For decades, medicine has been drawn to artificial intelligence (AI) as a means of enhancing healthcare practice and systems. In the Age of AI, algorithm-powered machine learning (ML) methods able to mimic human cognitive function seem limitless.

For now, real-world ML is limited. Applications such as detection, diagnosis, treatment guidance and prediction of outcomes are more theoretical than available, but promising. For all that AI can bring to medicine, the consensus across disciplines is uniform — it has no place in provider-patient relationships. ML may mimic human cognitive function through analytics, but only human providers can fill the inherently perceptual and contextual role of explaining diagnoses, procedures and treatment.

Preceding the Digital Age, radiology applied computer science to practice earlier than most disciplines. Radiologists, informed by their own informatics experts, arguably became the most digitally skilled healthcare professionals. In the 1980s, PACS technology introduced a masterful systems workflow standard for management, storage, and retrieval of medical imaging data. Close behind, Digital Imaging and Communications in Medicine (DICOM) enabled PACS image storage and transfer across wide-ranging networks and platforms.

In the Age of AI, algorithms feed volumes of data for recognizing patterns that will enhance radiological clinical practice. Convolutional neural networks (CNNs) — algorithms used to solve computer vision tasks — may increase application of AI to detection and identification. ML algorithms that can detect critical findings missed by even a trained radiologist’s eye are at a theoretical, early innovation stage.

Three ML experts have guided our Department through understanding the rapid advancement of AI in pre-clinical imaging development and clinical evaluation stages: 1) Assistant Professor and Chief Medical Information Officer Dr. Benjamin Mervak; 2) Associate Professor Dr. Pew-Thian Yap; and 3) Radiology Informatics Manager Dr. Jay Crawford.

Dr. Mervak began his practical training in imaging informatics during his fourth residency year at the University of Michigan. Since 2017, he has helped guide our direction at UNC by investigating pre-clinical ML innovations holding future promise. He’s also established other day-to-day radiological workflow improvements (eg, Epic@UNC /Radiant process improvements, critical results reporting, and automated clinical decision support).

At the UNC IDEA Lab, senior investigator Dr. Pew-Thian Yap oversees development of novel image analysis tools that could determine disease pre-determinants, diagnoses, treatment and prognosis once advanced to clinical evaluation.  Much of Dr. Yap’s work applies decades-old AI thinking to modern-era deep learning that analyzes massive, high-quality data sets to develop pattern recognition representations.

Dr. Yap’s work with deep learning innovation at the UNC Idea Lab has produced a range of promising ML methodology.  Such advancements include ML-driven MRI reconstruction that speeds up data acquisition by an order of magnitude, growth prediction of the human cortex in the first year of life, as well as automated real-time quality assessment of MR images.

Jay Crawford’s career has spanned imaging engineering, PACS, and informatics. As a biomedical engineer, he worked on the early generations of Digital Subtraction Angiography, MRI, and Digital Tomosynthesis. For the last 25 years, Dr. Crawford has been leading PACS teams and academic radiology departments through transformative changes in technology and workflow. Just as PACS changed the radiologists’ fundamental work environment, Crawford joins other informatics experts in preparing for AI as the next transformative change of radiology practice.

As Crawford notes, integrating AI algorithms within day-to-day workflow will require careful planning for accuracy, safety, and security in new technical applications. Informatics experts will also need to ensure radiologist workflow isn’t interrupted by AI servers ingesting large data sets for routing and analysis. Where appropriate, user interfaces for AI output must be designed to allow evaluation of results without undue delays, display overload, or excessive mouse clicks.

The collective knowledge of our informatics team on theoretical and commercial solutions has shaped where the Department is headed in pre-clinical innovation and on-market AI technology. The three have identified major imaging areas where AI application is either already at work, or theorized for ML development and application:

Commercialized Machine Learning Technology

Mervak: Methods that position studies for faster read, pre-emptive case evaluation and critical findings identification (eg, intracranial hemorrhage).

Yap: First FDA-approved deep learning clinical platform, to provide automatic ventricle segmentation on cardiac MRI.

Theoretical Machine Learning Innovation

Reliability, Accuracy, Detection and Quality

Mervak: Task-based observations to evaluate image quality, with higher reliability than signal-noise ratio measurements in evaluating and tracking CT and MRI.

Yap:

  • Accuracy – Up to 11 CXR-detected different pathologies
  • Diagnosis – Congenital cataract disease (90%+ accuracy; Skin cancer sensitivity and specificity (90%+)
  • Prediction – Stroke treatment and cognitive deficiency outcomes; referable diabetic retinopathy (90%+ sensitivity and specificity)
  • Detection – Identifying abnormalities on knee MRI scans and cerebral aneurysms on CT angiograms

Improving patient experience

Mervak:

  • Advanced image reconstruction techniques allow reductions in radiation dose (CT) or noisy data (MR)
  • Reduced patient screening time (eg, pre-MRI)
  • Auto-scheduling MRI patients through ML predicted study time, for scanner time optimization
  • Real-time pathology recognition to inform patient repositioning, adding/removing MR sequences, or changing gadolinium contrast used.

Yap:

  • Decreasing radiation dose and MRI scanner time

Systems efficiency

Mervak:

  • Decreasing no-shows — missed care opportunities (MCOs). ML prediction of patients at high risk due to cost, transportation, communication and cultural barriers
  • Algorithms shifting passive patient evaluation (current state) to active monitoring (future state) to correct errors (eg, allergy-related issues with contrast media, contrast versus no-contrast protocolling)

Yap:

  • Incorporating ML into conventional workflows to improve outcomes
  • AI-powered triage system reducing time to treatment for urgent cases
  • Processing, aided reporting, follow-up planning, data storage and mining
  • Enhancing technical protocols reproducibility
  • Increased opportunity for value-added tasks (radiologist)

The Department of Radiology embraces what AI brings to imaging methodology as an evidence-based, analytics-driven science. As imaging ML methods reach FDA approval from conceptualization, we face the pleasant dilemma of determining which ones will most enhance our patient services at UNC and how to move closer to precision medicine through AI-powered technology.

Matthew A. Mauro, MD, FACR, FAHA, FSIR
Chair – UNC Department of Radiology
Ernest H. Wood Distinguished Professor of Radiology