple retinal diseases.
• The system aims to screen for various retinopathies and track potential abnormalities over time.

Main Methods:

• A dataset of 4908 fundus images with annotations for diabetic retinopathy, age-related macular degeneration, cellophane maculopathy, pathological myopia, and healthy controls was used.
• A VGG-based feature extractor and convolutional neural network classifiers were employed for screening.
• Image alignment via affine transforms and heatmap generation using gradient-weighted class activation mapping++ were utilized for abnormality localization.

Main Results:

• The screening model achieved 99% accuracy, 93% sensitivity, and 97% specificity in distinguishing retinopathy from healthy controls.
• For differentiating retinopathy types, the model yielded an average accuracy of 80%, sensitivity of 78%, specificity of 94%, F1-score of 79%, and Cohen's kappa of 0.70.
• Visualization techniques provided reasonable identification of candidate retinopathy sites.

Conclusions:

• The proposed AI model effectively extracts diagnostic information and lesion locations for retinal abnormalities.
• This capability enables clinicians to focus on patient-centered treatment by highlighting pathological plausibility within deep learning models.