RESUMO
Background: Efficiently and accurately detecting cerebral microbleeds (CMBs) is crucial for diagnosing dementia, stroke, and traumatic brain injury. Manual CMB detection, however, is time-consuming and error-prone. This study evaluates a novel artificial intelligence (AI) software designed for the automated detection of CMBs using susceptibility weighted imaging (SWI). Methods: The SWI data from 265 patients, 206 of whom had a history of stroke and others of whom presented a variety of other medical histories, including hypertension, diabetes, hyperlipidemia, cerebral hemorrhage, intracerebral vascular malformations, tumors, and inflammation, collected between January 2015 and December 2018, were analyzed. Two independent radiologists initially reviewed the images to identify and count the number of CMBs. Subsequently, the images were processed using an automatic CMB detection software. The generated reports were then reviewed by the radiologists. A final consensus between the two radiologists, obtained after a second review of the images, was used to compare results obtained from the initial manual detection and those of the automatic CMB detection software. The differences of detection sensitivity and precision for patients with or without CMBs and for individual CMBs between the radiologist and the automatic CMB detection software were compared using Pearson chi-squared tests. Results: A total of 1,738 CMBs were detected among 148 patients (71.4±10.7 years, 100 males) from the analyzed SWI data. While the radiologists identified 139 cases with CMBs, the automatic CMB detection software detected 145 cases. Nevertheless, there was no statistical difference in the sensitivity and specificity of the automatic CMB detection software compared to manual detection in determining patients with CMBs (P=0.656 and P=0.212, chi-square test). However, the radiologist identified 93 patients without CMBs, while the automatic CMB detection software detected 121 patients without CMBs, exhibiting a statistically significant difference (P=0.016, chi-square test). In terms of individual CMBs, the radiologists found 1,284, whereas the automatic CMB detection software detected 1,677 CMBs. The detection sensitivity for human versus the automatic CMB detection software were 75.5% and 96.5% respectively (P<0.001, chi-square test), while the precision rates were 92.2% and 86.0% (P<0.001, chi-square test), respectively. Notably, the radiologists were more likely to overlook CMBs when the number of CMBs was high (above 30). Conclusions: The automatic CMB detection software proved to be an effective tool for the detection and quantification of CMBs. It demonstrated higher sensitivity than the radiologists, especially in detecting minuscule CMBs and in cases with high CMB prevalence.