Peter Shen
AI’s still-to-do list
October 17, 2022
By Peter Shen
Only a few years ago, radiology’s artificial intelligence solutions were touted to the rafters at medical imaging conferences. In 2022, the hype dust has settled. Certainly, the radiology community recognizes AI’s core benefits. It saves time by automating repetitive tasks. It also leverages pattern recognition to identify areas of concern that the radiologist might miss.
But the initial, pre-pandemic question of “What immediate value can AI bring to the radiologist?” has given way to new questions. They underscore how much AI must evolve to provide broader, more substantive value — to the radiologist, the healthcare institution, and the patient.
In addition to its established time-saving abilities and quantitative measurements, AI has begun to provide qualitative visualizations and guidance associated with suspected malignancies. While encouraging, this development raises a question: What is the rationale behind this qualitative guidance? A radiologist develops a reading methodology over a long career. Now, AI is introducing additional data points that may not even be relevant to that radiologist’s diagnosis. Suddenly, AI’s purported time-saving benefit is debatable. How does the radiologist contextualize this new data and incorporate it into a reporting style?
If AI offered imaging descriptors that elaborated on the rationale for its diagnosis, the radiologist could better decide whether to agree with AI’s findings. This more informed decision-making could improve the identification of false positives and increase diagnostic accuracy. AI-generated imaging descriptors also would help the radiologist explain a diagnosis to radiologists-in-training, who will ultimately incorporate AI into their routine reading and reporting practices.
Radiologists also want AI’s findings to be integrated into their routine diagnostic workflow. This integration would solidify AI’s practical value, which healthcare institutions eye closely. But AI has a potentially different value for a clinician in the emergency room than in the intensive care unit, and a different value when treating inpatients versus outpatients. Can we examine how AI’s value differs with reading environment? Furthermore, how often must AI find something that the radiologist would miss to be considered valuable? Should the radiologist only use AI for studies that he or she is unsure about? Must AI be applied to every case for it to have value?
As we drive toward more personalized diagnostic and treatment decisions, AI’s implicit bias is another area of concern. For example, a breast cancer algorithm can be trained on a large batch of patients from the Northeast. But patients across the country, much less the world, are hardly identical. Can vendors develop a breast cancer algorithm that is batch-trained using Northeastern patients and can also learn from new patient data as it is processed? This self-adjusting algorithm would be more complimentary to the patient population to which it is being applied — be it women from the South or South America.
The fact that radiologists are asking these questions, and providing feedback to vendors, is significant. Undoubtedly, this scrutinization of AI — and for that matter, its very adoption by healthcare institutions — would gain momentum if reimbursement incentives were tied to AI’s use in radiology. Reimbursement is likely forthcoming, but its form and arrival date are unclear. Lack of reimbursement must not weaken radiology’s collective resolve to improve AI and maximize its value.
AI has done much for medical imaging in the relatively short time that it has been commercially available. If radiologists and vendors address these questions, it will advance beyond its early promises of process automation and efficiency gains to enable more informed decision-making. AI will then become a more fully realized tool — one that is more fully embraced by the medical imaging community and, ultimately, more beneficial to the patient.
About the author: Peter Shen is the head of the digital & automation business at Siemens Healthineers North America.
Only a few years ago, radiology’s artificial intelligence solutions were touted to the rafters at medical imaging conferences. In 2022, the hype dust has settled. Certainly, the radiology community recognizes AI’s core benefits. It saves time by automating repetitive tasks. It also leverages pattern recognition to identify areas of concern that the radiologist might miss.
But the initial, pre-pandemic question of “What immediate value can AI bring to the radiologist?” has given way to new questions. They underscore how much AI must evolve to provide broader, more substantive value — to the radiologist, the healthcare institution, and the patient.
In addition to its established time-saving abilities and quantitative measurements, AI has begun to provide qualitative visualizations and guidance associated with suspected malignancies. While encouraging, this development raises a question: What is the rationale behind this qualitative guidance? A radiologist develops a reading methodology over a long career. Now, AI is introducing additional data points that may not even be relevant to that radiologist’s diagnosis. Suddenly, AI’s purported time-saving benefit is debatable. How does the radiologist contextualize this new data and incorporate it into a reporting style?
If AI offered imaging descriptors that elaborated on the rationale for its diagnosis, the radiologist could better decide whether to agree with AI’s findings. This more informed decision-making could improve the identification of false positives and increase diagnostic accuracy. AI-generated imaging descriptors also would help the radiologist explain a diagnosis to radiologists-in-training, who will ultimately incorporate AI into their routine reading and reporting practices.
Radiologists also want AI’s findings to be integrated into their routine diagnostic workflow. This integration would solidify AI’s practical value, which healthcare institutions eye closely. But AI has a potentially different value for a clinician in the emergency room than in the intensive care unit, and a different value when treating inpatients versus outpatients. Can we examine how AI’s value differs with reading environment? Furthermore, how often must AI find something that the radiologist would miss to be considered valuable? Should the radiologist only use AI for studies that he or she is unsure about? Must AI be applied to every case for it to have value?
As we drive toward more personalized diagnostic and treatment decisions, AI’s implicit bias is another area of concern. For example, a breast cancer algorithm can be trained on a large batch of patients from the Northeast. But patients across the country, much less the world, are hardly identical. Can vendors develop a breast cancer algorithm that is batch-trained using Northeastern patients and can also learn from new patient data as it is processed? This self-adjusting algorithm would be more complimentary to the patient population to which it is being applied — be it women from the South or South America.
The fact that radiologists are asking these questions, and providing feedback to vendors, is significant. Undoubtedly, this scrutinization of AI — and for that matter, its very adoption by healthcare institutions — would gain momentum if reimbursement incentives were tied to AI’s use in radiology. Reimbursement is likely forthcoming, but its form and arrival date are unclear. Lack of reimbursement must not weaken radiology’s collective resolve to improve AI and maximize its value.
AI has done much for medical imaging in the relatively short time that it has been commercially available. If radiologists and vendors address these questions, it will advance beyond its early promises of process automation and efficiency gains to enable more informed decision-making. AI will then become a more fully realized tool — one that is more fully embraced by the medical imaging community and, ultimately, more beneficial to the patient.
About the author: Peter Shen is the head of the digital & automation business at Siemens Healthineers North America.