Fostering a Healthy AI Ecosystem for Radiology: Conclusions of the 2018 RSNA Summit on AI in Radiology

Data Collection and Curation

Data collection and curation were ranked as the top two priority areas by summit participants. AI algorithms require large volumes of data to train a deep learning system effectively. This requirement underscores the need for scalable and efficient systems for data discovery, collection, curation, transfer, and manipulation (16). This technology gap presents an opportunity for RSNA to facilitate development in partnership with industry leaders.

Many summit participants focused on concerns about security, privacy, and limited interoperability that prevent sharing of image data between health care organizations or between health care organizations and AI vendors. Stakeholders face important trade-offs when considering data sharing. Greater portability and shareability of data for algorithm development can benefit patients, but at a greater risk to patient privacy.

A potential solution to the data-sharing problem could be to develop “sandbox” environments that enable institutions and their vendor partners to develop and test algorithms jointly so data security could be maintained and health care data access could be controlled. This approach offers a tenable solution for institutions that are hesitant to release patient data because of security and privacy concerns. Such environments also could be used to annotate, organize, maintain, and present training data from disparate sources. Automated pipelines will be essential to coordinate, aggregate, and store data from various sources in an efficient, sound, and reproducible manner.

Training on limited single-institution datasets usually does not yield algorithms whose accuracy generalizes across sites and patient populations. New research suggests that multi-institutional studies could be conducted without moving data by transferring the weights from an algorithm trained at one institution to one or more additional institutions for sequential training. These distributed learning methods can improve data security without significantly compromising algorithm accuracy (17).

Another major bottleneck in the development of AI algorithms is the need for labeled or annotated imaging datasets and requiring reliable normalization of imaging data across imaging sites. Normalization of images and their respective annotations is particularly challenging when the data originate from disparate sources, such as multiple research sites. The annotation bottleneck is particularly evident if bounding boxes or hand-drawn segmentation by a domain expert is required (18).

There is a need for open standards to facilitate exchange of datasets, algorithms, and validation protocols for multi-institutional collaboration by academic and industry partners. The standards also should allow investigators to compare algorithm performance in different sites and circumstances.

Educating the Radiology Community about AI

Education about AI was ranked highly (Table 3). RSNA can lead the development of medical imaging–specific educational assets that could serve multiple stakeholders, including radiology practitioners (ie, radiologists, radiology nurses, radiologic technologists, physicists, administrators, trainees), informatics and information technology personnel, software developers, radiology educators, researchers, and regulators. This is an appropriate role for RSNA’s educational capabilities.

There is an apparent chasm in communication between radiologists and the data scientists who develop AI algorithms. To bridge the gap, RSNA could develop courses and multidisciplinary groups to stimulate collaboration. Educational materials that could teach imaging to data scientists and data science to radiologists could fulfill this need.

Multimodal AI education is the best method to facilitate knowledge transfer effectively. In addition to journal articles, online webinars, and courses, RSNA could develop podcasts, symposia, workshops, hackathons, and educational fellowships, either on its own or in partnership with other organizations. Both active and passive learning sources developed jointly by radiologists and data scientists could have a significant impact on AI education (19).

These multimodal education materials could be guided by an AI curriculum developed for the radiology community. The curriculum could be located and maintained on the RSNA website as a free resource for members. An example of a joint curriculum resource is the National Imaging Informatics Curriculum and Course (20), a week-long online course jointly sponsored by the RSNA and the Society for Imaging Informatics in Medicine. The course currently is offered twice a year but could be expanded to adapt to the needs and feedback of its attendees. Such jointly sponsored courses also could be developed with other radiologic and data-science societies to offer greater opportunities for data scientists to interact with the radiology community.

AI Deployment in Clinical Practice, Quality Assurance, and Business Performance

AI applications in the clinical workspace will extend far beyond image interpretation and will have a profound impact throughout the imaging life cycle, including processes and events that precede image acquisition and interpretation. For example, AI can assist with clinical decision support and order entry, augment the protocoling and quality control process, produce scheduling efficiencies, facilitate change detection, and reduce contrast material and radiation dose (21). The implementation of AI technology in clinical practice falls into three categories: (a) clinical radiology practice, (b) quality assurance and auditing functions, and (c) evaluation of business performance with key performance indicators (KPIs) (22). We review each of these AI applications.

AI in clinical radiology practice.—The use of AI algorithms for clinical radiology practice is regulated by government organizations. The U.S. Food and Drug Administration (FDA) considers AI algorithms in health care “Software as a Medical Device” (23). Software as a Medical Device have a separate precertification pathway for FDA clearance, with their own definition, framework for risk stratification, quality management system, and clinical evaluation guidelines (24). The European Union (EU) requires the Conformité Européenne (CE) mark for a variety of medical and nonmedical innovations, including AI software. CE mark certification, like the FDA’s process, has stringent qualifying requirements, which can be difficult to navigate for novice developers. In addition to CE mark certification, the EU recently passed the General Data Protection Regulation (GDPR), a law that protects EU citizen data privacy (23). Laws such as GDPR put the power to share patient data into the hands of the patients themselves. The GDPR must be taken into account when designing data sharing and security systems but also when considering how specific EU citizens can use data. Similar laws may emerge in other countries, including the United States, as we gain more experience with AI development and use (25).

To help developers navigate the complexities of the FDA and CE mark processes, and the GDPR and any future legislation, the RSNA could educate radiologists and developers about the regulatory processes, as well as assist with connecting solutions (product) with a suitable need in the clinical space (market). Potential consumers of clinical AI algorithms need to remember that approval or certification by the FDA or CE does not guarantee a specific level of performance; it only signifies to the public that the product performed to a minimally acceptable level. Consumers and regulators must closely scrutinize postmarket monitoring of algorithm accuracy on real-world data. For example, the RSNA could assist by hosting registries for postmarket surveillance of AI algorithms.

Quality assurance and auditing features.—A major criticism of AI technology has been the “black-box” nature of the algorithms, whereby the inputs and outputs of a trained system are transparent to the user, but the decision processes in between remain hidden (1). Although the opacity of AI algorithms is unlikely to resolve completely, increased funding for Explainable AI is needed. Explainable AI methods enable the interrogation of AI models to determine which factors led to the model’s conclusion (26). This area of research has been vigorously supported by the Department of Defense and could be supported by other funding agencies that support AI research, including the RSNA.

Rigorous methods for evaluation, such as randomized controlled trials, can establish trust in an algorithm even when its methods are opaque, especially for AI products that claim to affect patient outcomes with improved diagnostic and detection capabilities. Making training datasets and algorithms freely available when possible also will improve trust. Some researchers advocate open-source publication of algorithms undergoing evaluation by randomized controlled trials to assure transparency. However, not all algorithms can be made available because the radiology software marketplace and industry ecosystem are predicated on the ability to develop proprietary algorithms.

Business performance.—The ability to measure business performance is crucial to organizational success. KPIs are financial and nonfinancial metrics that evaluate the success of a company or organization (27). Hundreds of KPIs exist in radiology (28) and can be divided into groups depending on their use and goals: (a) clinical performance (eg, complication rates, false-positive rates); (b) service-level performance (eg, missed appointments, patient wait time, report turnaround time); (c) utilization and productivity performance (eg, equipment idle time, repeat sequence rates, picture archiving and communication system downtime); and (d) financial performance (eg, net income, preauthorization effectiveness, claims rejection rate) (29).

AI tools can help to understand and optimize the business functions of radiology practices and aid radiology practice management by aiding in the development and evaluation of KPIs. For example, real-time AI evaluation of utilization and productivity KPIs (eg, equipment idle time, picture archiving and communication system downtime) could identify inefficiencies and anomalies that cause bottlenecks amenable to improvement. The veracity of the data being used to measure KPIs and to develop AI algorithms is essential to avoid downstream misleading or biased information (30), which ties to the need for better data collection and curation. RSNA’s role could include maintaining a database of radiology-specific KPIs and their evaluation methods with information on how AI algorithms could be used to measure these KPIs. Another role may be to offer a KPI benchmarking resource for comparison of best practices for business performance across institutions, aided by AI.

Ethical and Legal Implications of AI in Radiology

RSNA could take on a leading role in providing clarity about the ethical use of AI in radiology. A clear differentiation should be made between augmentation versus replacement of radiologist workflow tasks, with the former being an assistive goal and the latter being an adversarial goal (31). Making this distinction could foster guidelines about accountability and risk attribution for AI technology and its developers, radiologists, regulators, health care organizations, and patients (32). RSNA has joined with other organizations in North America and Europe to develop principles for the ethical use of AI in radiology.

Ambiguity about whether AI technology is assistive or adversarial can make it difficult to assign responsibility and liability for undesirable clinical outcomes arising from the use of AI in radiology. This ambiguity raises many unanswered questions: Who is responsible for an unfavorable outcome if an AI algorithm makes a mistake? What is the radiologist’s role and level of responsibility if an AI algorithm misdiagnoses a disease without input from the radiologist?

AI algorithm developers and users have an obligation to avoid bias. How do we ensure that AI technology is developed and used to avoid inherent bias? The possibility of inadvertently developing AI algorithms with sociodemographic, racial, and gender biases is a particular concern (33).