Radiation Oncology, AI and Machine Learning in Research
Radiation oncology appears to be a provocative sandbox for incorporating artificially intelligent computer systems as a tool for cancer prognosis and treatment.
The field is characterized by its wide variety and veracity of data types and consequential decision-making process on patients’ outcomes.
Overall, Frost & Sullivan projects 40% growth for incorporating artificial intelligence (AI) into health care and the life sciences markets. While the AI marketplace for healthcare applications was nearly $634M in 2014, the growth trajectory is estimated at $6.6 billion by 2021. Some of the earliest applications are expected to be in diagnosing disease and potentially prognosis of challenging cases.
However, within the field of radiation oncology as well as in other medical fields, the road to broad-based adoption of AI and machine learning is experiencing a few potholes.
- Artificially intelligent computer systems are hungry for data. As a result, they must be continuously fed with the right kind of data (standardized/structured) so they continuously learn and update their algorithms.
- However, cancer diagnosis and treatment represents multiple types of often non-standardized data sets, collected across multiple disciplines. For example, computer systems still are not able to incorporate unstructured data such as the pathology data which forms the basis of initial diagnosis.
- In addition, incorporating machine learning tools to pull from larger and potentially multi-center patient protocols challenges patient healthcare information privacy boundaries.
- Also, as AI is incorporated into cancer treatment planning systems, the impact of automation on existing workforce skillsets must be evaluated.
One area of promise in radiation oncology involves the role of AI within treatment planning systems.
The Department of Radiation Oncology at the University of Michigan is the birthplace of one of the first, and continuously nimble, treatment planning systems world-wide. Recently, I spoke with Dr. Issam El Naqa, PhD., Associate Professor, Radiation Oncology – Physics and Dr. Charles (Chuck) Mayo, PhD., Associate Professor, Radiation Oncology, University of Michigan Cancer Center. I asked them about their own research on the role of AI and big data in cancer research, diagnosis and treatment.
Dr. Mayo felt that many aspects of diagnosis and treatment can be automated. However, there is lack of standardization in how that information is interpreted and entered into the computer system. For example, CPT (Current Procedural Terminology) and ICD (International Classification of Diseases) codes as well as laboratory codes are widely adopted standardizations routinely used patient records. In contrast, lack of standardization for many other key data elements (e.g., patient outcomes measures) and variability in input from the human elements involved in large-scale healthcare systems undermines the ability to feed data-hungry AI algorithms the information needed to perform consistently and with accuracy required for clinical decisions.
As a result, Dr. El Naqa mentioned, the majority of AI involved in radiation oncology is found within small-scale, experimental systems. As a result, data integrity is well-defined and controlled by medical physicists and researchers who also are data scientists like he and Dr. Mayo are. These models involve computer simulations where the computer can learn the best treatment plan and learns to predict radiation therapy outcomes from the clinician’s point of view. As a result, as researchers collaborate with physicians on interpretation of findings, the MDs become more comfortable using the AI algorithm as a supportive decision-making tool.
For example, data from dose volume histograms of tumors can be compiled and compared to historical, statistical norms. While radiation oncologists become used to assessing certain patterns, AI turbo-charges their diagnostic capabilities with additional information from a larger database. Consequently, deviations are assessed to identify patterns. Furthermore, associations from these data are identified which may be most important to a patient’s personalized treatment protocol. Then, outcomes and findings are fed back into the AI treatment planning system and the algorithm is retrained.
When assessing and expressing recurrence of cancer in patient records, variability in interpretation exists within radiation oncology and other areas of cancer treatment.
Dr. El Naqa offers that recurrence of cancer, post treatment, involves human interpretation of tests, including images. While a data algorithm can be trained to indicate either Yes or No for recurrence, assessment of the nature and severity of recurrence often is up to the staff. In addition, the level and nature of experience of a medical team can vary.
Within clinical settings, Dr. Mayo discussed the variability introduced into the AI model when physicians manually enter new data to refine original information and clinical insights on recurrence. However, consider the impact of new data on the original data algorithm. If data scientists are not alerted when new data is manually entered by team members and the data algorithm is subsequently retrained, the model may perform inconsistently.
Ideally, the entire medical team of medical physicists, dosimetrists and data scientists is highly collaborative in recalibrating, updating and recalibrating the AI algorithm. These data scientists agree that in addition to constructing practical AI tools to augment decision making, an important part of their job is working with these teams to shift clinical culture to be more data- and AI-centric.
AI in radiation oncology reaffirms diagnoses rather than replaces physician- and team-based human diagnoses.
However, the machine-to-human handshake where physicians partner with AI challenges traditional healthcare culture and institutional thinking.
Potentially, Drs. Mayo and El Naqa extrapolate, radiation oncologists, dosimetrists and medical physicists may be able to solve problems with built-in AI tools. As a result, they solve and design as many as 80% of repetitive, planned treatments in a more timely, cost-effective and reproducible manner. Increased efficiency due to the AI system allows the team to concentrate on more difficult patient treatments.
Still, more broad-based adoption of artificially intelligent treatment planning systems in radiation oncology challenges the traditional roles of dosimetrists, medical physicists and physicians. For example, medical physics dosimetrists will be asked to develop higher level cognitive decision-making skills. Consequently, they would devote the majority of their day to addressing more complex and time-worthy problems with treatment planning.
From the perspective of the physician, the potential role of AI and big data in cancer therapy initially is perceived as challenging who decides what treatment plan to give to a patient. However, human knowledge and experience will not be replaced by an algorithm in cancer treatment any time soon.
Instead, AI becomes a potent and trusted supportive tool and partner, enhancing each team member’s cognitive abilities in providing customized, effective patient treatment.