Annals Academy of Medicine, January 2019

The real-world application of artificial intelligence (AI), machine learning (ML), and deep learning (DL), have generated significant interest throughout the computer science and medical communities in recent years. This interest has been accompanied by no small amount of hype. Though the term ‘ML’ was coined 50 years ago by Arthur Samuel, who stated that machines should have the ability to learn without being programmed,1 the advent of the graphics processing unit (GPU) has enabled much-improved processing power and enabled new possibilities with AI. DL—an approach that utilizes multiple neural networks to learn a representation of data using multiple levels of abstraction2—has revolutionized the computer vision field, and achieved substantial jumps in diagnostic performance for image recognition, speech recognition, and natural language processing.2 In the technical world, DL has been heavily used in autonomous vehicles,3 gaming4,5, and numerous smartphone applications. The availability of different software (e.g. Caffe, Tensorflow), and the off-the-shelf convolutional neural networks (e.g. AlexNet, VGGNet, ResNet, and GoogleNet) have removed barriers to entry for many academics and clinicians, resulting in the recent surge of interest within the medical settings. To date, this technique has shown promising diagnostic performance, across specialties including ophthalmology (e.g. detection of diabetic retinopathy [DR], glaucoma, and age-related macular degeneration from fundus photographs and optical coherence tomographs),6-11 radiology (e.g. detection of tuberculosis from chest X-rays [CXRs], intracranial hemorrhage from computed tomography of the brain),12-15 and dermatology (e.g. detection of malignant melanoma from skin photographs).16