Deep Learning in Medicine—Promise, Progress, and Challenges

Abstract

Recent years have seen a surge of interest in machine learning and artificial intelligence techniques in health care. 1 Deep learning 2 represents the latest iteration in a progression of artificial intelligence technologies that have allowed machines to mimic human intelligence in increasingly sophisticated and independent ways. 3 Early medical artificial intelligence systems relied heavily on experts to train computers by encoding clinical knowledge as logic rules for specific clinical scenarios. More advanced machine learning systems train themselves to learn these rules by identifying and weighing relevant features from the data, such as pixels from medical images, or raw information from electronic health records (EHRs).For machine learning approaches to work well in practice, feature engineering directed by experts is often required (eg, quantization of laboratory tests into discrete value ranges or the extraction of descriptors that characterize the texture of medical images). The newer deep learning techniques, which require less supervision, use an end-to-end learning mechanism to map raw inputs (such as laboratory test values or image pixels) to outputs without human-directed manipulation of the data. The resulting maps are composed of multiple layers of interconnected nonlinear processing units-the "neurons" of deep learning. Deep learning techniques may still require some expert involvement to design optimal model architectures and identify the best set of parameters.At present, the most successful applications of deep learning in medicine have been for analyzing medical images. Deep learning algorithms are capable of performing automatic and accurate detection of diabetic retinopathy 4 and skin cancer 5 from retinal fundus and skin images, respectively. The potential of deep learning to disentangle complex, subtle discriminative patterns in images suggests that these techniques may be useful in other areas of medicine. Substantial challenges must be addressed, however, before deep learning can be applied more broadly.

Machine learning applications in the diagnosis of leukemia: Current trends and future directions

“…For instance, Topol et al reported that out of the five studies using deep learning in pathology (with results compared to physicians), one study only was prospective in nature utilizing AHI in breast cancer metastases . Thus, there is a need to develop both data quality and quantity to potentiate the powers of ML tools . This study supports that the current quality of studies pertaining ML is suboptimal, and there is an imperative need to improve both quality and quantity of data before the prospective application of these models in medical practice.…”

The shortage of skilled workers in Europe: its impact on geriatric medicine

“…Moreover, inter-professional teaching would greatly benefit from the rapid development of education technology. Initiatives such as blended teaching [15], serious games [16], virtual patients [17], virtual team play [18] and very soon, machine learning [19] will totally transform the education of health professionals. Such fundamental changes are urgently needed to attract more young and promising learners of both genders into the field of health.…”

Neurons learn by predicting future activity

“…On the other hand, the increase in neuronal population activity also increases local blood flow leading to more glucose and oxygen entering a neuron (see review on neurovascular coupling: 21 ). This activity dependent energy supply can be expressed as: + 2 (∑ ) 2 , where xk represents spiking activity of neuron k from a local population of K neurons ( ∈ {1, … , , . .…”

“…However, networks that use backpropagation have multiple properties which are difficult to reconcile with biological networks. For example: (1) they require separate phases for sending a bottom-up 'sensory' signal and for receiving a top-down error signal, whereas in the brain top-down feedback is combined with incoming sensory information; (2) backpropagation requires symmetric weights, meaning that a synaptic connection from neuron A to B has to have exactly the same strength as connection from B to A; (3) neuron models are biologically unrealistic as they do not include, for example: spiking activity, internal neuron dynamics, or Dale's law (either excitatory or inhibitory neurons), among many other simplifications. This question of how to bridge the gaps between biological and artificial learning algorithms is a subject of rapidly growing research at the intersection of neuroscience and computer science [5][6][7] .…”