CV-AttentionUNet: Attention based UNet for 3D Cerebrovascular Segmentation of Enhanced TOF-MRA Images

Abstract

Cerebrovascular segmentation is quite a challenging task and has a crucial role in the diagnosis of cerebrovascular diseases. Due to lack of automated methods, the time of flight magnetic resonance angiography (TOF-MRA) is assessed visually, which is time consuming. For time-critical diagnosis using non-contrast TOF-MRA images, current methods are not end-to-end. Popularly used encoder-decoder architecture utilizes redundant features which eventually leads to extraction of low-level features multiple times while CNNs suffers from performance degradation when batch size is small with vanishing gradient problem occurring for deeper networks. In this paper we attempted to solve these limitations and proposed the 3D cerebrovascular attention UNet method, named as CV-AttentionUNet, for precise extraction of brain vessels. This approach first includes the vessel enhancement method for contrast adjustment of small vessels and uses 3D-UNet as a base architecture which extracts the features at multiple scales. Additionally, for combining low semantics to high semantics, we proposed the attention mechanism. This mechanism not only focuses on relevant associations and neglects irrelevant information, but also consider the small batch size problem by induction of Group Normalization (GN). Furthermore, inclusion of deep supervision incorporates different level of features proving beneficial for network convergence. Moreover, we provided a semi-automatic cerebrovascular segmentation labelling method dealing with the lack of dataset. We evaluated our framework on TubeTK TOF-MRA, where unlabeled images are labelled separately, using proposed method. The results indicate that our method performed better as compared to existing state-of-the-art methods on TubeTK dataset.