Purchase this article with an account.
Taylor Shagam, Minhaj Alam, Simon Bello, Laura Tracewell, Bryan Rogoff, Patricia Sha, Noli Graves, Dorothy Hitchmoth, Mary K Durbin, Niranchana Manivannan; Deep learning based binary image quality algorithm for low-cost fundus imaging system in remote care settings. Invest. Ophthalmol. Vis. Sci. 2021;62(8):100.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
An image quality (IQ) algorithm is important in assisting medical assistants with little to no experience in ophthalmology to judge whether an image is of sufficient quality for clinical diagnosis.In this study, we developed a binary IQ classification for providing real-time feedback in a remote care or primary care settings.
Data:5173 images acquired using VELARA™ 200 (Zeiss, Dublin, CA) camera from a retrospective study were graded by 2 graders and adjudicated by an optometrist). The images were acquired from normal and diseased subjects with various retinal pathologies were graded for quality of readable clinical information in a 1-5 scale (1-very poor and5-xcellent). The ground truth (GT) is determined by converting the gradings to binary: 0- Insufficient if IQ is <=2 (65.6% of 5173 images) and 1-Sufficient if IQ is >2 (34.4%). Small pupils, incorrect fixation, out of focus and other artifacts were the main causes of insufficient IQ.Algorithm:The dataset is split into three: i) training- 3646, ii) validation- 928 and iii) hold-out test – 599. The training set is augmented using flip and brightness adjustments. The deep learning architectures used to train the IQ classification are shown in figure 1. ImageNet weights were preloaded, sigmoid activation with binary cross entropy loss were used. The data was resampled and reweighted to account for the class imbalance. For hyperparameter tuning, all the networks were trained using Adam with cyclic learning rate scheduler and Stochastic Gradient Descent (SGD) with Nesterov momentum. SGD with Nesterov is chosen for the final model as it performed better. Sensitivity, specificity and the execution time were compared to select the final IQ model.
The performance of all networks in hold-out test set is shown in Figure 1. VGG-16 provided high sensitivity, specificity with lower execution time is selected as the final model. It achieved 99% sensitivity, 91% specificity and 220ms execution time in i5-10400H CPU. Figure 2 shows some of the examples results from IQ algorithm.
We developed an image quality algorithm with 99% sensitivity with an execution time to provide real-time feedback to the operator on whether to retake the fundus images.
This is a 2021 ARVO Annual Meeting abstract.
Figure 2. Sample results from proposed algorithm
Figure 1. a)Performance of various models b)ROC curve from final model
This PDF is available to Subscribers Only