Abstract
Purpose :
Purpose: An intraoperative Talbot-Moire interferometer uses Fourier analysis of captured images to generate exam data. The quality of the processed data can be degraded by the presence of artifacts in the images. A two-stage software module combined a convolutional neural network (CNN) front end with a deep neural network (DNN) back end capable of detecting in real time the artifacts appearing singly or in combination.
Methods :
Methods: The first stage of the detection module employed pre-trained weights for a VGG16 CNN model, and the second stage was a fully connected DNN. The CNN weights were fully specified. Only the DNN and weights needed to be determined. The CNN inputs the image to be analyzed and outputs a length-512 real-valued vector. This vector is a collection of image features suitable for classification tasks for a large range of image types. The DNN inputs the output vector from the CNN and produces a vector estimating the probability of the presence of each of the artifacts. The data set consisted of 287 hand-labeled interferometer images. Each image could contain one or more artifacts. These image artifact labels included: one or more bubbles present, one or more floaters present, and corneal glint present. The data set was randomly split into 229 images for training the model and 58 images for testing the model.
Results :
Results: 99% of the training set images were classified correctly for the presence of glint and 97% were classified correctly for the presence of floaters and bubbles. 91% of the test set images were classified correctly for the presence of glint, 95% were classified correctly for floaters, and 97% were classified correctly for bubbles. 97% of training images were correctly identified as being free from artifacts or having one or more artifacts, compared to 91% of the test images being correctly identified as being free from artifacts or not.
Conclusions :
Conclusion: The CNN/DNN module developed was reasonably successful at identifying the artifacts in the image test set. We believe further training with a much larger image data set will increase performance. In addition, optimizing the threshold for the output layer values would allow customization of likelihood of detecting a “bad” image at the expense of increasing the likelihood of rejecting an artifact-free image.
This is a 2021 ARVO Annual Meeting abstract.