Automated Analysis for Retinopathy of Prematurity by Deep Neural Networks

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinopathy of Prematurity (ROP) is a significant cause of childhood blindness in premature infants.
• Existing automated ROP diagnosis methods primarily focus on 'plus' disease, neglecting disease staging.
• Deep neural networks show promise for medical image analysis, including ROP detection.

Purpose of the Study:

• To develop and evaluate a novel deep neural network architecture for ROP diagnosis.
• To accurately recognize the existence and severity (mild vs. severe) of ROP per examination.
• To address the gap in automated ROP staging methods.

Main Methods:

• A novel convolutional neural network architecture with two sub-networks and a feature aggregate operator was designed.
• The first sub-network extracts high-level features from fundus images.
• Features are fused and fed into the second sub-network for classification of ROP existence and severity using a large RetCam 3 dataset.

Main Results:

• The proposed deep learning model achieved high classification accuracy in recognizing ROP.
• The architecture effectively differentiates between mild and severe ROP cases.
• Experimental results validate the model's effectiveness on a large dataset.

Conclusions:

• The novel deep neural network architecture is effective for automatic ROP diagnosis, including severity assessment.
• This approach holds promise for improving early detection and management of ROP.
• Accurate ROP staging can aid in timely intervention and potentially reduce childhood blindness.