American scientists have developed an algorithm that indicates aneurysm in the brain. The model based on the work of the convolutional neural network marks the possible place of the aneurysm on the image of the brain obtained using CT angiography, which greatly simplifies the diagnosis to radiologists. As reported in an article published in the JAMA Network Open, the use of a new automatic diagnostic method greatly improves the accuracy of the diagnostician.
The presence of an aneurysm (dilatation of the blood vessel) in the brain is a rather dangerous condition: its rupture can lead to hemorrhage, as a result of which various neurological disorders or even death can occur. The most effective way to prevent such consequences is early diagnosis and subsequent treatment to prevent rupture.
Computed tomographic angiography (CT angiography) is now used as one of the main diagnostic methods, which allows you to accurately visualize the blood vessels and assess the nature of the blood flow from a three-dimensional image of the necessary part of the body. However, aneurysms can be very small and can be difficult to examine.
To improve the diagnosis of aneurysm on CT angiography using automatic methods, researchers decided under the leadership of Allison Park (Stanford University). The HeadXNet algorithm they used is based on the analysis of three-dimensional images using convolutional neural networks. For training, researchers took images and results of 611 diagnostics: on the scans used there were both diagnosed aneurysms and their absence. The resulting model highlights the likely location of the aneurysm in one of the sections of the image in red.
After that, the resulting model was tested on 115 unpartitioned images and showed them to eight qualified radiologists (the experience of similar diagnostics of specialists ranged from 2 to 12 years) along with standard unmarked results of CT angiography. The accuracy of diagnostics when using the new algorithm significantly (p = 0.01) has increased in comparison with the usual diagnostics. At the same time, however, there were no significant changes in the time that diagnosticians spent on image analysis.
Despite promising results and high accuracy, the authors of the work clarify that it is not yet possible to use the new algorithm as the only diagnostic method. Any such automatic method should be accompanied by an additional assessment by an experienced radiologist – even if it is finalized.
Most automatic methods for diagnosing diseases are now being developed precisely in order to simplify the work of medical personnel.