By using pattern recognition technology, doctors may now be able to read deeper into computerized tomography (CT) brain scan imagery than ever before. For stroke patients, this new depth of understanding could inform critical treatment decisions.

Ischemic stroke, the most common type of stroke in humans, occurs when small clots interrupt blood flow to the brain. A highly successful treatment is available, but it carries with it the potential for deadly side effects. Intravenous thrombolysis (tPA) triggers symptomatic intracranial hemorrhaging (SICH) in six percent of stroke patients who take it.

Thrombolysis works by injecting a chemical into the patient's blood vessels to break up the clots and allow blood to flow freely again. It thins the blood, which is why it sometimes causes hemorrhage. Scientists look to CT brain scans to determine which patients are at risk for hemorrhaging, but the overall accuracy of their predictions (60-70 percent) may indicate that the information in CT scans is too complex for the naked eye to adequately interpret.
Reading between the voxels

Building off the technology airports use to catch international criminals, researchers at the Imperial College London set out to catch the SICH foreshadowed in the CT scans of stroke patients. In their study, published in the journal NeuroImage Clinical, they took the scans of 116 stroke patients who had been treated with tPA, (16 of whom consequently suffered SICH), and were able to retroactively predict the occurrence of SICH with 74 percent accuracy. Comparatively, experts were asked to evaluate the same 116 scans using the standard prognostic approach and achieved 63 percent success. In a second test they determined which one patient out of ten would suffer SICH from tPA with 56 percent accuracy, versus 31 percent accuracy for the standard prognostic approach. While these may seem like modest success ratios, they represent significant improvements.

The standard prognostic approach for determining a patient's tPA candidacy quantifies at least 12 variables (severity of stroke and subject's age are among them). The software used in the study, on the other hand, systematically interprets a vastly larger amount of information. By evaluating the entire optimized CT brain scan image, it processes 181,311 voxels (pixels with three dimensional value) of data per subject.

Researchers hope the accuracy will only improve as the software accumulates more data from which to discern patterns.

More Industry Headlines