Artificial Intelligence and Radiotherapy – Part One
Artificial Intelligence (AI) is one of the newest innovations reshaping the medical field to improve the quality and efficiency of patient care. Radiation therapy is one of the fields that will benefit greatly from this technology. Implementation of AI solutions has the potential to transform radiation oncology and the overall quality of radiation therapy for patients with cancer. These improvements are urgently needed, particularly in under-resourced healthcare environments, due to the increased incidence of cancer in recent years. Radiation oncology workforce shortages have led many vendors to include Deep-Learning (DL) tools to address this challenge. For example, Raystation and Varian Medical Systems are actively researching and adopting DL solutions.
Based on predictions, approximately 20% of clinical practices may be using DL-based tools clinically within the next few years. Thus, it is beneficial to understand how AI will be implemented, and the effects it will have on the future of radiotherapy. In this blog, the basic foundations of AI will be covered and applications of AI in radiotherapy will be reviewed.
Fundamentals of AI
Artificial intelligence (AI) involves the development and use of complex computer algorithms to perform tasks that normally require human intelligence, such as visual perception, pattern recognition, decision-making, and problem-solving, at a similar or improved level of performance.
Machine learning (ML) is a component of AI. ML is the study of computer algorithms that improve automatically through experience and by the use of data. Machine learning algorithms build a model based on sample data, known as “training data“, to make predictions or decisions without being explicitly programmed to do so. This process utilizes Artificial Neural Networks (ANNs). ANNs are created by Deep-Learning (DL) processes. Deep-Learning imitates the biological neural networks of the human brain to process data and create patterns for use in decision making. Each layer of the neural network builds on its previous layer with added data, leading to better analysis with experience or with newly added data.
For clarity, let’s examine AI in diagnostic radiology. A system with DL capability will be given data from thousands of assessed CT scans that inform the algorithm which patterns and shapes indicate cancer. Then the computer will be given CT scans, without the knowledge of which scans have abnormalities, and asked to determine which scans contain cancerous structures. A radiologist will then review the computer’s decisions, add data, and make corrections as needed until the computer has an exceptionally high success rate.
Applications of AI in Radiotherapy
AI could have particularly transformative applications in radiation oncology given the multi-faceted and highly technical nature of this field of medicine. The workflow in a radiation oncology center can be split into four conceptual domains: (1) CT simulation, (2) treatment planning, (3) quality assurance, and (4) treatment delivery. The process of diagnosing the patient with abnormal tissue growth occurs outside of the radiation oncology department; therefore, it is beyond the scope of this paper.
CT Simulation
A CT simulation for radiation therapy follows the initial consultation in the radiation therapy department but occurs prior to treatment planning. CT simulation includes a scan of the area of the body to be treated with radiation. The images acquired during the scan will be reconstructed and used to design the most precise treatment plan for the patient.
AI platforms might enable the personalization of radiotherapy by predicting the radiation sensitivity of the tumor prior to treatment planning. If a tumor is radiosensitive, the dose needed to eradicate the cancerous cells is decreased. In contrast, if a tumor is radioresistant it will require more dose to be eliminated. AI also has the potential to identify challenges that might be encountered based on the patient’s unique anatomy. Overall, AI detection of sensitivity and anatomical structures will optimize the treatment planning process.
CT scans require a greater dose for higher quality images; thus, there is a trade-off between dose reduction and image quality enhancement. The use of ANNs can lead to large improvements in image reconstruction so that similar quality images can be acquired with a lower dose. AI has been used to generate synthetic CT images from MRI images of the brain; therefore, fewer scans are required for treatment planning and the dose is reduced. Additionally, AI has the potential to reduce MRI scan times by enabling the reconstruction of fine details from under-sampled MRI data.
Treatment Planning
Treatment planning requires contouring of CT scans. In standard contouring workflows, the segmentation of tumor regions and other tissues is manually performed by the clinical staff on a slice-by-slice basis. As a result, the procedure is lengthy and subject to a high degree of variability, causing it to be one of the largest sources of uncertainty in treatment planning. The contouring of organs at risk (OARs) has shown to be successful through the use of ANNs. Furthermore, the challenging task of tumor segmentation by DL systems has shown considerable progress in recent years.
Once the scans are contoured, the optimal plan for each patient must be determined based on a set of parameters given by the radiation oncologist. Radiation treatment planning is considered a problem of optimization and requires multiple subjective trade-offs on the part of the oncologist and dosimetrist. As a result, conventional human-driven iterative workflows are often time-consuming with a high degree of variability. Auto-planning by AI systems mimics the process of iterative evaluations and adjustments made by experienced operators. Plans developed by AI processes are comparable to those manually generated while greatly reducing the planning time. The combination of contouring and treatment planning might enable full automation of the treatment-planning process in the near future.
Quality Assurance (QA)
After the radiation oncologist approves the treatment plan, the medical physicist performs plan checks and other QA checks to ensure that all the technical components involved in treatment delivery are functioning and set correctly to deliver the intended dose to the patient. QA involves repetitive and time-consuming measurements for each patient. AI tools have been shown to expedite this process and to detect rare errors in treatment delivery.
Treatment Delivery
AI has multiple applications for ensuring the correct delivery of treatment to patients. First proper positioning of the patient is necessary to ensure that the dose is not accidentally absorbed by healthy tissues. Currently, the integrated cone-beam CT (CBCT) device of the treatment machine is used to match the patient’s position to the position used for treatment planning. The challenge with CBCT is that it provides images of much lower quality than the planning CT images. Using similar processes described in the CT simulation section, AI has been applied to improve the image quality of CBCT to enable more accurate positioning of patients for treatment.
Complete immobilization of a patient is not achievable due to respiratory motion; therefore, the precision of irradiation is compromised. ANNs have shown to be effective in predicting tumor location from measurements of respiratory motion, thereby enabling adaptive beam realignment to occur in real-time.
Substantial changes in a patient’s anatomy between the planning appointment and delivery of treatment or throughout treatment can warrant re-planning. These changes often reflect tumor shrinkage or growth, or anatomical variations, such as a movement of the bowels, that could potentially result in altered doses to the tumor and organs. ML has been developed to identify candidate patients for re-planning intervention and the ideal time point at which it should occur.
Further Resources
Overall, AI has the potential to improve the accuracy, precision, efficiency, and overall quality of radiation therapy for patients with cancer. However, these improvements will result in significant changes for the radiation oncology workforce. Furthermore, multiple barriers to implementation will need to be overcome. Stay tuned for next month’s blog that will cover the necessary career shifts of Radiation Oncologists, Physicists, Dosimetrists, and Radiation Therapists due to the incorporation of AI. Additionally, there will be a discussion on the potential flaws and shortcomings of AI in this field. In the meantime, check out these links for more information.
https://www.nature.com/articles/s41571-020-0417-8
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6724618/