End-to-End Coordinate Regression Model with Attention-Guided Mechanism for Landmark Localization in 3D Medical Images

摘要

In this paper, we propose a deep learning based framework for accurate anatomical landmark localization in 3D medical volumes. An end-to-end coordinate regression model with attention-guided mechanism was designed for landmark detection, which combines global landmark configuration with local high-resolution feature responses. This framework regress multiple landmarks coordinates for landmark localization directly, instead of the traditional heat-maps regression. Global stage informs spatial information on the coarse low resolution images to regress landmarks attention, which improve landmarks localization accuracy in the local stage. We have evaluated the proposed framework on our Temporo-mandibular Joints (TMJs) dataset with 102 image subjects. With less computation and manually tuning, the proposed framework achieves state-of-the-art results.