Differentiating malignant and benign eyelid lesions using deep learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Early diagnosis of eyelid malignancies is crucial for effective treatment and decision-making.
• Artificial intelligence (AI) offers potential as a screening tool for eyelid lesions.

Purpose of the Study:

• To evaluate the diagnostic performance of deep learning models in classifying eyelid lesions using clinical photographs.
• To compare the AI models' accuracy against human ophthalmologists.

Main Methods:

• A retrospective cross-sectional study involving 4954 photographs from 928 patients.
• Two convolutional neural network (CNN) models (DenseNet-161, EfficientNetV2-M) were fine-tuned for ternary (malignant, benign, none) and binary (malignant vs. benign) classification.
• Performance was assessed by comparing diagnostic accuracies and Area Under the Curve (AUC) with nine clinicians.

Main Results:

• For binary classification (malignant vs. benign), EfficientNetV2-M achieved 92.5% accuracy, comparable to clinicians (85.8-90.0%).
• DenseNet-161 achieved 87.5% accuracy in binary classification.
• The mean AUC for the CNN models was 0.908 and 0.950, indicating strong discriminatory power. Gradient-weighted class activation mapping visually confirmed lesion localization.

Conclusions:

• Deep learning models demonstrate significant potential for differentiating malignant from benign eyelid lesions using clinical photographs.
• AI models achieved performance levels similar to human observers, suggesting their utility in augmenting clinical decision-making for eyelid tumors.