Computed tomography (CT) has been widely used for treatment planning in radiotherapy because of its ability to reveal anatomic structures in an accurate and noninvasive way. Accurate and consistent delineation of the heart and breast in CT images is critical for breast (lung) cancer radiotherapy. Currently, this is done manually by physicians in a slice-by-slice manner. The procedure is not only time-consuming but also inconsistent among different physicians. Automatic segmentation is increasingly important as the resolution of the CT scanner becomes higher and the amount of data become larger.
Although CT image usually has high resolution and signal to noise ratio, fully automatic segmentation of the heart is still a challenging problem due to similar intensity of neighboring organs and patient-specific variations. Also, the respiration motion and heart beating lead to blurry boundary, especially when non-gated imaging is used. Because of those difficulties, the use of the prior knowledge has been extensively investigated for medical image segmentation problem.
In this study, we make use of manual segmented training images to build shape and appearance model of the target we try to segment (e.g. breast and heart). The learned shape and appearance model is used to guide segmentation for testing images. The segmentation results are evaluated by Dice metric and mean distance error compared with manual segmentation. The proposed automatic segmentation methods together with automatic beam placement method are essential components towards fully automated treatment planning for breast radiotherapy.
This project was supported in part by Varian project Grant no. 15-C6000-33560-33570.