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

Objective: Microbial keratitis (MK) is one of the leading causes of blindness in low- and middle-income countries that often requires timely diagnosis and subsequent treatment. Literature has shown that poor-quality slit-lamp photos (SLPs) can negatively impact artificial intelligence algorithm performance in keratitis classification. In this study, we develop and validate a deep learning (DL) model to assess SLP quality in MK. Design: Deep learning training and validation on a novel image dataset of MK eyes from a prospective clinical study. Subjects Participants and/or Controls: Slit lamp photos of MK with 4 illumination types between July 23, 2020, and May 1, 2024, prospectively collected during the Automated Quantitative Ulcer Analysis study. Methods Intervention or Testing: Slit lamp photo quality was classified as either "good" or "poor" based on a standardized grading protocol. Five DL-based classification models (AlexNet, ResNet50, DenseNet169, InceptionV3, and MobileNetV2) were trained and evaluated using fivefold cross validation. Model performance was visualized with gradient-weighted class activation mapping. Main Outcome Measures: Accuracy metrics of the model include accuracy and F1-score, which is harmonic mean of precision and recall. Gradient-weighted class activation mapping heatmap distribution of the model was also assessed. Results: We collected a range of 247 to 264 images for each illumination type from 138 individuals. The mean age was 54.0 ± 19.1 years, 58.8% were female, 89.9% were White, and 4.6% were Hispanic/Latino. The proportion of good-quality images varied by illumination types: 63% for diffuse white light (158/252), 51% for diffuse blue light (125/247), 42% for sclerotic scatter (108/256), and 23% for slit beam (61/264). MobileNetV2 achieved the highest performance in predicting SLP quality, with accuracy scores of 83.73 ± 6.02% (81.56 ± 9.16% F1-score) for diffuse white light, 79.79 ± 3.83% (79.13 ± 4.86% F1-score) for diffuse blue light, 79.91 ± 4.47% (68.63 ± 20.06% F1-score) for slit beam illumination, and 71.88 ± 4.69% (and 69.64 ± 10.34% F1-score) for sclerotic scatter. Conclusions: This study highlights the complexity of assessing SLP quality in MK. Future research is needed to examine how image quality affects automated decision-making in MK. Financial Disclosures: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.