Artificial intelligence‐driven algorithm to detect and count cerebral microbleeds on Susceptibility Weighted Imaging

Abstract

AbstractBackgroundAmyloid‐related imaging abnormalities (ARIA)‐H, characterized by cerebral microbleeds (CMB) and superficial siderosis, is an important Magnetic Resonance Imaging(MRI) finding observed in the patients with Alzheimer’s disease, especially receiving the anti‐amyloid immunotherapy. Because the severity of CMB can potentially lead to discontinuation of the therapy, the detection and counting of the CMB would be critical tasks in MRI analysis. We present an algorithm which can detect CMBs and count the number of them, automatically and accurately.MethodWe employed datasets of both SWI and phase images, developed in collaboration with Eunpyeong St. Mary’s Hospital, Kyung Hee, and HanYang University Medical Center. The data set comprises 463 participants with CMBs (N = 247) and without CMBs (N = 116). Experts denoted gold standard of each CMBs as bounding boxes. A localizer is trained by selecting 2D axial slices, including expert‐confirmed annotations. A modified recurrent layer was applied for weight‐sharing between two images. An optimized region proposal network was then applied. To reduce FP regions across entire slices, we developed a classifier with attention‐based 3D convolution blocks. From the CMB localizer, candidate 3D CMB patches were used in the classifier.ResultCMB prediction was performed on the test set including 140 participants (CMB:24; Non‐CMB:116), with 263 expert‐confirmed gold standard (Figure 1). The lesion‐wise sensitivity was 0.913 [95% Confidence interval (CI), 0.863‐0.963] and the average FP rate was 0.672 per patient [95% CI, 0.290‐1.154]. Patient‐wise sensitivity and specificity were 0.862 [95% CI, 0.792‐0.932] and 0.907 [95% CI, 0.880‐0.934], respectively. When we divided the patients into four groups according to the number of gold standard, the model predicted the groups correctly by the sensitivity of 0.879 [95% CI, 0.851‐0.907] (Table 1).ConclusionOur model outperformed the results of recent studies by reducing FP rate with comparable sensitivity. Moreover, our model may suggest additional clinical benefit, because it can automatically count the number of CMB and classify the participants into 4 groups, used in the clinical investigation of anti‐amyloid therapy, according to the number of CMB. We expect this model may improve the workflow of the time‐consuming reading task and act as a failsafe for the detection of CMBs.