Multi-Task Learning for False-Positive Reduction and Segmentation of Cerebral Aneurysms in CTA Scans

Abstract

The computer-aided diagnosis for cerebral aneurysms consists of three major steps, which are lesion detection, false-positive reduction, and segmentation. Many methods based on deep learning technology have been designed for each of these tasks separately, without the shared information to further collaborate these models with each other, and therefore limit their further performance improvements. In this paper, we propose a novel framework to perform false positive reduction and aneurysm segmentation jointly in a multi-task manner. In this way, both false-positive reduction and segmentation networks can mutually share information between each other and facilitate together. We also incorporate the vessel segmentation information in the framework, which can provide important priors for false-positive reduction and segmentation. The proposed network is evaluated on a public dataset of cerebral aneurysms. Experimental results show that our vessel-guided multi-task model can achieve improved performance than separately training the false positive reduction and segmentation models for single tasks.