Deep Learning-Based Classification of Slit-Lamp Photograph Quality in Microbial Keratitis.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Microbial keratitis (MK) is a leading cause of blindness, particularly in low- and middle-income countries.
• High-quality slit-lamp photos (SLPs) are essential for accurate artificial intelligence (AI) analysis of keratitis.
• Poor SLP quality can significantly impair the performance of AI algorithms used for keratitis classification.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for assessing the quality of SLPs in patients with MK.
• To improve the reliability of AI-driven diagnostic tools for MK by ensuring high-quality input data.

Main Methods:

• A novel dataset of SLPs from MK patients was prospectively collected under four different illumination conditions.
• Five DL models (AlexNet, ResNet50, DenseNet169, InceptionV3, MobileNetV2) were trained and evaluated using fivefold cross-validation.
• Model performance was assessed using accuracy and F1-scores, with visualization via gradient-weighted class activation mapping.

Main Results:

• The MobileNetV2 model achieved the highest performance in predicting SLP quality across different illumination types.
• Accuracy scores for MobileNetV2 ranged from 71.88% to 83.73%, with corresponding F1-scores ranging from 69.64% to 81.56%.
• The proportion of good-quality images varied significantly by illumination type, highlighting the challenges in data acquisition.

Conclusions:

• Assessing SLP quality in MK presents complexities that impact AI performance.
• Further research is necessary to understand the influence of image quality on automated decision-making processes in MK diagnosis.
• The developed DL model shows promise in standardizing SLP quality assessment for improved AI applications in ophthalmology.