Publicación:

Improving corneal endothelium image analysis: advanced image annotation for image classification and segmentation

Resumen en español

The corneal endothelium, a single layer of hexagonal cells lining the inner cornea, is vital for ocular health. Assessing its condition relies on morphometric analyses of corneal endothelial cells from high-quality images. Unfortunately, disease presence or imaging artifacts frequently impede automatic analysis. This thesis introduces a deep learning and image annotation-based methodology for creating precise classification and segmentation models applicable to both normal and diseased eyes. We detail a comprehensive statistical analysis of corneal endothelial cell annotations by a group of graders, which aids in compiling datasets that can predict segmentation complexity from the quality of specular microscopy images. Subsequently, we estimate morphometric parameters along with the complexity report on categorized images. A convolutional regression network, trained to rate image quality on a scale from 0 to 6, informs a second regression network that qualitatively evaluates cell segmentations, signaling the dependability of parameter estimations. In conclusion, we discuss the outcomes, proposed goals, encountered challenges, and avenues for future research enhancement. Our work supports clinicians in evaluating corneal health for patients with or without Fuchs’ Dystrophy.