Use of convolutional neural networks in confocal microscopy of corneal nerve fibers

Relevance Corneal confocal microscopy (CCM) is a non-invasive method for assessing thin unmyelinated nerve fibers (TUNF). This technique has proven its diagnostic value in neurodegenerative diseases such as Parkinson's disease, consequences of coronavirus infection and a number of other pathologies [1–5]. The algorithm developed by the authors based on convolutional neural networks (CNN) made it possible to reduce the analysis time and objectively assess the "bead-likeness" of nerve fibers, previously determined manually. Objective To determine the performance, specificity and sensitivity of the method for assessing the state of corneal nerve fibers (CNF) using CNN in patients with neurodegenerative conditions and mechanical damage to CNF. Material and methods The study included 149 patients (268 eyes), average age 69.86 ± 4.4 years (81 men, 68 women). Three groups with NVR damage were identified: 1) Parkinson's disease (PD) (34 patients, 68 eyes); 2) post-COVID peripheral neuropathy (35 patients, 70 eyes); 3) mechanical damage after extracapsular cataract extraction (30 patients, 30 eyes). The control group consisted of 50 conditionally healthy volunteers (100 eyes). The "bead-likeness" of nerve fibers and other morphological parameters (KΔL, Ksym, length of nerve trunks, macrophages, microneuromas) were assessed using a CNN-based algorithm. The author's analysis algorithm was used to classify neurodegenerative changes. Results The analysis of confocal images revealed statistically significant differences between patients with neurodegenerative changes and healthy volunteers (p <0.05). The greatest changes were noted in patients with PD, including an increase in the number of nodular changes and additional inclusions that are not normally found. Analysis of the peripheral zone of the cornea confirmed the identified patterns: the density and length of nerve fibers in patients with neurodegeneration are reduced by an average of 75.6% relative to the central zone, while normally this figure is 65.7%. Patients with PD had stage 3-4 neurodegeneration, with post-COVID neuropathy - stage 2-3, after EEC - stage 2-4, which confirms the severity of degenerative processes. The developed diagnostic method demonstrated high sensitivity and specificity: PD - 97 and 94%; post-COVID neuropathy - 97.1 and 96%; EEC - 96.6 and 94%, respectively. These indicators confirm the high prognostic value of the method in various pathological conditions. Conclusion Hyperreflective inclusions in the peripheral zone of the cornea were detected for the first time in patients with stage 3 PD. The density of HRIs in the periphery of the cornea is reduced in neurodegeneration, which emphasizes the importance of studying changes in different zones of the cornea. The parameters of the "bead-likeness" of HRIs, reflecting the degree of morphological changes in neurodegeneration, were determined. Keywords: corneal confocal microscopy; neurodegenerative diseases; neural convolutional networks