Posters

Establishing a deep learning algorithm for the classification and segmentation of in vivo confocal microscopy images

Poster Details

First Author: N.Koseoglu TURKEY

Co Author(s): I. Kovler   M. Ozmen   A. Cohen   R. Soferman   P. Hamrah        

Abstract Details

Purpose:

In vivo confocal microscopy (IVCM) is a non-invasive imaging tool providing images of the cornea on a cellular level and also efficient in detection of corneal immune dendritiform cells (DCs). Visualization of DCs via IVCM has the potential to be utilized as a clinical tool to assess the effects of inflammation on corneal structures and function in inflammatory diseases such as dry eye disease (DED). However, the detection and analyses of DCs in IVCM images is highly subjective and time consuming. Therefore, in this study we propose a deep learning algorithm for analyses of DCs in IVCM images.

Setting:

Retrospective study conducted on images obtained from DED patients at the cornea clinic of Tufts Medical Center, Department of Ophthalmology.

Methods:

For classification of corneal layers, 2140 images obtained from healthy epithelium, subbasal nerve plexus, stroma, and endothelium were randomly selected. The algorithm was trained on 1540 images and validated on the remaining 610. A total of 258 images were utilized in the segmentation of DCs in IVCM images. The algorithm was trained on 233 and validated on the remaining 25 images. This process was repeated 5 times in random image sets as a standard cross validation strategy. Morphological parameters including DC density, area and perimeter were also analyzed.

Results:

The algorithm showed high sensitivity and specificity for the classification of corneal layers for all layers (>0,95 for both sensitivity and specificity, mean AUC >0,95). While the mean sensitivity for DC segmentation was 0,66, the specificity was recorded as 0,99 (mean AUC=0,84). The interclass correlation between semi-automated analyses and the deep learning algorithm were above 0,99 for all parameters(p<0,05).

Conclusions:

The proposed deep learning algorithm shows high sensitivity and specificity in the classification of IVCM images as well as segmentation and quantification of DCs. Our study suggests that artificial intelligence can rapidly evaluate IVCM images while maintaining a high degree of accuracy.

Financial Disclosure:

None