(Thursday, 8/2/2018) 10:00 AM - 12:00 PM
Room: Karl Dean Ballroom C
Purpose: Auto-segmentation in low contrast regions is challenging due to the limited image gradient. This study aims at developing deep inverse shape encoding (DISE) networks to transfer organ contours from prior CT images to low contrast regions in CT or CBCT.
Methods: For the CT-to-CT case, each image renders an attenuation map f with attenuation constancy assumed between the two images. Forsaking reliance on noisy gradients, we use the robust counterpart, invShape(Na) = {x|f(x) in Na}, the inverse shape of a small range Na. Such set gives a definitive shape to, and holistic distribution of, the voxels in the same range and adjacent in the gradient field. The image domain is dissected into non-overlapping inverse shapes induced by a range partition. The first DISE layer specifies one or more range partitions, shared by both images. The next layers encode inverse shapes, with an integral scheme for sparse sampling, sparse representation of shape signatures, spatial geometries and certain statistics, and for fast and robust mappings of corresponding inverse shapes between the two images. The contour transfer results from the mappings. For the CT-to-CBCT case, we map the transformed attenuation functions in the cone-beam projection space for certain advantages, despite the loss of attenuation constancy. Two or more inverse shape configurations must be used for mutual evaluation and iterative correction. Finally, only the contour mappings are back-projected, robustly, to the image space.
Results: The organ contours are successfully transferred between anthropomorphic phantom XCAT-CT images at two respiratory phases. The warped contours agree well with the ground truth, with Dice coefficient greater than 0.97.
Conclusion: Contour transfer with DISE networks promises an efficient, robust auto-segmentation method in low contrast regions. Moreover, the method is extendable to cross-modality contour transferring via a common space where both attenuation functions can be represented in non-quality-degrading transformations.
Funding Support, Disclosures, and Conflict of Interest: This work is supported by NIH grant R01 CA-184173.
Keywords
Segmentation, Nonlinear Image Warping, Geometric Transformation
Taxonomy
IM/TH- image segmentation: Multi-modality segmentation
Contact Email
