Abstract
Purpose :
Almost every patient encounter will involve performing an eye examination to give clues about the patient’s diagnosis. Neural network models have proven adept at processing image data, and it is quite cheap, fast, and simple to take a high resolution image of a patient’s eye. Processing these images to yield certain parameters or findings from the exam would increase the efficiency of collecting this data in clinical practice.
Methods :
A dataset of images of partial faces with a clear view of one eye was collected and processed into a format suitable for training neural networks on. The “PyTorch” python library was then used to further prepare the data into tensors and then to construct a neural network model and train it. This model was then tested on another set of images to asses its performance. The input provided to the model was a scaled-down gray-scale version of the original image. The target output the model was trained to estimate was the ratio of the pupillary diameter to the iris diameter.
Results :
The model created, with two convolutional and max-pooling layers followed by three linear fully-connected layers was able to achieve a mean deviation of 0.01+-0.01 on measuring the pupillary-to-iris-diameter-ratio. RMSE was 0.014. Stated as a percentage deviation, it was within 5.8% +- 11.3% of the actual target value.
Conclusions :
A neural network model can be used to determine certain parameters that are part of a standard eye examination with reasonable accuracy. It has been demonstrated in the case presented of pupillary dilation, yielding a numerical value that is accurate enough to represent the current state (miosis,mydriasis). This validates the approach to some extent and encourages further development to encompass other exam parameters.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.