Artificial intelligence (AI) and predictive radiomics models that will be a boon to radiation oncologists in their daily practice were topics emphasized at ASTRO’s 61st annual meeting held Sept. 15-18 in Chicago. Mary Feng, MD, vice chair for clinical research and professor of radiation oncology at the University of California-San Francisco, presented an overview of practical applications of AI that will aid radiation oncologists in their workflow and clinical decision making.
Dr. Feng noted that this new field is attracting much research attention throughout the world. But after development, AI algorithms will require significant clinical validation in real-world medicine. Only then will they be feasible and appropriate to implement into healthcare practices.
AI-driven robots – “machine doctors” – will not displace radiation oncologists, she said. Rather, they offer conditional automation to help direct clinical decisions and require professionals to administer their use and make final decisions relating to patients.
An enormous amount of data is required for machine learning, and efforts are underway to make data easy to share. In 2017, for example, the American Association of Physicists in Medicine (AAPM) Task Group 263 published guidelines for standardizing nomenclatures in radiation oncology. But standard data entry must be quick and easy, and more work is needed, stressed Dr. Feng. “We must be proactive about consistent data entry,” she said, “so we can ‘harvest’ the data with confidence in the future.”
AI can be useful in many facets of radiation oncology workflow and in aiding such decision-making as: whether a potential patient should be treated, when metastases may occur and if radiation therapy treatment could extend remission, whether the patient can be cured with radiation therapy and, if that is not likely, whether survival and/or quality of life may improve. AI-driven decision support tools could be a great aid to physicians, patients and their families.
AI and predictive radiomics models also could facilitate treatment planning. Autosegmentation, for instance, would save time during simulation and contouring. Knowledge-based planning would help radiation oncologists and dosimetrists rapidly understand tradeoffs prior to planning and help them know what was achievable in the past. Specifically, they could identify when a tumor would be underdosed or organs at risk overdosed. They could also recommend treatment changes on the fly based on tumor response and/or organ dysfunction, harnessing the collective experiences of thousands of similar patients.
Additionally, AI could provide impartial “advice” to help determine when a treatment plan was “good enough,” and when it is appropriate to stop tweaking a plan with the hope of further improving it. Other knowledge-based planning tools could automate treatment planning for many patients, an “easy button” that might create better plans and do so faster than those created manually.
In quality control, AI models could predict such things as outliers, patient collision risk, modality down time, linear accelerator imaging issues, and intensity-modulated radiation therapy (IMRT) quality assurance failures. AI could automate the routine, enabling the professional staff and clinical resources of a radiation oncology department to focus its “saved time” on high-risk patients and patients who may require a unique treatment.
But AI-drive functions and radiomics models can only go so far. “Robotic AI and radiomics models do not have intuition,” said Dr. Feng. “They can’t beat my clinical judgement and experience.”
On a scale of 0 (no automation) to 5 (full automation), Dr. Feng predicts that AI and radiomics will comfortably fit midway, in a state of conditional automation where automated systems will drive and monitor aspects of radiation therapy planning and treatment, but always rely on a human driver – the radiation oncologist – for backup.
“In patient care, AI can help us every step of the way, from consultation through follow-up. But we must continue to be skeptical and perform sanity checks,” she urged. “Today, we must partner with data scientists to build interpretable models that are practical to use and be proactive in testing these models.”
Models predicting acute toxicity in head and neck cancer patients
Most patients with head and neck cancers receive radiation therapy treatment, and nearly all experience one or more acute toxicities, especially dysphagia. Difficulty in swallowing often leads to weight loss and use of a feeding tube. Patients may also require rehydration, nutritional support, and/or hospitalization for pain management. Managing these and related issues requires decisions on timing and methods of intervention, and AI can help. “The application of big data and predictive analytics will help tailor specific customized treatments to patients and offer expected outcomes and toxicities,” said Jay Paul Reddy, MD, PhD, assistant professor of radiation oncology at MD Anderson Cancer Center in Houston, TX, during an ASTRO 2019 scientific session.
Researchers at MD Anderson are working to develop predictive models of acute toxicity during radiation for head and neck cancer patients to identify patients most likely to experience significant weight loss during radiation therapy, who may require feeding tube placement, and who may experience unplanned hospitalization within 90 days of treatment. Dr. Reddy described research testing of three models using data from 2121 consecutive courses of radiation treatment between May 2016 and August 2018.
The patient cohort of the MD Anderson study was predominantly male (75%), and ranged in age from 55 to 70 years. Dr. Reddy said that radiation therapy treatment sites represented a typical mix of head and neck cancers, primarily the oropharynx (35%), followed by the oral cavity (15%), skin (11%), larynx (8%), salivary gland (6%), and thyroid (5%). The patients were prescribed a radiation dose of 30 Gy to 69.3 Gy in 9 to 33 fractions, with a median of 60 Gy administered in a median 30 fractions.
The researchers’ approach used structured data for multiple variables including clinical and pathologic characteristics, outcomes, and acute toxicities recorded in the hospital’s oncology information system, electronic health records, and tumor registry. They extracted more than 700 clinical and treatment variables for each patient. The training set for the three models included the first 1896 radiation therapy courses; the remaining 225 radiation therapy courses were used as the validation set.
A model was considered clinically valid if it had an AUC performance of more than 0.7. The gradient-boosted decision tree models achieved clinical validity with respect to predicting significant weight loss and the need for feeding tube placement in the validation patient cohort. None of the models were able to predict unplanned hospitalizations at the 0.7 AUC performance level.
“This study demonstrates the feasibility of employing precision oncology to predict acute radiation toxicities,” said Dr. Reddy. “This type of model may help us determine if we need to be proactive in placing a feeding tube in a patient up front, thus minimizing the potential for more severe acute toxicity to occur. Radiomics tools will help us in many aspects of decision-making.”
Adaptive radiation therapy prediction model for high-risk NPC patients
Radiation therapy, most generally IMRT, is the standard treatment for nasopharyngeal carcinoma (NPC), a rare type of head and neck cancer in the upper part of the throat behind the nose. Its location adjacent to the brainstem, parotid glands, and spinal cord make surgery impractical.
Even with the most precise pre-treatment IMRT planning, weight loss may cause patients to experience significant radiation dose deviations to the target tumor and surrounding health organs during treatment. To address this problem, adaptive radiation therapy (ART) offers clinical and dosimetric benefits for NPC patients by compensating for the dosimetric impacts caused by anatomic and geometric variations in patients as they undergo radiation therapy. However, ART planning is time-consuming and resource intensive, requiring repeat imaging, re-contouring, re-planning, and analysis of dosimetric impacts between new and previous treatment plans. Thus, ART is not feasible in a typical radiation oncology department to use with every NPC patient.
An algorithmic tool that could help radiation oncologists identify high-risk NPC patients who would benefit the most from optimal personalized ART treatment could help streamline resource management in a radiation oncology department and increase its use with select patients. Researchers at Hong Kong Polytechnic University’s Department of Health Technology and Informatics are using radiomics to develop such a model. They describe radiomics features extracted from a 3-dimensional (3D) tumor volume containing predictive biomarkers for tumor shrinkage following cancer treatment using multiparametric magnetic resonance imaging (MRI), which could help identify NPC patients eligible for ART.1
Their study included 70 patients who received primary radiation therapy with or without chemotherapy at Queen Elizabeth Hospital in Hong Kong between April 2015 and February 2016. They were divided into training (51 patients) and validation (19 patients) cohorts.
During treatment, patients were weighed weekly, and imaging was performed daily to correct for positional variations and assess anatomic or geometric changes. Lead author Ting-ting Yu and colleagues reported that when a patient had lost >10% of their pre-treatment body weight, the original radiation therapy treatment plan was recalculated to determine whether a new, modified plan was needed. This occurred for 39 patients, of whom 16 had a new plan created, most often during week 4-5 of treatment. Plans were also performed if a change in body or neck contour, neck tissue loss, and/or lymph node shrinkage were seen. Most patients had stage III (61%) and IV (34%) NPC cancer.
All patients also underwent MRI imaging with authors extracting 479 radiomic features from the images, including shape, texture, and first-order intensity features. The features were applied to three models, of which a joint T1-T2 model with six selected radiomic features had the best predictive capabilities. Research is continuing.
AI Growth: From big data farming to harvesting results in radiation oncology. Appl Rad Oncol.