A deep learning approach to identify blepharoptosis by convolutional neural networks.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Blepharoptosis is a common cause of reversible vision loss.
• Accurate diagnosis of blepharoptosis can be challenging, particularly for non-specialists.
• Current automated diagnostic methods often disrupt clinical workflows.

Purpose of the Study:

• To develop an artificial intelligence algorithm for accurate blepharoptosis identification from clinical photographs.
• To overcome limitations of existing automated diagnostic techniques.

Main Methods:

• Trained convolutional neural networks (CNNs) on 500 clinical photographs of patients with and without blepharoptosis.
• Utilized images sourced from a tertiary ophthalmic center.
• Compared performance of various CNN architectures, including DenseNet121 and Resnet34, with and without pre-training.

Main Results:

• Most trained CNN models achieved reasonable accuracy in identifying ptosis.
• The DenseNet121 architecture without pre-training demonstrated the highest performance (90.1% sensitivity, 82.4% specificity).
• Models without pre-training showed slightly higher accuracy (85.8%) compared to pre-trained models (82.6%), though pre-trained models trained faster.

Conclusions:

• AI can accurately diagnose blepharoptosis from clinical photos without requiring reference markers or user input.
• Several CNN architectures are effective for this task.
• DenseNet121 and Resnet18 architectures without pre-training performed best on the developed dataset.