Artificial Intelligence: A Primer for Breast Imaging Radiologists

Abstract

Artificial intelligence (AI) is a branch of computer science dedicated to developing computer algorithms that emulate intelligent human behavior. Subfields of AI include machine learning and deep learning. Advances in AI technologies have led to techniques that could increase breast cancer detection, improve clinical efficiency in breast imaging practices, and guide decision-making regarding screening and prevention strategies. This article reviews key terminology and concepts, discusses common AI models and methods to validate and evaluate these models, describes emerging AI applications in breast imaging, and outlines challenges and future directions. Familiarity with AI terminology, concepts, methods, and applications is essential for breast imaging radiologists to critically evaluate these emerging technologies, recognize their strengths and limitations, and ultimately ensure optimal patient care.

Keywords: artificial intelligence; breast imaging; deep learning; machine learning; mammography.

Figure 1.

Hierarchy of artificial intelligence fields (28).

Figure 2.

Structure of a neural network. A neural network is composed of groups of nodes with consecutive layers—an input layer, one or more hidden layers, and an output later.

Figure 3.

Comparison of traditional ML and DL (3,36). Most published models in the breast imaging literature utilize supervised learning, in which the computer is provided with labeled data (eg, inputs are mammographic images that are labeled as “benign” or “malignant”). The training processes for traditional ML models and DL models differ in that the traditional ML model is based on human-engineered features, whereas the DL model learns the features that are necessary to classify the mammographic images as “benign” or “malignant” without human input. Once trained, the traditional ML model and the DL model could then classify a previously unseen mammographic image as “benign” or “malignant.” Abbreviations: DL, deep learning; ML, machine learning.

Figure 4.

Example of a confusion matrix. The reader can visualize if the model is “confusing” two classes (ie, if the model is mislabeling one class as another one). In this example, the true label of 0 is predicted with 100% accuracy, the true label of 1 is predicted with 75% accuracy, the true label of 2 is predicted with 87% accuracy, and the true label of 3 is predicted with 92% accuracy.

Figure 5.

Example of a receiver operating characteristic curve. The green line represents a perfect classifier (with an area under the curve [AUC] of 1.0), and the dotted red line represents a random classifier (with an AUC of 0.5). Models with AUCs above 0.5 (blue line) have at least some ability to discriminate between classes, with better models having AUCs closer to 1.0.

Figure 6.

Examples of heat maps or saliency maps. Mediolateral oblique views of the right breast demonstrate an invasive ductal cancer in the superior aspect of the breast at posterior depth (A, arrow), with overlying malignant heat map in red (B) and overlying benign heat map in green (C).