Automatic Whole Heart Segmentation Using a Two-Stage U-Net Framework and an Adaptive Threshold Window

Whole heart segmentation is an important medical imaging method used to enable clinical applications. However, automatic segmentation of the heart is still a challenging task due to the complexity and particularity of medical images, especially when the heart is segmented into substructures. In this study, we present a training strategy that relies on a two-stage U-Net framework and an adaptive threshold window to automatically segment a whole heart and its substructures. The two-stage U-Net framework consists of a region of interest (ROI) detection of the whole heart and accurate segmentation of the heart substructures. The adaptive threshold window method is used to remove the noisy parts of the data while preserving the anatomical relationships between local regions. Experiments were performed on a dataset from the MM-WHS Challenge 2017. The proposed approach resulted in a high segmentation accuracy with a 79.3% and 95.5% Dice similarity coefficient for the whole heart and ascending aorta segmentation, respectively, using limited GPU computing resources and small amounts of annotated data. The full implementation and configuration files in this paper are available at https://github.com/liut969/Whole-Heart-Segmentation.