3D Multi-Scale Residual Network Toward Lacunar Infarcts Identification from MR Images with Minimal User Intervention

Abstract

Lacunes or lacunar infarcts are small fluid-filled cavities associated with cerebral small vessel disease (cSVD). They contribute to the development of lacunar stroke, dementia, and gait impairment. The identification of lacunes is of great significance in elucidating the pathophysiological mechanism of cSVD. This paper proposes a semi-automated 3D multi-scale residual convolutional network (3D ResNet) for lacunar infarcts detection, which can learn global representations of the anatomical location of lacunes using two multi-scale magnetic resonance image modalities. This process requires minimal user intervention by passing the potential suspicious lacunes into the network. The proposed network is trained, validated, and tested using five-fold cross-validation using data, including 696 lacunes, from 288 subjects. We also present experiments on various combinations of multi-scale inputs and their effect on extracting global context features that directly influence identification performance. The proposed system shows its capability to differentiate between true lacunes and lacune mimics, providing supportive interpretations for neuroradiologists. The proposed 3D multi-scale ResNet identifies lacunar infarcts with a sensitivity of 96.41%, a specificity of 90.92%, an overall accuracy of 93.67%, and an area under the receiver operator characteristic curve (AUC) of 93.67% over all fold tests. The proposed system also achieved a precision of 91.40% and an average number of FPs per subject of 1.32. The system may be feasible for clinical use by supporting decision-making for lacunar infarct detection.