Multi-scale convolutional neural network with enhanced feature fusion for voxel-level detection of cerebral microbleeds

Cerebral microbleeds (CMBs) are small areas within the brain, that appear as dark spots on MRI scans of the brain, Gephart et al. (2022) (Fig. 1) and can provide multiple neurological insights for numerous conditions affecting cognition, mood, and function. Identification of CMBs is challenging due to their small size, low severity, and potential confusion with nearby blood vessels or MRI artifacts. This paper presents a system called Multi-Scale Deep Convolutional Neural Network (MS- DCNN) that employs a multi-layered deep neural network to automatically identify CMBs within neurological magnetic resonance imaging (MRI). The architecture of our neural network will aggregate features of different "brain sizes" (for example: 3×3, 5×5, and 7×7) which allows for the identification of CMBs in diverse areas of the brain. Furthermore, to increase performance we add an optional attention mechanism which enables the examination of the refined feature maps and allows the model to favour regions of clinical importance. Heat maps show identified regions that indicate clinically relevant areas. We developed numerous preprocessed items that serve as essential components in our model. These include skull stripping, intensity normalization, contrast enhancement (e.g., CLAHE), and a final examination which emphasizes the low penetration and commonly low visibility of cerebral microbleeds. Furthermore, to improve the model performance we incorporated distinct augmentation (e.g., rotation, flipping, and noise) to increase generalization. The model is trained as a binary cross-entropy loss function incorporating balanced class representations to account for a highly imbalanced situation. We evaluate our model on a publicly available annotated dataset, achieving strong performance metrics: Accuracy: 96.2%, Sensitivity: 93.8%. Importantly, the false positive rate per scan is reduced to 1.7, significantly outperforming baseline CNN and U-Net architectures. These results demonstrate that the proposed MS-DCNN provides a highly accurate, robust, and interpretable framework for assisting radiologists in the early detection of cerebral microbleeds.