Necessity of Local Modification for Deep Learning Algorithms to Predict Diabetic Retinopathy

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning (DL) algorithms show promise for diagnosing diabetic retinopathy (DR).
• Current DR detection models often rely on global datasets, potentially limiting their applicability in diverse populations.
• Assessing the need for model localization is crucial for universal DR screening.

Purpose of the Study:

• To evaluate the necessity of modifying deep learning algorithms for diabetic retinopathy screening across different regions.
• To compare the performance of Inception-v3, ResNet101, and DenseNet121 models using both global and local datasets.

Main Methods:

• Developed a DR severity detection model using the Kaggle Diabetic Retinopathy Detection dataset.
• Validated three DL architectures (Inception-v3, ResNet101, DenseNet121) on a local dataset from Taipei City Hospital.
• Evaluated model performance using the quadratic weighted kappa score (κ) and confusion matrix analysis.

Main Results:

• Inception-v3 performed best on the global dataset, while DenseNet121 excelled on the local dataset.
• All models showed a 5-8% higher performance on the local dataset compared to the global dataset.
• DL models tended to overestimate DR severity compared to local ophthalmologists' diagnoses.

Conclusions:

• Deep learning models trained on global data require local adaptation for accurate diabetic retinopathy diagnosis in clinical settings.
• Model localization is essential to ensure the reliability and applicability of AI-driven diagnostic tools.
• Further research into regionally optimized DL algorithms is needed for equitable healthcare outcomes.