Computer-aided diagnosis of eyelid skin tumors using machine learning

Area of Science:

• Ophthalmology
• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Eyelid skin tumors require accurate and timely diagnosis.
• Current diagnostic methods can be time-consuming and may require specialized expertise.
• Automated diagnostic tools can potentially improve efficiency and accessibility of care.

Purpose of the Study:

• To develop and validate a machine learning (ML) framework for the automated diagnosis of malignant eyelid skin tumors.
• To compare the diagnostic performance of the ML model against human evaluation.

Main Methods:

• The study utilized a dataset of eyelid lesion images from Sheba Medical Center, fine-tuned after pretraining on the International Skin Imaging Collaboration (ISIC) 2019 dataset.
• A total of 559 images (373 benign, 186 malignant) were used for model training and validation.
• Gradient-weighted Class Activation Mapping (Grad-CAM) was employed for model interpretability.

Main Results:

• The developed ML model achieved a sensitivity of 93.8% (95% CI 80.0-100.0%) and a specificity of 73.7% (95% CI 60.0-87.1%) in diagnosing malignant eyelid lesions.
• The model's decision-making process was visualized using heatmap techniques.
• Performance was comparable to human diagnostic capabilities.

Conclusions:

• The study presents a reliable model-aided diagnostic technology for eyelid skin lesions.
• This ML framework demonstrates potential to assist clinicians in differentiating between benign and malignant lesions.
• The technology could reduce healthcare system burden and enhance clinical efficiency, particularly in remote regions.