How Is AI Being Used in Daily Neuroradiology Practice?
Experts explore potential benefits and challenges involved in integrating AI into diagnostic and treatment workflows
AI tools are beginning to impact daily neuroradiology practice, with real implications for clinical workflow, diagnosis and patient care. Experts emphasize that understanding both the potential and limitations of these tools is essential for safe, effective and responsible adoption in daily practice.
“AI has the potential to impact basically every stage in the radiology imaging cycle,” said Jason Talbott, MD, PhD, professor of clinical radiology and biomedical imaging at the University of California, San Francisco, and Zuckerberg San Francisco General Hospital. “As these technologies continue to integrate into our workflow, it’s critical that we stay informed and prepared to evaluate and adopt them responsibly.”
Specifically, AI has been increasingly used in the field of neuroradiology to evaluate acute neurological changes, improve neuroradiologist efficiency, and accelerate imaging of the brain and spine.
“We are at the frontier of AI in medicine, where its integration into clinical practice is no longer a future prospect but an evolving reality,” said Amish Doshi, MD, professor of radiology and neurosurgery, division chief of neuroradiology and vice chair of radiology ambulatory services at the Icahn School of Medicine at Mount Sinai in New York City. “As AI continues to gain traction, radiologists are discovering its potential as a powerful tool to enhance patient care.”
Automating Radiology Reports
One of the most talked-about AI applications in radiology is the use of large language models to support report generation. These tools can generate human-like text and have the potential to, at least partially, automate neuroradiology report generation, according to Dr. Talbott who presented on this topic at an educational session during RSNA 2024.
He pointed to a recent Radiology study in which researchers used Generative Pre-trained Transformer 4 (GPT-4), a natural language processing model, to convert fictitious free-text radiology reports into structured templates.
“GPT-4 was very effective for post hoc, standardized, structured report generation,” Dr. Talbott said. “It was highly scalable, and there’s potential for structuring vast amounts of radiology data into more structured formats.”
The use of GPT-4 presents some potential patient data privacy concerns, however. Dr. Talbott also cautioned about false statements made confidently by LLMs, referred to as “hallucinations.” Other hurdles include biases in training data and integration challenges with radiology systems.
AI In Brain Tumor Imaging
While the potential of AI in neuroradiology is broad, its real-world adoption—particularly in specialized areas like brain tumor imaging—remains limited.
“Despite thousands of studies on AI in neuroradiology, only about 126 FDA-cleared products exist, and few of them relate to brain tumor imaging,” said Mariam S. Aboian, MD, PhD, an attending radiologist at Children’s Hospital of Philadelphia.
As Dr. Aboian noted, one possible area where AI tools could be helpful is volumetrics, measuring whole tumor volume and assessing it over time. She also emphasized that radiologists should consider how any of their AI tools would integrate with other specialties.
Another important consideration is generalizability. “If you develop an AI that works in your hospital and maybe three other hospitals through an external validation data set, does it mean it's going to work in the entire U.S.?” Dr. Aboian asked. “Does it mean it's going to work in the entire world? Most likely not. So, you really want to evaluate the model generalizability.”
To help assess model accuracy, Dr. Aboian suggested looking at metrics such as the Dice coefficient and Hausdorff distance. She also urged radiologists to consider the impact of a given algorithm on workflow efficiency and medical decision making.
AI Triage in Brain and Spine Imaging
Beyond diagnosis and reporting, AI is making strides in triaging radiologic images via computer-aided diagnosis tools (CADt), Dr. Doshi noted. It can also assist with diagnosing conditions such as brain hemorrhage, stroke, aneurysms and spinal fractures.
“A commonly used AI algorithm can detect any hemorrhage in the brain,” Dr. Doshi said. “In general, these algorithms have shown pretty high sensitivity and specificity.”
He noted there are several commercially available triage algorithms for detection of brain aneurysms and vessel occlusion in stroke on CT angiography scans of the brain. Additionally, Dr. Doshi noted the use of AI in assessing fractures on CT scans of the cervical spine.
“Reported sensitivities for brain and spine AI triage algorithms range from 88% to 95%,” he said.
Dr. Doshi referred to the results of a paper on the winning algorithms of the RSNA 2022 Cervical Spine Fracture Detection Challenge. “Overall, the authors concluded that AI models can detect and localize cervical spine fracture on CT scans with high performance outcomes,” he said.
Highlighting the challenges associated with AI technology, including performance and generalizability, Dr. Doshi said, “We must evaluate this technology in our practice to understand the benefits and challenges of each of these AI algorithms. It will take time, but improvements in AI and our understanding of how best to use it, will drive greater use down the line.”
Deep Learning for Image Reconstruction
AI is also enhancing image quality and efficiency through advanced reconstruction techniques that improve both CT and MR imaging, according to Lawrence N. Tanenbaum, MD, former vice president, chief technology officer and director of advanced imaging at RadNet, Inc.
Deep learning reconstruction (DLR) for CT is powerful enough that a radiologist can look at a one-millimeter cut that has the noise behavior of a five-millimeter cut using older methods, Dr. Tanenbaum noted.
“If you are knowledgeable about CT, you know that there were complaints about iterative reconstruction—some said the texture of the image changed,” he said. “Deep learning doesn't have that problem. It's very smooth, crisp and sharp.”
In MR, DLR provides a strong benefit by reducing scan time while maintaining quantitative integrity, Dr. Tanenbaum said.
“Realize that if you buy software and just turn it on, you don't get any additional speed,” he noted. “You have to do something to accelerate the scan. One of the things you can do is reduce the number of excitations or acquisitions. Of course, in doing so, you sacrifice your signal-to-noise ratio. Using DLR restores it.”
“You can also push the parallel imaging to the point where noise becomes problematic, and DLR will fix that,” he continued. “Some of the more sophisticated algorithms actually correct aliasing, where parts of the image appear in the wrong location or get overlapped, permitting even greater acceleration.”
The increased speed that comes with AI has an obvious impact on workflow and patient comfort as well as economic benefits to the facilities, Dr. Tanenbaum noted.
As AI continues to evolve across multiple fronts in neuroradiology, clinicians are urged to take a measured, evidenced-based approach when deciding which tools to adopt.
“It’s important to understand the limitations of these technologies, assess what your needs are and trial some of these tools to see if the investment—which is not insignificant—is actually worth the benefit you're expecting in real practice,” Dr. Talbott concluded.
For More Information
Access the Radiology article, “Leveraging GPT-4 for Post Hoc Transformation of Free-text Radiology Reports into Structured Reporting: A Multilingual Feasibility Study.”
Access the Radiology: Artificial Intelligence article, “Performance of the Winning Algorithms of the RSNA 2022 Cervical Spine Fracture Detection Challenge.”
Access the AppliedRadiology article, “Deep Learning: Promising to Revolutionize Image Reconstruction.”
Read previous RSNA News stories about neuroradiology: