Neural networks in medical imaging

Since its conception, based on principles of connectionism, neural networks have progressed from simple electrical circuits to the current deep learning-based models, which have shown great performance in analyzing medical imaging. Neural networks are structured as a series of computational algorithms which endeavor to identify underlying relationships in a set of data. These algorithms include activation functions, and weights which transform the data from input to output. The self-learning capability of these networks rely on the loss, backpropagation, and gradient descent algorithms. Convolutional Neural Networks (CNNs) are a type of neural network that has been used extensively in computer vision and performs well in extracting features from an image. Training a neural network can be manually supervised, unsupervised, or a combination of both approaches and can be tailored based on the availability and features of the training data. The training process commonly presents with challenges of overfitting, vulnerability to adversarial examples, and the black box effect. Multiple applications of CNNs exist in the literature, for tasks such as classification, detection, segmentation, image registration, and image processing. The use of neural networks for these tasks holds the potential to reduce image acquisition times, improve overall image quality, and minimize error and variability. An exciting evolving development is the use of CNNs for software control of the radiological machines, such as a CT or magnetic resonance imaging (MRI) scanner, to ensure consistent, error-free high quality standardized results. This application may be used to address the chronic shortage of highly skilled operators in developed and developing nations.