key: cord-0962618-71tbgm52
authors: Kang, John; Thompson, Reid F.; Aneja, Sanjay; Lehman, Constance; Trister, Andrew; Zou, James; Obcemea, Ceferino; El Naqa, Issam
title: NCI Workshop on Artificial Intelligence in Radiation Oncology: Training the Next Generation
date: 2020-06-13
journal: Pract Radiat Oncol
DOI: 10.1016/j.prro.2020.06.001
sha: 52c9c987122a1a9a60f2cb49eb2990d14981b45b
doc_id: 962618
cord_uid: 71tbgm52

Abstract Artificial intelligence (AI) is about to touch every aspect of radiotherapy from consultation, treatment planning, quality assurance, therapy delivery, to outcomes modeling. There is an urgent need to train radiation oncologists and medical physicists in data science to help shepherd AI solutions into clinical practice. Poorly trained personnel may do more harm than good when attempting to apply rapidly developing and complex technologies. As the amount of AI research expands in our field, the radiation oncology community needs to discuss how to educate future generations in this area. The National Cancer Institute (NCI) Workshop on AI in Radiation Oncology (Shady Grove, MD, April 4-5, 2019) was the first (https://dctd.cancer.gov/NewsEvents/20190523_ai_in_radiation_oncology.htm) of two data science workshops in radiation oncology hosted by the NCI in 2019. During this workshop, the Training and Education Working Group was formed by volunteers among the invited attendees. Its members represent radiation oncology, medical physics, radiology, computer science, industry, and the NCI. In this perspective article written by members of the Training and Education Working Group, we provide and discuss Action Points relevant for future trainees interested in radiation oncology AI: (1) creating AI awareness and responsible conduct; (2) implementing a practical didactic curriculum; (3) creating a publicly available database of training resources; and (4) accelerate learning and funding opportunities. Together, these Action Points can facilitate the translation of AI into clinical practice.

Artificial intelligence (AI) is a longstanding field of study that has attempted to emulate and augment human intelligence. In the last several years, AI has been reinvigorated by advances in computer technology and machine learning (ML) algorithms, which aim to teach computers to learn patterns and rules by using previous examples. ML builds on experiences from computer science, statistics, neuroscience, and control theory, 2 among many other disciplines. ML has benefited from recent availability of large datasets and developments in computers' hardware and software for solving large-scale optimization problems. Most notably, deep learning (DL) techniques have demonstrated significant successes in computer vision and language processing.

These advances are most visible in consumer quality-of-life improvements such as self-driving cars and voiceactivated virtual assistants. The umbrella term "informatics" includes practical applications of any of the above areas of study; for example, bioinformatics for biology and clinical informatics (or biomedical informatics) for clinical practice. The term "data science" refers to the general study of data analysis, which has recently focused on ML methods. A schematic of the relationships between common terminologies is shown in Figure   1 .

Many fields such as finance, manufacturing, and advertising have already incorporated AI into their workflows to improve efficiency and perform supra-human tasks. While AI has been adopted more slowly in the clinic due to multiple competing factors-including a lack of training-the perception and engagement of AI in medicine has been improving. The American Medical Association (AMA) adopted a policy in June 2019 to integrate training in AI augmentation 1 . The National Institutes of Health (NIH) Big Data to Knowledge (BD2K) Initiative has established several Centers of Excellence in Data Science, and is focused on enhancing nationwide training infrastructure in biomedical data science as well as data sharing 2 . Radiation oncology holds significant promise for AI-powered tasks-described in several perspectives and reviews 3-8 -not just for optimizing workflows or diagnosis, but also more rewarding tasks such as prognostic prediction and personalized treatment recommendations.

AI applications in radiation oncology span the domains of both medical physicists and radiation oncologists.

Some applications, such as auto-segmentation and automated treatment planning, will be human-verifiable; in other words, a human can check the work of a computer prior to deployment. Other applications-survival prognostication, decision support, and genomics-based treatment planning are not human-verifiable at an individual scale and will thus require careful model development and validation. As applied research in these applications grows in radiation oncology, a commensurate growth in education is necessary to be able to build and validate trustworthy AI models that can be applied to the clinic.

Separate surveys of trainees in radiation oncology and radiology reveal that the majority are interested in additional training in the AI or informatics 9,10 .

In the radiation oncology survey, the American Society of Radiation Oncology (ASTRO) queried chairs and trainees in 2017 to assess their perception of training and research opportunities in genomics, bioinformatics, and immunology. 10 Among the three areas, bioinformatics received the most enthusiasm: 76% believed that bioinformatics training would "definitely or probably" advance their career. 67% expressed interest in a formal bioinformatics training course and 88% of chairs reported they would "probably or definitely" send faculty or trainees to such a course, reflecting an unmet need in training opportunities. Though the ASTRO survey did not specifically ask about AI/ML, we believe the high interest in bioinformatics accurately reflects interest in quantitative analysis in line with AI/ML methods 11 . In recognition of the need for radiation oncologists with specialized informatics training, the NCI, Oregon Health & Science University, and MD Anderson Cancer Center have each created training programs specific for radiation oncology fellows/residents aimed for careers as medical director in informatics and/or formal board certification in clinical informatics 12 .

The radiology survey supports a sentiment towards AI that is similar to that in radiation oncology 13 . A single institution survey of a radiology department revealed concerns about job security but also enthusiasm to learn about AI/ML 9 . This survey showed that 97% of trainees (residents and fellows) were planning to learn AI/ML as relevant to their job (vs. 77% of attending radiologists). In fact, 74% of trainees (vs. 60% of attendings) were willing to help create or train an ML algorithm to do some of the tasks as a radiologist. National radiology societies have been responsive to these sentiments. The American College of Radiology (ACR) Data Science Institute (https://www.acrdsi.org/) recently launched the ACR AI-LAB™ to allow radiologists to create, validate and use models for their specific local clinical needs. The Radiological Society of North America (RSNA) and the Society for Imaging Informatics in Medicine (SIIM) co-sponsor the National Imaging Informatics Course, held twice a year (in its 3rd year); the majority of residency programs have participated (https://sites.google.com/view/imaging-informatics-course/). The RSNA annual meeting hosts several AI refresher courses and coding challenges in the new "AI Pavilion" with residents encouraged to participate. The SIIM Resident, Fellow, Doctoral Candidate, Student (RFDS) society hosts monthly journal clubs and promotes mentoring opportunities (https://siim.org/page/rfds_community_inter).

In this perspective article written by the Training and Education Working Group of the NCI Workshop on AI in Radiation Oncology (Shady Grove, MD, April 4-5, 2019) 14 , we propose an overall action plan for radiotherapyspecific AI training that is comprised of the action points outlined in Table 1 . We cover each action point (AP) in detail in this article.

In the last decade or so, we have seen several examples of ethical concerns and biases magnified by AI. When there are biases in the training data (e.g., certain populations or scenarios are over represented), then an algorithm that models correlations could propagate or even amplify these biases, leading to undesirable outcomes in deployment 15 . This is particularly problematic as AI is sometimes viewed as being "objective" without consideration for the data generation process, which is often unknown.

The European Union (EU) has recently released a seven-point action plan towards so-called "trustworthy" AI.

This plan focuses on the ethical aspects of AI and includes: human agency and oversight; robustness and safety; privacy and data governance; transparency; diversity, non-discrimination and fairness; societal and environmental well-being; and accountability (https://ec.europa.eu/futurium/en/ai-allianceconsultation/guidelines). Similarly, the Food and Drug Administration (FDA) has taken similar steps towards regulation of AI applications in medicine 16 . A key component of improving awareness is to be transparent and clearly document where and when an AI algorithm is used in any part of the clinical workflow. And in cases

where AI is applied, researchers and physicians should also clarify whether the AI is an ML system-which are 5 the more recent type of AI trained on large data and tend to be less interpretable-or an older rule-based system. ML and rule-based AI have different behaviors. For example, neural networks-a type of ML systemare vulnerable to adversarial attacks 17, 18 .

There are currently no educational guidelines for AI training in radiation oncology or medical physics residents.

Serendipitously, there is an active discussion within the field about revising the radiation oncology resident training curriculum. While in depth discussion of all the factors at play is beyond the scope of discussion here, we refer readers to a pair of editorials by Amdur and Lee 19 and Wallner et. al 20 . In July 2020, the Accreditation Council for Graduate Medical Education made several changes to the radiation oncology residency curriculum. 21 The revisions are notable for mandating education in several new areas, including clinical informatics. We are pleased that the ACGME has the foresight to update training curriculum to include informatics and hope that this paper can serve to provide high-level guidance.

In Action Point 2, we propose a high-level overview of a curriculum draft for trainees in medical physics and radiation oncology to adequately grasp the basic principles of AI. These principles are generalizable to medicine as a whole and have particular significance for interventional and informatics-heavy specialties such as radiation oncology.

Responsible conduct of AI, bias, and disparities 2.

Methodology: data science basics 3.

Interpreting data and models 4.

Practical experience and applications 5.

Data sharing: logistics and culture

There is increasing concern that AI models influenced by bias will further perpetuate healthcare disparities for patients. The underlying reason behind why bias is retained in AI models is often related to training data which fails to represent the entire population equally. Because AI algorithms do not have a concept of "fairness", 6 surveillance of inherent bias with AI is typically left to those who designed the system. As noted by the EU/FDA in Action Point 1, proper application of AI should aim to enhance positive social change and enhance sustainability and ecological responsibility. Particularly in medicine, rules and regulations should be put in place to ensure responsibility and accountability of AI systems, their users and their appropriate utilization. In the computer science and ML communities, there has been increasing efforts to improve the teaching of ethics and human-centered AI in coursework (https://stanfordcs181.github.io/) 22 . A complementary area of work is to develop methods to audit AI systems in order to identify potential systematic or cultural biases.

Trainees must develop an appreciation for these critical complexities and potential limitations of AI.

Data features, structures, and algorithms form the foundation of AI applications. Unfortunately, quantitative analysis and critical data appraisal are not universally emphasized in medical or post-graduate education, particularly for physicians. As many ML techniques become published in general medical or oncology journals, it is incumbent upon editors and readers alike to have some basic facility with the techniques. Building a working knowledge of basic statistical concepts such as hypothesis testing, confidence intervals, and basic performance metrics will need to be introduced before more data structures and model-agnostic techniques such as data cleaning, cross validation, model fitting, bias-variance trade off, and advanced performance metrics, such as the widely-used but poorly-understood receiver operating curve 23 . To de-mystify many of these topics, there are existing high-quality online courses made broadly available, which will be further discussed in Action Point 2 and Action Point 3.

For proper clinical application of AI tools, physicians should be able to assess the validity of the data and the model-generation process. So-called "black box" models have such internal complexity that they are conceptualized as inputs mapped to outputs without any intent to understand how the mapping occurs.

Several ML methods, including deep learning (DL) and most ensemble methods, fall into this categorization.

While black box AI models can have excellent performance during training and internal validation, they often encounter problems generalizing when widely deployed. Understanding why a problem occurred can be difficult with "black box" models and is currently a very active area of AI research 18, 24 .

One way to demonstrate data and model interpretability is through "use cases." In medical research, there are well-known examples of the potential dangers of black box models related to confounding 25 . Fortunately, researchers were able to catch these issues before deploying their models, which may not always be the case for complex datasets with nonobvious confounders.

There is an ongoing discussion on the necessity of AI interpretability by the FDA 16, 26 and the informatics community 27 . All authors would agree that elevating the knowledge base of clinicians and physicists will certainly enable more innovation regardless of final regulatory plans.

For trainees interested in applying data science to clinical practice, these opportunities should be encouraged and promoted.

While medical physics and radiation oncology AI curricula could have significant overlap, there will necessarily be focuses on separate domains. In medical physics, instruction may cover methods for auto-segmentation, automated/adaptive treatment planning, and quality assurance. Radiation oncology trainees may be more interested in prognostic predictions and clinical decision support. In the future, as AI takes more of an augmented intelligence role, there should be instruction for physicians for how to decide whether to accept, interrogate or reject recommendations. For example, physicians may need to determine whether there is sufficient rationale to accept an automatically generated plan or treatment recommendation using clinical and dosimetric information.

Several radiation oncology departments have AI/ML researchers who could contribute to a training curriculum. These courses should be jointly taught to both physicists and physicians. We anticipate that common courses and collaboration between trainees in medical physics and radiation oncology will improve 8 translation of AI methods into the clinic. Given that medical physicists already have quantitative training in methods with significant overlap with ML, we anticipate close collaboration between physicists and physicians. Indeed, this is the current status quo in most radiation oncology departments performing AI/ML research.

For departments without access to sufficient resources, online education using so-called MOOCs ("massive open online courses;" a misnomer as they are not necessarily massive or open) and workshop models (see Action Point 4) may be more educational to trainees than co-opting faculty without training in AI/ML.

For advanced practitioners, we will discuss data science hackathons and crowdsourcing in Action Point 3.

One of the key aspects of creating robust predictive models is being able to show generalization to novel datasets through a process called external validation, which requires institutions to share data among themselves. The data sharing culture in medicine has been historically tribalistic but has gradually become more collaborative. This dynamic was well exemplified by the backlash to an infamous 2016 editorial (coauthored by then-editor-in-chief of the New England Journal of Medicine) that was viewed as anti-data sharing 28, 29 . Unlike in academic medicine, academic AI researchers have a strong open-access culture where pre-print archiving of publications is the norm and datasets are simultaneously published with papers to invite validation. Notably, patients are generally supportive of the sharing of their data and would likely embrace scientific reuse of their data to improve the lives of future patients 30 , though we recognize that there are many regulatory limitations to widespread data sharing of this sort. Finding a path for controlled data sharing amongst trusted parties, or more broadly with de-identification schemes could be an important first step in improving the accuracy of AI algorithms.

In this curriculum, we hope to emphasize the efforts in medicine and oncology to promote data sharing ( Table   2 ). The NCI is keen on improving data sharing protocols and resources. In 2018, the NCI Office of Data Sharing One approach to overcome data transfer medicolegal/PHI issues is through distributed or federated learning.

In this approach, analysis is performed locally and models are transferred (e.g., feature weights) instead of data; this decentralized approach has shown equivalent performance to that using central pooling of data 33, 34 .

Such innovative approaches for anonymization could be part of a training curriculum to help overcome barriers to data sharing.

sharing, along with formalization of key principles in data sharing, namely that data should be FAIR: findable, accessible, interoperable, and reusable. These FAIR guiding principles for scientific data management and stewardship 35 are of utmost importance, and should be discussed with and endorsed for all trainees. In line with FAIR, several radiation oncology academic centers and cooperative groups have contributed datasets to the TCIA [36] [37] [38] . Open access journals with a focus on radiation oncology include BMC Radiation Oncology, the Frontiers in Oncology section on radiation oncology, and Advances in Radiation Oncology, which was launched by ASTRO in 2015.

Several coordinating efforts present opportunities to pool ideas and data to promote collaborating, increase power for discovery, and avoid redundancy. Within imaging, these efforts include the aforementioned ACR Data Science Institute for AI in medical imaging, which aims to identify clinically-impactful use cases in radiation oncology, such as auto-segmentation and MRI-derived synthetic CT scans 39 . Within genomics, the Radiogenomics Consortium is a transatlantic cooperative effort pooling American and European cohorts to find genomic markers for toxicity to radiotherapy 40 . Several groups within the consortium are interested in creating ML models to predict toxicity response in radiotherapy [41] [42] [43] [44] .

Through the proposed curriculum draft of Action Point 2, we hope to build a core of trainees for the next generation who can understand and apply data science fundamentals while also understanding ethical considerations and data sharing principles. As there is more interest, trainees will likely want to be involved in practical research projects. Given that AI expertise is not evenly distributed, both intra-and inter-institutional collaborations can be fostered. In this respect, trainees can provide a valuable service by annotating data for research. At the same time, they would also benefit from the service of others by receiving annotated data for model and skills development. A model for this can be seen in eContour (https://www.econtour.org), a free web-based contouring atlas. In a randomized trial, eContour improved nasopharynx contours and anatomy knowledge compared to traditional resources 45 

Developing and maintaining resources described in Action Point 3 will require accelerated learning of particularly motivated trainees who will need institutional infrastructure and funding mechanisms to be successful.

While MOOCs provide consistency and quality of education, for accelerated training, the radiation oncology community could adopt the intensive weeklong workshop model that is widely used by oncology organizations 11 . Examples include separate workshops on clinical trial development by ASCO/AACR , and what support from institutions, societies, and funding agencies is required (AP4). We hope that this paper will spark further discussion on the future of trainee education in radiation oncology. One concrete path forward could be for radiation oncologists and medical physicists to collaboratively apply for fellowships/funding and develop workshops (AP4) for creation of data science educational curricula (AP2) and resources (AP3), while being mindful of the ethical concerns in AI implementation (AP1).

The contents do not represent the views of the U.S. Department of Veterans Affairs or the United States Government. and statistical ML; ML includes support vector machines and neural networks; DL includes deep neural networks and convolutional neural networks; big data can be described as data having volume, velocity, variety, veracity/variability, and value (https://www.ibm.com/blogs/watson-health/the-5-vs-of-big-data/); data analytics refers to the process of making meaningful predictions and models, as exemplified by the work of several authors referenced in this paper 24, 25, 33, 41, 43, 47 . 

The National Institutes of Health's Big Data to Knowledge (BD2K) initiative: capitalizing on biomedical big data

Predicting outcomes in radiation oncologymultifactorial decision support systems

Machine Learning Approaches for Predicting Radiation Therapy Outcomes: A Clinician's Perspective

Big Data and machine learning in radiation oncology: State of the art and future prospects

El Naqa I. Big Data Analytics for Prostate Radiotherapy

Artificial intelligence in radiation oncology: A specialty-wide disruptive transformation?

The Role of Artificial Intelligence in Diagnostic Radiology: A Survey at a Single Radiology Residency Training Program

Assessing the Training and Research Environment for Genomics, Bioinformatics, and Immunology in Radiation Oncology

Career Enrichment Opportunities at the Scientific Frontier in Radiation Oncology

Regard to Wallner

Artificial intelligence in radiology: friend or foe? Where are we now and where are we heading?

NCI Convenes Workshop on Artificial Intelligence (AI) in Radiation Oncology | News & Events | DCTD

AI can be sexist and racist -it's time to make it fair

Food & Drug Administration. Clinical and Patient Decision Support Software Draft Guidance

Explaining and Harnessing Adversarial Examples

Interpretation of Neural Networks is Fragile

A Call for Change in the ABR Initial Certification Examination in Radiation Oncology

The meaning and use of the area under a receiver operating characteristic (ROC) curve

Intelligible Models for HealthCare: Predicting Pneumonia Risk and Hospital 30-day Readmission

Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: A cross-sectional study

The Challenge of Regulating Clinical Decision Support Software After 21(st) Century Cures

AMIA Urges More Work on FDA's Decision Support Guidance | AMIA

NEJM Editor Backtracks on Data-Sharing "Parasites

NEJM Editor Flip Flops On Data Sharing After Social Media Firestorm. CardioBrief

Patient-Led Data Sharing -A New Paradigm for Electronic Health Data. NEJM Catalyst

NIH's DRAFT Data Management and Sharing Policy: We Need to Hear From You! Office of Science Policy

American Association of Physicists in Medicine Task Group 263: Standardizing Nomenclatures in Radiation Oncology

Developing and Validating a Survival Prediction Model for NSCLC Patients Through Distributed Learning Across 3 Countries

Distributed deep learning networks among institutions for medical imaging

The FAIR Guiding Principles for scientific data management and stewardship

Data From NSCLC-Radiomics

Data from NSCLC-Cetuximab

Imaging and clinical data archive for head and neck squamous cell carcinoma patients treated with radiotherapy

Artificial Intelligence in Radiation Oncology Imaging

Establishment of a Radiogenomics Consortium

Computational methods using genome-wide association studies to predict radiotherapy complications and to identify correlative molecular processes

Machine Learning on a Genome-wide Association Study to Predict Late Genitourinary Toxicity After Prostate Radiation Therapy

Genomics models in radiotherapy: from mechanistic to machine learning

Multi-institutional Randomized Trial Testing the Utility of an Interactive Three-dimensional Contouring Atlas Among Radiation Oncology Residents

Machine Learning Applications in Head and Neck Radiation Oncology: Lessons From Open-Source Radiomics Challenges

Use of Crowd Innovation to Develop an Artificial Intelligence-Based Solution for Radiation Therapy Targeting

Opportunities and challenges to utilization of quantitative imaging: Report of the AAPM practical big data workshop

State of Data Science in Radiation Oncology Workshop. In: NIH National Cancer Institute

The European Organisation for Research and Treatment of Cancer, State of Science in radiation oncology and priorities for clinical trials meeting report

Predict overall survival and docetaxel discontinuation in mCRPC

Kaggle Data Science Bowl 2016 Predict heart ejection fraction MICCAI radiomics challenges

HPV (2) local control in oropharyngeal cancer Kaggle Data Science Bowl 2017 Detect lung cancer via National Lung Screening Trial DICOMs TopCoder Lung Cancer Challenge

Digital Imaging and Communications in Medicine; DREAM: Dialogue for Reverse Engineering Assessments and Methods; HPV: human papillomavirus; MICCAI: Medical Image Computing and Computer Assisted Intervention Society; mCRPC: metastatic castrate resistant prostate cancer