Clinical Adoption of AI Across Image Producing Specialties

With the rapid increase in the number of artificial intelligence (AI) publications across medical image-producing specialties, clinical communities are becoming increasingly interested in whether these AI solutions will eventually be adopted in routine clinical workflows and, if so, what steps are required to get there. On October 11, 2022, a group of clinical and AI experts got together for a web-based panel hosted by the Society for Imaging Informatics in Medicine (SIIM) to debate the answers to these questions and share their perspectives on many aspects of incorporating AI into clinical workflows.

Kirti Magudia, MD, Ph.D., Marilyn Bui, MD, Ph.D., Roxana Daneshjou, MD, Ph.D., and Camilo Brandao-de-Resende, MD, Ph.D. were the four distinguished guests who, respectively, represented the communities of radiology, dermatology, pathology, and ophthalmology during the SIIM webinar titled “Clinical Adoption of AI Across Image Producing Specialties.” The virtual panel, co-hosted by SIIM Machine Learning Education Sub-committee Co-Chairs, Felipe Kitamura, MD, Ph.D. and Timothy Kline, Ph.D., composed of two major sections. In the first segment of the panel, each speaker offered an overview of the current state of AI in their respective clinical field, and in the second segment, all panelists responded to the co-hosts’ more specific questions. Each of the two sections of the panel is summarized in the following paragraphs.

Part 1. Current State of AI in Medical Specialties:

Dr. Kirti Magudia, Assistant Professor of Radiology at Duke University School of Medicine, recognized opportunistic screening of pathologies, the use of AI in education, and the mitigation of AI bias as the three most intriguing applications of AI in radiology, and emphasized the importance of developing standards for incorporating AI into clinical practice.

Dr. Roxana Daneshjou, a Clinical Scholar at the Stanford University School of Medicine, named photo capture using light photography as the most common imaging modality in dermatology. She noted, however, that photo capture using light photography is not standardized and this obstacle, as well as siloed datasets of imaging data in dermatology, lack of biopsy-proven labels, and limited availability of skin tone descriptions have impeded the widespread application of AI in clinical dermatology.

Dr. Marilyn Bui, Professor of Pathology and Scientific Director of the Moffitt Cancer Center, opened her discussion by noting that present AI technologies should be considered more as help for clinicians than as their replacements. Therefore, “augmented intelligence” may be a more appropriate term than “artificial intelligence” to use in medical areas. She continued by mentioning pathology detection, classification, prediction, quantification, and prognostication as the five main use-cases of AI in clinical pathology, and while pointing to the yearly increasing number of AI publications in this field, emphasized that up to now, there exists only one FDA-approved AI algorithm in pathology. She concluded by citing collaboration between different stakeholders as the key  element in expanding the utility of AI models in clinical practice.

Finally, Dr. Camilo Brandao-de-Resende, an Honorary Research Fellow at the Moorfields Eye Hospital in London, and a senior machine learning specialist, introduced external eye photos, fundus photos, angiographies, optical coherence tomographies, corneal topographies, and visual fields as the main sources of imaging data in ophthalmology. He mentioned three FDA-approved AI algorithms for detecting diabetic retinopathy, glaucoma, and age-related macular degeneration, as well as two promising areas of AI research in ophthalmology: 1) the use of augmented reality to predict outcomes or diagnoses during eye surgery and examination; and 2) oculomics, which investigates ophthalmic biomarkers of systemic and central nervous diseases.

Part 2. Question and Answers:

Q1. Where is AI in terms of deployment and clinical practice in your clinical field?

The best implementation, integration, and monitoring of AI systems in the context of clinical settings were cited by Dr. Magudia as the current objectives for AI application to radiology. She also underlined that the idea that AI would eventually replace radiologists has passed, and that instead we should concentrate on reducing the biases of AI systems and making them more valid and reliable in actual use. According to Dr. Daneshjou, there is currently just one FDA-approved AI device for dermatology, so it is too early to discuss the use of AI in this sector. Dr. Bui stressed the need for establishing infrastructures (such as to digitize the pathology slides) and mentioned inter-stakeholder collaboration as the key to achieving this aim with regards to applying AI in pathology. Last but not least, Dr. Brandao-de-Resende mentioned that the methodological and generalizability concerns are the current obstacles to using AI in ophthalmology. However, since AI is a young area, these barriers will be overcome as time goes on.

Q2. What are the main barriers to successfully implementing AI in the clinics?

Although panelists mentioned the need for the development of more valid and generalizable models as one of the obvious barriers to the deployment of AI in clinical settings, almost everyone concurred that other significant groups of barriers to the deployment of AI are found in the infrastructure domain. On the one hand, difficulties with logistics and a lack of practical interfaces for AI tools were cited as major obstacles. The present models for using AI in clinical practice have some drawbacks, including unfriendly ways for end users to interact with AI technologies, the paperwork required to deploy those systems, and the additional labor required to monitor their effectiveness. On the other hand, it was indicated that the quality of medical imaging data greatly depends on the devices accessible in various institutions. This heterogeneity in devices frequently results in diverse quality of input data to the AI models and, in the end, their unreliable performance. The panel’s ultimate recommendation was that starting small is preferable to trying to be superman and handle all of radiologists’ problems at once. We should take small steps when using AI, get input from various stakeholders, and expand our workflows and algorithms more gradually.

Q3. Should clinicians be feared of being replaced with AI?

The idea that AI will displace physicians was rejected by all of the panelists.They all highlighted that the purpose of AI is to support medical professionals, and that many of the recent and current discussions about AI displacing clinicians are based on exaggerated media reports. In fact, assertions that AI performs better than physicians should be compared to more reliable study designs, such as randomized clinical trials and fairness or bias studies. If this is done, there will be a large number of physicians eager to accept the help that AI can provide in their profession.

 Closing considerations

Most of the lessons learned in Radiology could expedite the development of AI in other specialties. However, there are unique challenges in each specialty.. For example, dermatology has the potential to develop direct-to-consumer AI products, while pathology needs to add an extra step to digitalize the whole slide images before we consider developing AI models.

 The panelists and co-hosts shared their feeling that SIIM is in a unique position to promote discussions about similarities and dissimilarities in AI implementation across different image-producing medical specialties, which is not only interesting but also potentially fruitful.