Abstract
Purpose :
Estimate the biomechanical properties of the corneal tissue is crucial for understanding ectatic diseases, and diagnosing glaucoma. It is essential for enhancing predictability in surgeries as corneal crosslinking, intrastromal ring segment implantation, and various forms of refractive surgery. However, current approaches to in vivo characterize the cornea's mechanical properties are indirect, influenced by the intraocular pressure (IOP) and they do not consider corneal anisotropy. We propose a deep learning-based methodology to estimate the patient-specific mechanical properties of the corneal tissue in vivo from the corneal deformation during a Non-Contact Tonometry (NCT).
Methods :
An extensive dataset of corneal deformations during NCT was generated through Monte Carlo simulations, following a Fluid Structure Interaction (FSI) model developed previously. 21 healthy patient-specific corneal geometries with thickness ranging from 484 to 588 μm were constructed based on Pentacam data. 300 FSI simulations were executed, varying the mechanical properties of corneal tissue within the range found in literature, coupled with intraocular pressure (IOP) values ranging from 8 to 30 mmHg. Subsequently, a non-intrusive reduced-order model for corneal deformation was developed. A predictive tool of the mechanical properties of the corneal tissue was constructed; it considers the corneal geometry, the intraocular pressure, and the corneal deformation during the air jet as input variables.
Results :
Artificial Neural Networks (ANNs) are employed to build a reduced order model (ROM) of the corneal deformation. Specifically, it approximates the corneal deformation during the air puff of a given patient with specific mechanical properties and IOP. ROM computes the corneal deformation in 1 ms, against the 48 hours needed by the FSI simulations. The model was validated against FSI data not used during the training phase. The deformation obtained with the ANN-based model led to a maximum of 4.23% error against the original FSI solution. Given these fast computational times, it is possible to compare the clinical Corvis ST data with the results of the ROM and estimate accurately the mechanical properties of the corneal tissue.
Conclusions :
The deep learning algorithm constructed could estimate the mechanical properties of the corneal tissue in vivo.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.