The applications of artificial intelligence (AI) in healthcare are potentially numerous, clearly going beyond the field of medical imaging alone. AI is not a new scientific discipline, having its origins in the Dartmouth conference of 1956 [
1]. Although multiple layer perceptrons (MLPs), considered as the origin of “deep” convolutional neural networks (CNNs), were studied since the 70’s, major developments came in the 90’s [
2]and early 00’s [
3] concerning the learning rules establishing how weights in MLP can be updated. AlexNet [
4], winning in 2012 the ImageNet competition performing visual object recognition from photographs, introduced a major breakthrough in neural network performance bringing AI to the forefront of interest on computer vision and imaging applications. A more detailed historical overview can be found elsewhere [
5]. Growing numbers of patients, higher demands for quality like early detection and personalized therapies and an increasing workload for medical and nursing staff creates a demand for automation and the need for extracting more information from acquired data. Potential advantages of AI are already visible in screening routines in which a high number of patients (and associated data) are investigated for the presence or absence of disease, with results that are not worse than human performance. For example, McKinney et al. were able to show non-inferiority of their algorithm for screening of breast cancer as compared to experienced radiologists [
6]. At the same time, AI results are being criticized because of the lack of transparency and consequently a potential lack of reproducibility [
7].
The introduction of AI into the operation of radiology departments has led to optimizing resources [
8]. Such operational AI should prove even more relevant in nuclear medicine (NM), which deals with radioactive isotopes, whose shelf-life is limited. Patient scheduling, management of preparation of radiopharmaceuticals, report generation and recovering and organizing previous NM and imaging studies are examples of tasks where AI could contribute to streamlining the operation of a department.
We must however admit that AI still has little place within NM so far. No doubt, this may be related to the fact that smaller patient numbers pass through a NM department every day as compared to, e.g. a radiology department. However, this is underestimating the potential of AI as on the one hand, each patient image represents numerous—be it correlated—data [
9], and on the other hand, AI methodology has been shown to be able to adjust to smaller datasets by utilizing knowledge obtained in larger ones.
In this paper, we provide a short and concise review of the current state of the art in the field for both more physically and more clinically oriented components of AI applications in imaging. For more detailed reviews on each specific topic, readers are directed to other, more specialized articles [
10‐
17]. Finally, some ideas on the introduction and use of AI in NM are discussed.
The basics of AI: statistics
It is sobering to consider how AI is connected to classical, established statistical modelling. At the same time, it claims a right of its own since it mitigates a number of often overlooked but critical weaknesses in classical statistical procedures which are commonly used in the medical literature. Normal practice is to test the possible difference in the value of a biomarker (e.g. image intensity reflecting blood flow, oxygen extraction or metabolism) across two groups or conditions using a
t-test. This statistical procedure quantifies the (un-)likelihood of the experimental data arising from identical distributions. However, it does not provide insight to the differentiability of the two groups based on the value of the biomarker. If a biomarker is claimed to be an early marker of disease, it is less relevant at which statistical significance level the values differ between prodromal patients and a control group than it is as to how accurately the marker itself differentiates between groups. Hence, a low
p-value can lead to an artificially high confidence in a biomarker’s ability to actually detect disease. The statistical concept of
p-values in itself must be interpreted with care because it tells little about how replicable a result is. In fact, it can be shown that if a statistical test yields a
p-value of 0.05 (i.e. the standard level of significance), the probability of replicating this result is only 50% [
19]. The probability increases only to 80% for a
p-value ten times smaller,
p = 0.005. Together with procedural flaws, such as serially adding variables to control for in regression, referred to as “researcher degrees of freedom,” it leads to an inflation of false-positive results and the ongoing replication crisis [
20,
21].
In contrast, machine learning (ML) conceptually aims to optimize accuracy and reproducibility and in the case of AI even avoids human procedural bias in parameter selection. Interestingly, a ML approach does not necessarily venture beyond classical statistical modelling, it only takes a different approach to interpretation.
Linear regression is a classical statistical model but may also be seen as one of the simplest types of ML. Whereas in classical applications the interest is in determining the effect of certain covariates on a target, often reduced to a p-value as outlined above, the ambition of a ML application is to ensure the best possible fit to the observed data, and therefore predictive performance, while at the same time optimizing replicability.
AI diverts from classical statistics where it promotes more complex models in order to increase predictive accuracy. In particular, in case of deep learning (DL), classical statistical models such as logistic regressions are merged into hierarchies in such a way that the results of initial logistic regressions are used as input to others in so-called neural networks. This architecture implies that complex interactions between input variables are effectively utilized to increase model performance. It also implies that raw data, such as images, may naturally be input to the model, instead of summary features (such as for instance tumour volume, texture, etc.), which depend on human interaction.
Hence, AI, in comparison to classical statistical models, seeks to optimally identify feature differences between groups to optimize model performance while promoting generalizability beyond the study data initially available and has the ability to operate on raw healthcare data, such as images or health record information, without the need for manual feature extraction and summary, yielding a potentially more unbiased approach to knowledge discovery [
22].
Applications of AI in NM
For the sake of brevity and clarity, two major components in which AI can play a part can be discerned. The first component concerns image formation and image processing tasks and will be referred to as the “physics” component in the following sections. The second is largely application driven and hence will be referred to as the “clinical” component. It concerns routine workflow and final clinical endpoints, e.g. diagnosis, prognosis and prediction of response to therapy. Although these two components will be treated separately in this review, both are ultimately connected, the convolutional neural network (CNN) concepts being similar and very often associated within a single imaging paradigm. An example within this context may be the radiomics pipeline proposed by Hatt and Visvikis et al., where the development of AI-based image formation and segmentation algorithms serves to accurately determine regions of interest and extract imaging biomarkers which can subsequently be used in diagnosis and/or patient therapy stratification as well as follow-up [
1,
23].
In addition, the available data from multimodality devices such as PET/CT or PET/MR and from the emerging total body PET technology is expected to largely increase with the development of (multi)parametric imaging. Within this context deep learning (DL)-based reconstruction and analysis, algorithms are potentially more efficient to deal with the increasing volumes of acquired data.