Automated Detection of Cerebral Microbleeds in Susceptibility-Weighted MRI Using a Multi-Scale Deep Learning Framework
Abstract
Cerebral microbleeds (CMBs) are small hemosiderin deposits associated with cerebral small vessel disease and are considered important biomarkers for stroke risk, dementia, and neurodegenerative disorders. Manual identification of CMBs in susceptibility-weighted imaging (SWI) or T2*-weighted MRI is time-consuming and prone to interobserver variability. This paper introduces a ground-breaking deep learning model that can automatically detect microbleeds from MRI data. We engineered a model that combines multiresolution convolutional neural networks with attention-based refinement of features to improve our sensitivity to small vascular pagiotic lesions and drive down the rate of false positives attributed to vessels and calcified regions of the brain. The experimental results obtained from large open-source datasets demonstrate that the average sensitivity of our system achieved 0.86, specificity 0.91, precision 0.83, and F1 score 0.84 with an area under ROC curve (AUC) 0.92, thereby outperforming traditional machine learning algorithms as well as baseline convolutional neural networks by 6-10% in detection accuracy. Moreover, our method produced approximately an 18% decrease in the rate of false positives when compared to patch-based CNNs. The results demonstrate that our model can serve as a reliable computer-assisted diagnosis tool to aid radiologists in performing their duties to the patient.