Deep learning-based covert brain infarct detection from multiple MRI sequences

Abstract

Automatic and accurate detection of covert brain infarcts may help identify individuals at risk of cognitive decline, dementia, and vascular events who could be eligible for early preventive measures or enrollment in clinical trials. We propose a novel deep learning-based framework to detect covert brain infarcts from multiple MRI sequences, including T1-weighted and T2-weighted fluid attenuated inversion recovery (FLAIR) scans. First, we design a simple yet effective cross-sequence registration method to register T1 and FLAIR by slice-level and pixel-level alignment. The accurate registration enables different sequences to share the infarct annotations. Second, we employ a fully convolutional one-stage object detector for each sequence to obtain infarct candidates. The exploitation of the contextual information of adjacent slices and the elimination of predefining anchor boxes and proposals can achieve high sensitivity with computational efficiency. Finally, we propose a multi-sequence fusion strategy with attention mechanisms to jointly combine different sequences so that their complementary representation can be explored to reduce false positives. The attention mechanisms are designed to consider the importance of different sequences, spatial locations, and channels. To evaluate the effectiveness of the proposed method, we construct a novel dataset with 264 cases and 738 brain infarcts with pixel-level annotations. Extensive experiments are conducted on this dataset and the results demonstrate that our method achieves state-of-the-art infarct detection performance with a sensitivity of over 80% and less than 5 false positives per case.