Artificial intelligence (AI) — the ability of a machine to perform cognitive tasks to achieve a particular goal based on provided data — is revolutionizing and reshaping our health-care systems. The current availability of ever-increasing computational power, highly developed pattern recognition algorithms and advanced image processing software working at very high speeds has led to the emergence of computer-based systems that are trained to perform complex tasks in bioinformatics, medical imaging and medical robotics. Accessibility to ‘big data’ enables the ‘cognitive’ computer to scan billions of bits of unstructured information, extract the relevant information and recognize complex patterns with increasing confidence. Computer-based decision-support systems based on machine learning (ML) have the potential to revolutionize medicine by performing complex tasks that are currently assigned to specialists to improve diagnostic accuracy, increase efficiency of throughputs, improve clinical workflow, decrease human resource costs and improve treatment choices. These characteristics could be especially helpful in the management of prostate cancer, with growing applications in diagnostic imaging, surgical interventions, skills training and assessment, digital pathology and genomics. Medicine must adapt to this changing world, and urologists, oncologists, radiologists and pathologists, as high-volume users of imaging and pathology, need to understand this burgeoning science and acknowledge that the development of highly accurate AI-based decision-support applications of ML will require collaboration between data scientists, computer researchers and engineers.
Applications of machine learning (ML) to prostate cancer care are rapidly growing owing to the many technological platforms involved in its diagnosis, prognosis and treatment.
In diagnostic imaging, ML is applied to perform low-level image analysis tasks such as prostate segmentation and fusion of different modalities (for example MRI, CT and ultrasonography) and high-level inference and prediction tasks such as prostate cancer detection and characterization.
ML algorithms are able to enhance prostate cancer treatment by augmenting the surgeon’s display with information such as cancer localization during robotic procedures and other image-guided interventions and could be used towards autonomous manipulation of tools for assistance in the operating room.
Computer-assisted diagnosis of prostate cancer in histopathological slides could be achieved by ML in order to optimize accuracy, reproducibility and throughput and to further enhance health-care delivery by enabling the use of customized precision-care pathways.
ML methods are used to identify genes or groups of genes for which expression specificity to predict outcomes of prostate cancer is high and could be used for screening, developing diagnostic tools, determining optimal individualized treatment and producing targeted drug regimens.
Collaboration between urologists, data scientists, computer researchers and engineers is required to ensure that artificial intelligence (AI)-based decision-support applications are properly trained, operated and regulated.
Access optionsAccess options
Subscribe to Journal
Get full journal access for 1 year
only $17.75 per issue
All prices are NET prices.
VAT will be added later in the checkout.
Rent or Buy article
Get time limited or full article access on ReadCube.
All prices are NET prices.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This work was funded by a Prostate Cancer Canada grant (D2016-1352), by the Canadian Institutes of Health Research (MOP-142439) and by the C. A. Laszlo Chair of S.E.S. G.N. is a recipient of a Prostate Cancer Canada Postdoctoral Research Fellowship Award (PDF2016-1338).
Nature Reviews Urology thanks O. Vermesh and Q. Wang for their contribution to the peer review of this work.
A classifier in machine learning refers to the placing of a new observation into the appropriate category among those categories that were based on trained data sets of known observations.
- Ground truth
Labels or annotations that were determined by an expert, considered to be the correct targets and used in training, testing and evaluating machine learning algorithms.
- Naive Bayes classification
A supervised learning method that is based on a probabilistic approach and relies on Bayes’ theorem. The learning process involves parameter estimation of the probability distributions based on the data. For a new sample, the class with the maximum probability would be predicted on the basis of the probability of its features.
- Support vector machines
(SVMs). Supervised learning methods for classification that learn the optimal ‘separation’ between the features of each class. The predicted class of a new sample would be based on the ‘region’ in feature space that the sample features occupy.
- Random forests
A supervised learning method for classification that is based on decision trees. It consists of multiple trees, each with a random subset of the features, and tries to optimize the split values of the branches.
- k-Means clustering
An unsupervised learning method for clustering. The algorithm iteratively assigns each data sample into one of k classes on the basis of the ‘distances’ between features.
- Principal component analysis
An unsupervised learning method for dimensionality reduction.
Types of neural networks that are trained to encode an input into a lower dimensionality such that the reverse decoder can reconstruct the encoded sample as similar as possible to the original input. These models can be used for unsupervised learning of the most descriptive features of the data, for example, for dimensionality reduction.
- Artificial neural networks
A collection of units that are connected to each other (typically) as layers and inspired by biological neural networks in the brain. Each unit, also referred to as an ‘artificial neuron’, has an output that is a function of the weighted sum of its multiple inputs and is ‘activated’ if that sum is higher than a threshold (bias). Given the data as the input to the first layer, the weights and biases of the neurons are optimized to match the output of the last layer with some target (for example, minimizing a classification error).
- Convolutional neural network
(CNN). A type of artificial neural network in which a neuron in a layer is connected to a few adjacent neurons from the previous layer, and the next neuron in that layer is connected to the next adjacent neurons from the previous layer and so on. Such a network architecture is commonly used in learning vision tasks such as image classification.
- ‘Leave-patient-out’ validation
An approach, also known as a k-fold cross-validation, to evaluate the performance of a classifier by training it on all data except the samples of one or more patient or patients and then testing it on the left outpatient or outpatients. The process can be repeated over all patients, each time with a different (subset of) patient or patients left out, and averaging the results.
- Statistical shape modelling
A representation of a set of shapes by modelling their geometry with a typically small number of parameters that control their main modes of variation and are derived using statistical methods.
- Sørensen–Dice similarity coefficient
A value that measures the similarity between two sets, with a value of zero when the two sets are unique and a value of one when they completely overlap. If the sets are pixels within two shapes, the Sørensen–Dice similarity coefficient can measure the amount of overlap of the shapes.
- Recurrent neural network
(RNN). A type of artificial neural network in which neurons are connected to other neurons at previous time steps and can, therefore, learn temporal patterns in sequential data.
- Dynamic contrast-enhanced
(DCE). A modality of MRI that measures parameters of tissue perfusion (ktrans) in the presence of a contrast agent.
- Diffusion tensor imaging
(DTI). An MRI technique that maps the diffusion of water molecules in the tissue.
- Generative adversarial networks
(GANs). Types of artificial neural networks in which there are two paths of layers: one that generates samples from a random input and another that tests the similarity of the generated samples to real samples. The generative network is, therefore, trained to generate samples that mimic the real samples. Such networks are used to generate synthetic images that are visually similar to real images of their class.
When a machine learning model is trained to perform well on a limited data set and performs worse when it is applied to new data, it is said to be overfitted to those data. In order to avoid overfitting, a data set that is large enough to represent the real-world diversity should be used for training and the training should be stopped before fully converging.