Machine learning: applications in ophthalmology

Abstract

<p>Deep learning (DL) technology has revolutionised the screening, diagnosis, and management of eye diseases. Pattern recognition, through direct and indirect visualisation of the eye and adjacent structures via clinical examination and technological adjuncts, forms the cornerstone of ophthalmology. The imagecentric nature of ophthalmology makes it the perfect candidate to benefit from DL algorithms and as a test bed for clinical incorporation and advancement of intelligent algorithms within clinical medicine. Machine learning (ML) refers to pattern-recognition algorithms trained on datasets that are labelled (as in supervised learning), unlabelled (unsupervised learning), or contain a mixture of both (reinforcement learning).DLis a subset ofML; it mimics the hierarchical neural networks of the human cortex to learn the input with multiple levels of abstraction and generate predictions automatically. Compared to traditional techniques, DL has demonstrated superiority in image recognition, classification, and segmentation. Traditional techniques necessitate manual extraction of a set of features that are unique to each image within the training set in a step known as 'feature extraction'. This requires expert analysis and lengthy finetuning. In addition, the accuracy and applicability of the model would be undermined by variability in human anatomy and clinical presentation, as well as technological complications such as motion or reflection artefacts, and discrepancies in lighting conditions. On the other hand, DL utilises a stack of nonlinear hidden layers to complete an end-to-end learning process with an annotated dataset and the classification as the input and output, respectively. The automatic feature extraction, transformation, and decoding mitigate the problem of overfitting data points and unconscious biases granted that the DL models are well trained with various diverse datasets of adequate size.<br></p>