Abstract
Purpose :
NIDEK GS-1 (NIDEK CO., LTD. Japan) captures 360° exams of the iridocorneal angle. The angle is split into 16 sectors each acquired on 17 focal planes. Incorrect alignment conditions, patients’ movements during exam acquisition, difficulty in keeping the eye open may degrade the quality of acquired images possibly preventing a correct pathology assessment. The aim of this work is to use Deep Neural Networks (DNNs) for assigning a quality index to GS-1 images.
Methods :
A total of 1835 images (RGB, 1280x960) taken from healthy and pathologic eyes were used. For each image a 224x224 ROI centered on the most relevant part of the angle (trabecular meshwork or irido-corneal interface for closed angles) was considered as input to the DNN. The dataset was manually split into two classes: “good” and “bad”. The former enclosed all images whose ROI was not saturated and rich of details, whereas the latter contained the remaining images (see examples in Fig 1). Since good and bad classes were unbalanced (1-to-10 ratio), 5 partially overlapping ROIs for good images and 1 ROI for bad images were considered, respectively. Transfer Learning was chosen in place of training a DNN from scratch. Hence, a pretrained network was chosen and partially retrained for the new quality-index task, with the advantage of reducing training time and computational requirements. For this study, the VGG16 and Xception DNNs trained on ImageNet were considered. The input dataset was divided into training, validation, and test sets having 70%, 20%, and 10% of images, respectively.
Results :
The Xception-based DNN gave the best results with ~95% accuracy and fast convergence (<15 epochs). The quality index has been assessed both on the test set and by visually comparing the best-quality image with the best-focus image in a focus sequence (see Fig 2).
Conclusions :
Transfer learning based on off-the-shelf DNNs can be successfully used for giving a quality index to GS-1 images. The proposed algorithm could reduce the time spent by clinicians for selecting the best images on which to perform the clinical assessment and could be also used for automatically retaking the examination if an insufficient quality is detected improving the effectiveness of the device.
This is a 2021 ARVO Annual Meeting abstract.