Abstract
Purpose :
Six ordinary differential equation (ODE) tear film breakup (TBU) models are created to capture evaporation, osmosis, and different types of flow. A convolutional neural network designed after Su et al. (IEEE, 2018) and trained on over 40,000 image patches successfully identifies TBU and non-TBU in fluorescent (FL) images gathered in vivo with an accuracy of 92%. The ODE models are fit to the automatically identified TBU FL intensity data by optimizing TBU quantities such as evaporation and tangential flow rates. Best-fit determination of TBU parameters suggests which mechanisms cause the thinning.
Methods :
The FL intensity data was originally recorded from 25 normal subjects with 20 trials taken over two visits (Awisi-Gyau, Indiana University thesis, 2020). We extract the experimental FL image data from the centers of TBU regions identified by a convolutional neural network. These are fit with the models designed to mimic evaporation-driven, tangential flow-driven, and combination thinning. We estimate parameters using a least squares minimization of the difference between experimental and computed intensities using the trust region-reflective method. Theoretical intensity dependent on tear film (TF) thickness and FL concentration was based on Nichols et al (2012, IOVS, 53:5426). Separate procedures were used to estimate initial FL concentration and localized TF thickness (Wu et al, IOVS 2015, 56:4211). The fits use up to four parameters: evaporation rate v, a (steady) and b1 (decaying) extensional flow rates, and decay rate b2. All computations are via custom Python programs.
Results :
Optimal evaporation rates fall near or within experimental ranges (Nichols et al, IOVS, 2005). An example fit is shown in the Figure. The best-fit model determines that the evaporation rate is -3.25 μm/min and that the instance exhibits strong, outward tangential flow that decays, allowing evaporation to take over in importance.
Conclusions :
Intensity decay in automatically identified TBU areas can readily be fit with simplified models that capture essential thinning dynamics and yield physically relevant quantities. This procedure can be applied to a wide range of instances to obtain statistical information that cannot be directly measured during breakup.
This is a 2021 ARVO Annual Meeting abstract.