Technical improvements in radiotherapy delivery have been remarkable. The first time a patient was treated with a linear accelerator was in 1953.1 Initially, the radiotherapy beam was shaped with collimator jaws and lead blocks, which is labor-intensive and offers limited freedom to adjust the beam shape. The introduction of the multileaf collimator (MLC) was a giant step towards precision radiotherapy as it allowed to change the field shape during the treatment continuously. The MLC opened up the opportunity for intensity-modulated radiotherapy (IMRT) and later volumetric modulated arc therapy (VMAT). In the latter, the gantry rotates around the patient while the gantry speed, dose-rate, and aperture shape vary. These radiotherapy delivery improvements allow for the highly conformal dose distributions currently used.

Conformal dose distributions are only beneficial if the dose is accurately delivered to the target and simultaneously avoids normal tissues. To that end, one has to localize the patient's anatomy and the tumor before and during radiation therapy. Medical imaging is a powerful and non-invasive method to localize the patient's anatomy repetitively and can be achieved by a growing amount of imaging modalities. Before treatment, multimodal and 4-dimensional imaging modalities are used for target delineations. During treatment, patient setup relative to the treatment machine is monitored by, for instance, cone-beam CT (CBCT). After treatment, imaging again plays a vital role in the response assessment.

In radiology, it is still common practice to extract the complementary information of multiple medical images by careful visual inspection and “mental fusion”. Image registration allows for accurate image fusion (ie, displaying a combination of aligned images) and quantifying anatomical changes between scans. In radiation oncology, tasks such as target delineation, image guidance, and adaptive radiotherapy require image registration. The registration describes how an anatomical location in one scan relates to the same anatomical location on another scan. It is the link between knowing what to treat and implementing this treatment. Accurate registration is of utmost importance because inaccuracies can lead to a geographical miss of the target or are compensated for by an expansive radiotherapy margin. Therefore, improvements in image registration may lead to improved disease control and/or less toxicity.

Automatic image registration originates from the 1970s in the field of medical imaging.2 In the 1990s, the first rigid registration algorithms were developed in the radiation oncology field using chamfer matching for multimodal image alignment.3 Since then, many image registration algorithms have been developed and found their way into commercially available software and clinical applications.4 While registration methods increase in complexity and degrees of freedom, more powerful optimization algorithms are necessary to obtain an accurate and timely solution. To this end, artificial intelligence (AI) is a potent tool that opens up new registration methods and clinical opportunities.

This paper provides an overview of AI-based image registration methods and current and future applications in radiotherapy. First, we describe image registration principles and the main image registration applications in radiation oncology. Second, an overview is presented of AI strategies and applications. Finally, we describe current challenges and future perspectives.