Abstract
Purpose :
NIDEK GS-1 (NIDEK CO., LTD Japan) provides a 360° view of the iridocorneal angle by acquiring multiple images on different focus planes over the whole angle depth. The device also provides a feature-based algorithm (FB) to automatically detect the images with the best focused Target Region (TR), i.e. the trabecular meshwork or, in case of angle closure, the iridocorneal interface. In this work, we developed a deep learning based algorithm (DLB) to improve the detection accuracy of the images with the best focused TR (henceforth best focus shots).
Methods :
The algorithm includes two steps: first, the extraction of a ROI centered at the TR (not part of this work), then the processing of image pairs by a Siamese Convolutional Neural Network (SCNN) that predicts the probability that one shot has a more focused TR than the other. The SCNN is coupled with a concatenation and a fully-connected layer with a softmax activation function, and optimizes a categorical cross entropy loss. The SCNN is trained inside a Cross-Validation (CV) on 288 stacks of 17 images each, and a Bayesian optimization is performed over the CV folds to tune the SCNN hyperparameters. To test the predictive performance of our solution, both DLB and FB algorithms are run on 96 image stacks that were not considered for training, and the results are compared with the manually identified best focus shots. In detail, the performance is assessed by computing the distance between predicted and actual best focus shots.
Results :
DLB achieves better mean performance over the stacks with respect to FB, being both mean and median distance between predicted and actual best focus shots smaller (Tab. 1, Fig. 1). Moreover, DLB’s smaller standard deviation suggests a more stable performance over the stacks than FB (Tab. 1).
Conclusions :
Our algorithm has the potential to improve the detection of shots with the best focused TR; consequently, our solution increases the automation of the examination, making it more accessible to non-expert personnel, and providing an exam quality that is less operator dependent.
This is a 2021 ARVO Annual Meeting abstract.