We know that COVID-19 does have an effect on the eye; however, the specific COVID-19 signature/pattern in the eye is not yet known to the medical profession. There is mounting evidence that Artificial Intelligence (AI) can help the clinician in all aspects of medicine, from formulating a diagnosis to making therapeutic decisions and predicting patient outcomes. Within AI, Deep Learning (DL) is a set of computational methods that allow an algorithm to program itself by learning from a large set of examples that demonstrate the desired behavior, removing the need to specify rules explicitly. In this project a DL algorithm was created to automatically detect Covid-19 in eye images. A specific type of artificial neural network (a large mathematical function with millions of parameters) optimized for image classification called a deep Convolutional Neural Network (CNN) was trained to classify eye images into one of two classes (also known as labels). The function computes Covid-19 positive class or Covid-19 negative class from the intensities of the pixels in an eye image.
Data preprocessing protocols, hyperparameter search, neural network training protocols, and results are described here in narrative form with quantitative information presented in the attachments. A specific CNN named the Inception Resnet is used in this research. Data preprocessing was performed to address the problems of class imbalance (less images of COVID-19 positive than negative), small dataset and biased learning (the CNN predicts everything as COVID-19 negative). Data augmentation was used to balance the datasets, with more augmentations applied to the class with less data. Geometric rotation, injection of Gaussian noise and focal loss were used as augmentation techniques, with focal loss providing additional control on the problem of bias. In AI, a hyperparameter is a parameter whose value is used to control the learning process. By contrast, the values of other parameters are derived via training. One of the most important part of neural network modeling is selection of quality hyperparameters, or more simply: the selection of parameters that the model is unable to change and optimize on its own, such as learning rate and momentum. The difficulty with making a quality selection of hyperparameters is the breadth of the search space, which is nearly infinite. This important part of the research was carried out using the Bayesian hyperparameter optimization algorithm to find the best set of hyperparameter values. Bayesian techniques differ from other, less efficient techniques like random and grid searches in that they use past evaluation results to choose the next values to evaluate. In practice, this allows the Bayesian technique to find better values of hyperparameters in less time. Another important contribution
of this research is the interpretation of the results of the CNN. In general, to develop trust in an AI technique with explaining the predictions of the model, it is important to understand the underlying mechanics of that technique, and any potential pitfalls associated with it. The Local Interpretable Model- agnostic Explanations (LIME) method was used to interpret the inner workings of the CNN when predicting the label of an image. The LIME method essentially produces a graphical overlay of a heatmap on the original image to highlight portions of the image that influence its decision in labeling that image as either COVID-19 positive or negative.
Results and Further Research
Convolutional Neural Network (CNN), an artificially intelligent system was trained to classify eye images into Covid-19 positive class or Covid-19 negative class from the intensities of the pixels in an eye image. It was found that the CNN had high accuracy, precision and sensitivity in learning to detect the pattern of COVID-19 in the human eye. These quantitative results are very encouraging, and they strongly suggest that further
research is warranted to determine the feasibility of applying this algorithm in the clinical setting and to determine whether use of the algorithm could lead to improved care and outcomes compared with current methods of COVID19 assessment.