A thoracic CT volume is an input to the algorithm. The first step is rough segmentation of the lung, and the second step is body cavity segmentation. Third, a location of lung is roughly estimated, whose results are used to generate the patient- specific shape. Finally, the multi-shape graph cuts with neighbor constraints and an adaptive weight refine the segmentation results. The difference from the algorithm [185] is the third step, or rough location estimation of lung, and the final step, or fine lung segmentation with the proposed adaptive weight.

Rough Location Estimation of the Lung

This step estimates top and bottom axial slices of the lung, ztop and zbottom, where the z axis is parallel to the craniocaudal axis. Subsequently, landmarks are automatically defined according to the extracted axial slices and bounding box of a body cavity extracted in the previous step. The landmarks are forwarded to a radial basis function-based nonlinear registration between the SSM to an input volume, and the registration results are used to generate the patient-specific shape. In the previous method [185], ztop and zbottom slices were determined based on the result of a CT value-dependent rough segmentation only, which frequently failed in segmentation caused by severe pathologies and/or postmortem changes, in particular zbottom. As a result, the estimated positions of zbottom slices would greatly deviate from the true positions. In this study, a linear predictor of zbottom slice was presented based on the z coordinate of the top axial slice of the liver segmentation result [182] explained in the previous subsection. The details can be found in the paper [186].

< Prev   CONTENTS   Source   Next >