Processing large datasets using artificial intelligence is a promising approach in disease diagnosis and monitoring that focuses on improving research algorithms for existing technologies. Interest in studying corneal nerve fibers (CNFs) arises not only from the need to understand the pathogenesis and progression of various ocular diseases but also from the potential for thin, unmyelinated corneal nerves to be used as biomarkers for systemic polyneuropathies.

PURPOSE

This study evaluates the preliminary results of using a neural network-based algorithm for analysis of confocal images of CNFs.

MATERIAL AND METHODS

The comparative study of CNFs was conducted in a group of 50 healthy volunteers (100 eyes) aged 25 to 55 years without concomitant ocular or systemic diseases. Confocal microscopy of the central cornea was performed to assess the state of CNFs. Image analysis and nerve recognition were carried out using special software (Liner calculate, Liner 1.2S) and a newly developed neural network-based algorithm.

RESULTS

The study considered suitable encoders for image processing, including ResNet_50, VGG_16, and InceptionResNetV2. Among these, images processed with the VGG_16 encoder in Imagenet mode demonstrated the highest quality. Quantitative CNF parameters (length and density of the main trunks, macrophage count, anisotropy and symmetry coefficients) were comparable between the regular software and the neural network-based algorithm.

CONCLUSION

The results indicate the potential of using neural networks, particularly the VGG_16 encoder family, for structural assessment of the CNFs. Key advantages of the proposed algorithm include improved quality of image interpretation and reduced time required for analysis.