To adapt and deploy AI models faster for assessing specific patient populations, GE HealthCare and Mass General Brigham are researching the potential of medical imaging foundation models, which utilize self-supervised learning to bypass the laborious, manual tasks required for this feat.

When introducing an AI application to a new regional or local group of patients, teaching the model about the specific care needs and conditions that the population faces requires manually labeling and inputting large data sets into it, a process that is laborious and time-consuming. It also increases costs and complexity and makes broad adoption of these technologies less appealing in healthcare.

Through self-supervised learning, also known as pretraining, foundation models learn patterns from unlabeled data sets and computational resources that not only provide them with insights into specific patient populations but act as a basis when coming across new findings, allowing them to adapt to tasks and care regimens quickly with fewer labeled examples. This is especially helpful in cases that are rare or complex to label at scale.

"Incorporating responsible AI practices into this phase, we are committed to ensuring these innovations adhere to guidelines, prioritize patient safety and privacy, and promote fairness and transparency across all applications,” said Parminder Bhatia, chief AI officer of GE HealthCare, in a statement.

The introduction of foundation models into their research builds on both organizations’ prior 10-year commitment made in 2017 for developing and exploring sustainable AI uses for diagnostic and treatment regimens.

Most recently, the two rolled out the Radiology Operations Module, the first in a series of joint predictive AI innovations for improving patient scheduling. The ROM detects so-called missed care opportunities, where patients are late, fail to schedule a follow-up, or miss an appointment, allowing clinicians to allocate more time to patients and reducing administrative burdens.

The addition of foundation models is expected to have a similar impact, streamlining imaging workflow and accelerating diagnoses, according to Dr. Keith Dreyer, chief data science officer at Mass General Brigham.

“I think we are all optimistic that foundation models may actually complement and enhance the work we have been doing with convolutional neural networks over the past few years. Hopefully, this work will help make healthcare delivery more efficient for our practitioners, more accessible for our patients, and more equitable for our diverse communities.”

Back to HCB News

Business Affairs Homepage