Assessing the Impact of Image Quality on Deep Learning Classification of Infectious Keratitis

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Corneal ulcers are a significant cause of visual impairment.
• Accurate and timely diagnosis of bacterial and fungal keratitis is crucial for effective treatment.
• Convolutional neural networks (CNNs) show promise in analyzing medical images for disease detection.

Purpose of the Study:

• To evaluate the influence of specific corneal photograph quality parameters on the diagnostic performance of a CNN for keratitis.
• To identify which image quality factors, if any, affect CNN predictions in classifying bacterial and fungal keratitis.

Main Methods:

• A CNN model trained for keratitis classification was tested on ulcer photographs.
• Images were assessed for 5 quality parameters: gaze direction, eyelid position, exposure, focus, and light reflection.
• CNN performance was quantified using ROC and precision-recall curves; heatmaps visualized influential regions.

Main Results:

• CNN performance was significantly improved when light reflection or obscuring eyelids were present.
• No other tested image quality parameter demonstrated a significant impact on CNN performance.
• Gradient class activation heatmaps indicated the corneal infiltrate was the primary focus for CNN predictions.

Conclusions:

• The developed CNN demonstrates robust, expert-level performance in diagnosing keratitis, largely independent of common image quality variations.
• Future research should explore smartphone-based imaging and datasets with broader image quality ranges to further validate CNN performance.