A new deep-learning-based artificial intelligence system can identify a variety of acute neurological disorders in patient CT scans within seconds. The tool might enable neurologists to rapidly identify patients who need urgent attention and prioritize diagnosis and treatment in these individuals.

Swift diagnosis and treatment of acute neurological diseases such as stroke is crucial to minimize progressive neurological damage. Computerized analysis of patient brain images has the potential to help clinicians to diagnose and treat patients more rapidly than by human judgement alone.

“I was motivated by my experiences of taking care of patients with acute neurological illnesses, in which any possible way of reducing the time it took for us to reach these individuals could have improved their outcomes,” explains corresponding author Eric Oermann. “One day, Joseph Titano and I said to ourselves that we needed to solve this problem with deep learning.”

The researchers used a database of over 37,000 head CT scans featuring a variety of neurological findings. An artificial intelligence system known as a 3D convolutional neural network was trained to identify whether or not scans contained “critical findings” signifying the presence of acute neurological disease such as stroke, haemorrhage or hydrocephalus.

Compared with human judgement, the computerized system could correctly detect critical neurological findings with the same sensitivity (79%) but a lower specificity (48% versus 85%). From these results, Oermann and colleagues predicted that their system could be used to prioritize patients who were most likely to require urgent intervention, prompting physicians to perform rapid diagnosis in these individuals.

To test this idea, the team conducted a randomized controlled trial in a simulated clinical environment to compare the abilities of humans and the automated system to triage brain scans. The automated system processed each image in an average of 1.2 s — 150 times more quickly than the human average. Importantly, in the computer-prioritized list, scans of individuals requiring urgent attention were assigned a significantly higher queue position than scans of individuals with routine findings.

The endgame is to help patients, and that means translating academic results into actual tools that contribute to patient care

The team now hope to develop and test their tools in a real clinical setting. “The endgame is to help patients, and that means translating academic results into actual tools that contribute to patient care,” comments Oermann.