Segmentation Of Dry Keratitis From Fluorescein Staining Video Examinations Using Deep Learning Techniques

This research investigates the performance of an automated approach for evaluating dry eye keratitis and compares its performance with the expert grading method using the Oxford scoring system. By examining these techniques, we strive to establish their correlation and better understand the potential benefits and limitations of the automated method compared to the traditional expert grading approach. Our goal is to contribute to the ongoing advancement of dry eye keratitis assessment and explore the possibilities of integrating artificial intelligence to improve accuracy, efficiency, and treatment monitoring.

In this study, we collected a dataset of 43 fluorescein staining video examinations using blue light illumination and a yellow filter. These examinations were acquired with a slit lamp equipped with a camera (ION imaging system, Quantel(R)). We selected four representative frames from each video that accurately represent the keratitis condition (two for each eye). We randomly split these videos into a training set (20 videos), a validation set (9 videos), and a test set (9 videos).

In this study, we performed manual segmentation of the cornea on a fraction of the training set frames and then proceeded to train a U-Net-like model. We then processed the entire dataset of 152 frames to extract the corneal region. The areas affected by keratitis were annotated manually to create a corresponding mask for each frame. Using the training set, we further trained a model to predict keratitis-affected areas. Once we obtained predictions for the test frames, we used two distinct methods for analysis: first, we counted the predicted keratitis points in a manner consistent with the Oxford grading system (0 to 5), and second, we counted the total number of pixels identified as keratitis on the corneal surface.

The correlation between the Oxford score estimated by the ophthalmologist and the number of keratitis points estimated by our algorithm is high with a Pearson coefficient r= 0.84 (p<0.001). The correlation between the Oxford score estimated by the ophthalmologist and the pixel area corresponding to the keratitis areas is also significant with a Pearson coefficient r= 0.78 (p<0.001). The difference between the Oxford score estimated by the ophthalmologist and the one estimated by the algorithm was significant but remains low (0.41, p<0.05). There was no significant difference (p>0,05) when excluding the Oxford score class with the higher number of confluent points (i.e. Oxford score = 5). 

Our proposed deep learning segmentation algorithm, based on Convolutional Neural Networks (CNN), offers a promising solution and satisfactory results demonstrating a strong correlation with the established Oxford score. However, it is important to note that the presence of merging or coalescing lesions may result in an underestimation of the Oxford score for patients with more severe ocular damage. Despite this limitation, the proposed CNN-based algorithm holds significant potential for further development and optimization, ultimately enhancing dry eye keratitis assessment methods and paving the way for more accurate and efficient diagnostic processes.