Pictures into numbers

How AI and quantitative assessments deliver insights and value for medical imaging

By | 4 minute read | September 2, 2020

One of the key advantages artificial intelligence (AI) brings to medical imaging is its ability to transform images into numbers. The concept and use of quantitative imaging has reached prevalence over the past few years by enabling computers to do more than just display digital images. Numerical measurements such as volume, dimension, shape, growth rate and more can be extracted from imaging data, placed into databases and analyzed to reveal anomalies, patterns, relationships and insights that strictly qualitative, human analysis might not.

But how does AI support this process? What AI principles underpin the technology? And how does it deliver value?

A good place to start is by simply understanding what is meant by artificial intelligence — and specifically in a medical imaging context. IBM® cites the popular usage of AI as “…the ability of a computer or machine to mimic the capabilities of the human mind — learning from examples and experience, recognizing objects, understanding and responding to language, making decisions, solving problems…”

In imaging, three types of AI are typically applied:

  • Neural networks are based on the way neurons in the brain operate. These networks can be trained to identify images by being fed data about those images. For example, a neural network system can learn to identify an apple by being given images labeled as “apple” or “not apple.”
  • Machine learning is the application of neural networks that enable computers to learn how to identify images on their own — without explicit programming. By being given more data — more images of apples — machine learning teaches itself to become better at identifying an apple.
  • Deep learning is a subset of machine learning that uses layers of choices — each refining the previous — to perform specific tasks with ever-increasing accuracy. Deep learning systems can handle large data volumes and address complex tasks such as prioritizing criteria most important to reaching a decision.

These techniques — and the quantitative output they generate — are well suited to the general and specific applications and challenges faced in cardiology and radiology today:

  • Data overload presents two challenges: first, the sheer number of images and, second, the volume and complexity of data in electronic health records (EHR).

A single study can produce thousands of images. Multiply that by millions of studies conducted each year and the figures are staggering. The good news is AI thrives on data. The more images machine learning applications are provided, the more effective their training becomes — and the more potential they have to help clinicians uncover nuances and gain insights.

The amount of information in health records is massive, largely unstructured — such as clinical notes — and difficult to review. It’s also critical to informing clinically beneficial readings. Natural language processing, a field of machine learning, can understand semantic information in EHRs and summarize it to support better-informed and timely decisions. Hardin Memorial Health, for example, is using an AI solution from IBM Watson Health® to reduce the effort radiologists spend sifting through patient data in EHRs.

  • Identifying high-value findings and avoiding errors have become critical as image volumes increase. IBM reports that radiologists view over 84 million studies and 90 billion images each year, and according to Dr. David Gruen at IBM Watson Health, “this pace is not sustainable, even for the most experienced radiologists working under ideal conditions… there simply aren’t enough of us radiologists to make this formula work over the long term.” AI’s ability to analyze and quantify imaging data makes it an ideal partner for overworked clinicians. Machine and deep learning techniques can discover, isolate, prioritize and report specific anomalies that humans might miss or not initially consider.
  • Workflow and prioritization considerations are vital to making quantitative analysis actionable. AI techniques can triage reviews based on data analytics — and present that information as part of existing systems and workflows.
  • Human bias in readings can skew diagnosis and take treatments in sub-optimal directions. It typically stems from heuristic principles — using assumptions and past experiences to speed up decisions. This can lead to two basic types of errors: perceptual — unseen abnormalities — or interpretive — identified abnormalities that are incorrectly interpreted. Quantitative assessments are less susceptible to heuristic expediencies and can help physicians make critical decisions with comprehensive numerical data. For example, ensuring that all injuries are identified for patients in trauma centers, identifying cardiovascular abnormalities or detecting subtle variations to enhance musculoskeletal treatments and diagnoses.1

The use of AI in medical imaging and its ability to support quantitative assessments is established and growing. In fact, the American College of Radiology’s Data Science Institute (ACR DSI) has established a set of use cases that address abdominal, breast imaging, cardiac, musculoskeletal, neurology, oncology, pediatric and other applications. According to ACR DSI the use cases “provide a previously missing understanding with end users on what an AI solution is to deliver by providing functional requirements and enabling better analysis and test models.” The use cases are updated and freely available at the ACR DSI website.


Still wondering how AI can help you? Attend a webinar series that shows concrete, clinical AI applications that can support the reading activities of physicians in radiology and cardiology.