The latest AIMed Clinician Series on June 29th featured a special focus on imaging – a subspecialty that’s undergoing a rapid paradigm shift in performance, driven by AI. Yet, in his keynote address, Alok Gupta, Vice President and Global Manager of Imaging at IBM Watson Health, warned of the challenges ahead before the technology becomes fully adopted.
“What we are seeing is that digital transformation has already started even before COVID-19, with healthcare providers and physicians, the two agents of change,” said Alok Gupta, Vice President and Global Manager of Imaging at IBM Watson Health in the opening keynote session. “They are trying to simplify processes, establish financial strengths, implement new and effective ways to deliver high quality patient care. It’s bringing a lot of excitement but is also surfacing many challenges.”
Overall, as Gupta described, we are seeing exciting solutions for physicians in four major categories. First, AI is extracting clinically relevant information from electronic health records with clinicians reading images to gain deeper insights in order to start examining patients with more context. Next, algorithms running in the background are pre-accessing studies before the clinicians review them and raising alerts when needed.
AI is turning visual observations into quantitative clinical assessments, saving physicians’ time by offering quantification of anatomical structures or conditions. Deep learning has especially evolved to support many applications across organ and lesion segmentation tasks. Finally, detection algorithms identify abnormal findings or lesions in images. These algorithms directly impact clinicians’ perceptions and have the potential to make the greatest impact. Thus, extensive validation studies are essential.
Despite a more robust performance, with the availability of AI tools democratizing technological development, Gupta thinks it will still be a challenge to build AI systems that can perform consistently across various clinical settings. “Many accept that the present AI development is time-consuming, expensive and uncertain,” he said. “When in fact, taking algorithms to the market and going through regulatory clearance is just as time-consuming, expensive and uncertain.”
Indeed, regulatory barriers safeguard the healthcare system and its patients, but slows down the process of bringing new AI products into the market, with only a small number of cleared algorithms impacting the workflow. This, in turn, fragmented the AI imaging field with no single vendor being fully accountable for how imaging AI should operate. Under this landscape, IT departments may not be able to deliver imaging AI at an affordable cost, especially with multiple small vendors concentrating on separate algorithms that are not integrated into the core workflow, affecting the overall interoperability.
“Interoperability requires physicians to capture and share images, reports and a broad set of data. Ideally, it allows them to collaborate on patient cases with peers using industry standards,” Gupta said. “However, what we have on the market today is AI that addresses specific tasks on a specific modality. Altogether, the entire task of selecting, contracting, implementing, and using the AI algorithms is cumbersome and not value-adding.”
As such, Gupta believes for AI to produce trusted insights, special care is needed in the design and training of the algorithms. “Building an AI vendor requires the data scientists having access to meaningful datasets, the knowledge to productize ideas into commercial innovation and deep involvement of clinical experts,” he noted. “The data used to train algorithms should be collected thoughtfully and represent different geographies, ethnicities, and conditions including multiple types of lesions, diseases -especially the rare ones, and a variety of vendors and modalities.”
In addition, Gupta noted the need to monitor overfitting and generalization in the training dataset to ascertain the quality of data and work on the inclusion and exclusion criteria before using them to train the algorithms. “This is just the start,” Gupta said. “With each new AI algorithm, everything will need to be adjusted again. All these continue to add to the complexity of the workflow, making AI adoption at scale particularly challenging.
“This brings us to a solution of orchestrating relevant algorithms and aggregating the results for workflow integration,” Gupta continued. “In this way, information about a patient is sent to an ‘orchestrator’, who will exercise rigorous business logic to determine the best qualified AI algorithms to proceed with the analyses. The results from the selected algorithms are then aggregated into a single package and returned to the physician’s workflow. The physician would comfortably review and weigh the new insights to complete the interpretation of the case.”
Gupta believes the solution allows for a vendor neutral approach to standardize the presentation of AI results and follow through in the workflow to ensure a seamless flow of data across diagnostic imaging systems for proper integration. “In the end, no matter where your imaging application is running, all components should be communicating without obstruction,” Gupta said. “This will support non-stop business continuity and the consistent work experience.”