Could A Deep Learning Algorithm Be A Better Neurological Diagnostic Tool?

The CT scan is a common technique for assessing internal physiological damage across the world.  The head CT scan, specifically, helps to diagnose emergencies as well as detect acute brain hemorrhages.  Of course, this type of diagnostic imaging—3D stack of greyscale images notarized by poor soft-tissue contrast, a low signal-to-noise ration, and a high artifact incidence— requires highly-trained expertise.  And that means there is also a high risk for overlooking subtle abnormalities, which can sometimes be life-threatening. 

In order to increase the efficiency of these results—and, hopefully, accuracy as well—University of California-San Francisco scientists developed a fully complex [digital] neural network—called PatchFCN—which can potentially identify these head CT scan abnormalities.  This technique enables physicians to more closely examine the scan results and make a more accurate therapy recommendation. 

Study co-author Esther Yuh explains, “We wanted something that was practical, and for this technology to be useful clinically, the accuracy level needs to be close to perfect.  The performance bar is high for this application, due to the potential consequences of a missed abnormality, and people won’t tolerate less than human performance or accuracy.”

In this project, the researchers trained the AI neural network with nearly 4400 head CT scans that were collected at UCSF.  With this information uploaded, the team then used the algorithm to interpret another independent set of 200 scans and then compared their performance with that of four US board-certified [human] radiologists.  ON average, the algorithm took only one, single second to evaluate each stack of these images and more importantly, perform abnormality delineation at the pixel-level. 

The PatchFCN technique, then, demonstrated the highest accuracy ever for this type of clinical application.  In other words, the technique exceeded the performance of two of the four radiologists (so, at least 50 percent better), in some instances identifying very small subtleties that some had missed in their human analysis.  

CT interpretation is a core skill for radiology training, with the most skilled readers demonstrating 95 percent accuracy.  The PatchFCN tool, then, could be a suitable screening tool with a mostly human ability as it exhibited 100 percent sensitivity at some specificity levels and 90 percent at others. 

Reply