The growing availability of digital data and computational power enables the training of deep neural networks useful for real-world applications.
Deep neural networks are typically composed of multiple convolutional layers used to efficiently extract information from high-dimensional inputs, pooling layers to reduce dimensions, and fully connected layers to aggregate neuron activations into output values.
Neural networks learn to approximate a function by forward propagating the input layer by layer, calculating a loss comparing the current output to the ground truth, and then updating the network’s weights and biases through backpropagation.
Deep architectures are selected based on the nature of the input and desired output data to perform tasks such as classification, detection, semantic segmentation, and temporal recognition.
Deep-learning strategies to guarantee the security of sensitive medical data, train networks using less supervision, increase model explainability, and deliver real-time predictions in the operating room are being developed to generate value in surgery.
Instead of trying to produce a programme to simulate the adult mind, why not rather try to produce one which simulates the child’s? If this were then subjected to an appropriate course of education one would obtain the adult brain.
Alan Turing, 1950, “Computing Machinery and Intelligence”
The question of whether machines could be capable of thinking and learning like humans has always allured mankind. As briefly discussed in Chapter 1, ancient mythology dating back more than 2500 years ago already made references to modern concepts such as self-moving objects, robots, and what we now refer to as artificial intelligence (AI). The recent surge in deep learning is contributing to turning these antic fantasies into today’s reality. Deep-learning breakthroughs in fields ranging from image and speech recognition to game playing have been widely covered by the press, generating enthusiasm in the general public as well as among businesses and funding agencies. Replicating biologic neurons on silica chips is by no means a novel idea; however, the incredible amount of digital data we now ubiquitously generate—together with the growing availability and decreasing cost of computational power—makes training deep neural networks practical. Furthermore, open source programming frameworks have lowered the barrier to entry, allowing quick prototyping for real-world applications.
Deep-learning models are already in use for applications such as automatic recommendations of media content and smart assistants in our homes and self-driving cars. Health care is a very active field of research for AI given its social and economic relevance. In this sector, surgery is a particularly interesting and challenging subfield because it is a high-stakes discipline where multiple people interact, make quick decisions based on a large amount of sparse information, and act to alter a patient’s anatomy. Despite clear opportunities and widespread hype, deep learning has yet to impact surgical patients. Thus, it is the ideal moment for surgeons to understand the intuitions behind neural networks, become familiar with deep-learning concepts and tasks, grasp ...