This resource is only available for ASCRS Members. Please log in using the teal "Login" button in the top right-hand corner to log in.
To conduct a pilot study and create a custom machine learning image recognition model solely based on pre-classified photokeratoscope images of normal eyes and eyes with keratoconus.
One hundred labeled photokeratoscope images of normal eyes and one hundred photokeratoscope images of clinically confirmed eyes with keratoconus were uploaded into Google Cloud Project Platform. AutoML Vision was enabled. The data was then labeled and categorized, with 80% of the images from each category utilized as a training set (to create initial parameters) and the remainder as a validation set (to tune the hyperparameters to improve overall recognition generalization). The metrics evaluated include the score threshold, sensitivity, specificity, precision, recall, and the area under the precision-recall curve.
The model used an additional 69 images as a prediction set. It was able to correctly identify 22 out of 25 photokeratoscope images of eyes with keratokonus, and 41 out of 44 photokeratoscope images of normal eyes. The overall accuracy of the model is 91.3%. The sensitivity and specificity are 88% and 93.2%, respectively. The positive and negative predictive values are respectively 88% and 93.2% as well.
We were able to create a custom machine learning image recognition model using the AutoML Vision software to identify photokeratoscope images of normal eyes and eyes with keratoconus with an overall accuracy of 91.3%.