KFWC: A Knowledge-Driven Deep Learning Model for Fine-grained Classification of Wet-AMD.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Wet Age-related Macular Degeneration (AMD) is a leading cause of blindness.
• Accurate diagnosis of wet-AMD subtypes, Neovascular AMD and Polypoidal Choroidal Vasculopathy (PCV), is challenging due to image similarity and data scarcity.
• Existing deep learning models struggle with fine-grained classification of wet-AMD subtypes.

Purpose of the Study:

• To develop a deep learning model for accurate fine-grained classification of wet-AMD subtypes.
• To address the challenges of insufficient data and high image similarity in wet-AMD diagnosis.
• To improve automated detection of Neovascular AMD and PCV.

Main Methods:

• Proposed a Knowledge-driven Fine-grained Wet-AMD Classification Model (KFWC).
• Implemented a two-stage approach: pre-training on 10 lesion signs and classification using human knowledge.
• Leveraged prior knowledge to enhance feature extraction for improved classification accuracy.

Main Results:

• KFWC achieved an AUC score of 99.71% on a clinical dataset.
• Outperformed the best baseline by 6.69% and human ophthalmologists by 4.14%.
• Demonstrated good interpretability and alleviated data collection/annotation pressures.

Conclusions:

• The knowledge-driven KFWC model effectively addresses data limitations and image similarity in wet-AMD fine-grained classification.
• The approach reduces the burden of data collection and annotation in medical image analysis.
• KFWC surpasses previous methods and human performance in diagnosing wet-AMD subtypes.