Automatic segmentation of the human spinal canal using an intelligent digital atlas

Abstract

To protect the spinal cord from radiotoxicity, it is necessary to carefully delineate the spinal canal when delivering therapeutic radiation to any target contained in the thorax, abdomen, or pelvis. The current clinical practice of manually segmenting the spinal cord on selected images and applying interleaved extrapolations to the entire spinal cord is not always accurate and frequently requires manual adjustment. We previously demonstrated the feasibility and validation of a fully automatic atlas-based segmentation system to the brain in patients with intracranial tumors [1D’Haese P. et al.Int J Radiat Oncol Biol Phys. 2003; 57: S205Abstract Full Text Full Text PDF Google Scholar]. Here we introduce an innovative method to contour the spinal canal with complete automaticity. The purpose herein is to determine the feasibility of this system for treatment planning in CT images of oncology patients undergoing radiotherapy. Previous atlas-based, non-rigid registration methods have been readily applied to anatomic structures regular in shape and characteristically stationary. The brain has worked well for these algorithms due to its immobility in the cranial compartment and its relatively uniform shape. To employ automatic segmentation techniques to wider anatomical regions, we have innovated a method that accounts for local variations in plasticity and rigidity. Central to this system was to develop an intelligent atlas that can be trained to know which anatomic regions are relatively pliable and consequently liable to change shape (lungs, muscle, adipose tissue) and which anatomic regions are relatively rigid and stationary (fibrous tissue, bone). Thus, the deformation algorithm automatically adjusts the stiffness of the transformation to permit larger displacements over compliant regions and smaller displacements over less compliant regions [2Duay V. et al.Int Symp on Biomed Imaging. 2004; Google Scholar]. From a single CT image set, we created a digital atlas by manually contouring the cervical-, thoracic-, and lumbar spinal column and subsequently training the atlas for regional pliability. We then employed this completely automatic segmentation system to deform the atlas and provide contours of the entire spinal canals in CT images of 3 patients being planned for radiotherapy. The resulting contours were reviewed by a radiologist and radiation oncologist. The automatic segmentation system was easily applied to the 3 patients in this study. The figure shows representative CT images of the spinal column sagittally (panel A) and axially (panel B) with automatically-generated contours shown in dotted lines. The contours were deemed extremely satisfactory by the reviewing radiologist and radiation oncologist and were subjectively ascertained to require minimal, if any, adjustment for treatment planning. This novel automatic segmentation method proved effective for contouring the spinal canal in a small group of radiation oncology patients. Having a robust, reliable system to automatically and efficiently segment the spine is especially significant, considering the advent of real-time radiation planning and the importance in monitoring dose delivery. These results impart broader implications for utilizing automatic segmentation in the regions of the thorax and abdomen, where the anatomical architecture displays varying levels of plasticity.