Abstract
Purpose :
To detect the Keratoconus using machine learning and deep learning algorithms.
Methods :
The diagnosis of the Keratoconus is done via investigation of corneal topography image of the cornea. Our dataset consists of 1281 topography images labeled by corneal topography imaging device. In each of these images, 7936 points are selected on patient’s eye and saggital slope, tangential slope, and perpendicular distance from eyeball of both front and back of the corneal surface as well as the corneal thickness are measured. Topography images are classified into five groups: Abnormal or Treated, Keratoconus Compatible, Myopic Post-Operation, Keratoconus Suspected and Normal. We splited the dataset into train and validation sets with the fraction of 80% to 20%. To benchmark we trained some machine learning algorithms on the train dataset and measured the accuracy of classification in the validation dataset.
Results :
We measured the accuracy of classification in the validation dataset.(Table 1) Our current network model can accurately classify myopic post operated, normal and keratoconus compatible patients. However, it still struggles to classify abnormal or treated and keratoconus suspected patients. The development process for neural network is still ongoing. Our current network has 90% accuracy on the validation set and its’ confusion matrix on the validation set. (Table 2)
Conclusions :
We will try to optimize neural network to achieve better results especially in the subgroup of the keratokonus suspected and treated patients. Furthermore, we aim to develop an attention layer and obtain the location of the key features that plays crucial role on the classification, so that we can generate a decision and support making system for doctors to report easily.
This is a 2021 ARVO Annual Meeting abstract.