Semi-automatic anatomy segmentation in thoracic dynamic MRI based on Segment Anything Model

Anatomy segmentation in thoracic dynamic MRI (dMRI) plays a crucial role in the quantitative analysis of regional respiratory structure and function in patients with thoracic disorders. Although several semi-automatic and automatic segmentation methods based on traditional image processing models and deep learning have been proposed with good performance, the accuracy of these methods are lower when processing thoracic dMRI scans using different acquisition protocols. In this paper, we present a novel semi-automatic anatomy segmentation approach based on a foundation model. In the first stage, we propose a corner-points and linear interpolation based interactive ROI detection strategy to extract the anatomic structure ROI from thoracic dMRI slices, which can reduce the negative influence of tissue spatially located far away from the anatomic structure of interest. In the second stage, we input the coordinate values of ROIs and the dMRI slices into the segment anything model (SAM) network to segment the anatomic structures of interest. The illustrative and quantitative results demonstrate that our approach achieves high accuracy and stability for single object and multi-object segmentation in thoracic dMRI.