Improved Training Efficiency for Retinopathy of Prematurity Deep Learning Models Using Comparison versus Class...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Retinopathy of prematurity (ROP) is a leading cause of childhood blindness.
• Accurate ROP diagnosis is crucial for timely intervention.
• Medical image classification using deep learning shows promise but faces challenges with data variability and scarcity.

Purpose of the Study:

• To compare the efficacy and efficiency of training neural networks for ROP image classification.
• To evaluate the use of comparison labels (relative disease severity) versus diagnostic class labels.

Main Methods:

• Deep learning neural networks were trained on expert-labeled retinal images from the i-ROP cohort.
• Networks were trained using either class labels or comparison labels indicating plus disease severity.
• Performance was evaluated using binary classification tasks (normal vs. abnormal, plus vs. nonplus) and AUC measurements.

Main Results:

• Networks trained with comparison labels achieved significantly higher AUC values in both classification tasks across datasets, indicating greater efficiency and accuracy.
• Comparison learning demonstrated superior performance even with fewer images.
• The performance gap between label types narrowed as training set size increased.

Conclusions:

• Comparison labels are more informative and abundant per sample than class labels for ROP image classification.
• This approach offers a potential solution to data variability and scarcity in medical AI training.
• Comparison-based learning enhances neural network performance in medical image analysis.