Abstract
Purpose :
Clinically distinguishing between active corneal lesions that require treatment and those in remission that do not poses a challenge in bacterial keratitis. This study was aimed to develop a deep learning-based model to classify the active and remission states of bacterial keratitis using three types of anterior segment images: broad-beam, slit-beam, and scatter.
Methods :
Image data were collected from patients who were diagnosed and treated for bacterial keratitis at Samsung Medical Center from January 1, 2008, to December 31, 2022. Two ophthalmologists (Y.K. and D.H.L.) confirmed diagnosis by corneal culture tests and classified active state and remission state with annotating lesions on images. We generated patches from the original data by using patch technique to preprocess the data, resulting in a new patch dataset. We set VGG16 as a baseline Convulution Neural Network classifier. The model was trained on the patch dataset and the original dataset, and the performances were compared. For evaluation metrics, Area Under the Receiver Operating Characteristic (AUROC), accuracy, sensitivity, and specificity were used.
Results :
A total of 903 images from 63 patients were utilized. The data were divided into training and evaluation sets in an 8:2 ratio based on the number of patients for each image type. By applying the patch technique, the highest AUROC of the model trained from the original was improved from 0.802 (Table 1) to 0.897 (Table 2). The slit image alone exhibited the highest AUROC in both dataset when compared to broad-beam, scatter, and any two to three images combined.
Conclusions :
We developed a deep learning-based classification model for bacterial keratitis activity. The implementation of slit-beam images with patch dataset yielded superior classification results when compared to various combinations of anterior segment image types, as well as the original dataset.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.