Enhancing automated strabismus classification with limited data: Data augmentation using StyleGAN2-ADA.

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence

Background:

• Designing deep learning systems for medical diagnosis, such as automated strabismus diagnosis, is often hindered by severely limited datasets.
• Traditional data augmentation methods may not be sufficient to address extreme data scarcity in medical imaging.

Purpose of the Study:

• To propose and evaluate a generative data augmentation technique to improve deep learning-based automated strabismus diagnosis systems.
• To assess the effectiveness of generative models in overcoming data limitations and mitigating overfitting.

Main Methods:

• Implemented a generative model based on the StyleGAN2-ADA architecture.
• Utilized two classifiers to assess strabismus classification performance.
• Compared the proposed generative method against traditional data augmentation techniques.
• Validated generative samples using classifier performance and explored their correlation with ophthalmologist diagnosis agreement.

Main Results:

• The generative data augmentation technique demonstrated substantial performance enhancement compared to traditional methods.
• The approach effectively mitigated overfitting issues in deep learning model training.
• Generative samples were validated for practicality beyond standard metrics like FID.
• A relationship between ophthalmologist agreement and generative model performance was explored.

Conclusions:

• Generative model-based data augmentation significantly improves quantitative performance in scenarios with extreme data scarcity for strabismus diagnosis.
• This technique offers a practical solution to enhance deep learning model robustness and accuracy when training data is limited.