Hybrid attention densely connected ensemble framework for lesion segmentation from magnetic resonance images

Abstract

White matter hyperintensities (WMHs) are associated with various neurological and aging diseases, and morphological analysis plays a crucial role in the assessment of disease progression. In this article, a novel hybrid attention densely connected ensemble framework is deployed for WMH segmentation from multi-modality magnetic resonance imaging (MRI). On the one hand, hybrid attention densely convolutional network (HA-DCN) is designed with a novel hybrid attention module embedded. The hybrid attention module can further improve the precision of lesion localization by extracting the complementary information of high-level features and low-level features from the spatial domain and channel domain. On the other hand, the focal Tversky loss function and generalized dice loss function derived from dice similarity coefficient are ensembled into the proposed framework, which achieves a trade-off between specificity and sensitivity. As a result, the volume of automation lesion segmentation is more agreeable to the manual lesion segmentation by experts. The proposed framework was evaluated online and offline in the MICCAI 2017 WMH segmentation challenge. A quantitative experiment has further demonstrated the effect of multi-modality and the effectiveness of the proposed hybrid attention densely connected ensemble framework. Furthermore, the challenge dataset consists of three scanners, reflecting the flexibility and robustness of the model. It also exhibits its potential for real-world clinical practice.