Abstract Archives of the RSNA, 2010
Antonio Criminisi, Abstract Co-Author: Employee, Microsoft Corporation, Cambridge, United Kingdom
Albert Montillo PhD, Presenter: Researcher, Microsoft Corporation
Jamie Shotton, Abstract Co-Author: Employee, Microsoft Corporation
Sayan Dev Pathak, Abstract Co-Author: Employee, Microsoft Corporation
John Winn PhD, Abstract Co-Author: Employee, Microsoft Corporation
Khan M. Siddiqui MD, Abstract Co-Author: Employee, Microsoft Corporation, Redmond, WA
Co-founder, iVirtuoso, Inc, Baltimore, MD
The growing use of MDCT scans has introduced the need for radiologists to read and interpret a huge number of images in their routine workflow. An efficient, fully automatic algorithm for anatomy segmentation is introduced, which promises to improve such workflow by enabling semantic navigation and quantitative image analysis at the touch of a button.
Decision forests are the basis of the algorithm, which associates each voxel with a probability of belonging to one of 13 anatomical structures: heart, liver, spleen, left lung, right lung, l. kidney, r. kidney, aorta, gall bladder, l. femur, r. femur, l. pelvis, r. pelvis. The algorithm is general and can be trained on a different set of structures.
Multiple decision trees are trained on a database of 40 voxel-wise annotated 3D CT scans from different patients and scanners, with varying resolutions, pathologies, cropping etc. (see figure). At run-time, the trees are applied to each voxel of a previously unseen scan and the relevant organs automatically recognized and segmented. Spatial relationships between structures are captured via new, efficient visual features which build upon state-of-the-art computer vision research.
The labelled database (45 scans) has been split into training (15 scans) and testing subsets. Quantitative results are computed on the testing subset only to avoid over-fitting. A voxel-wise recognition accuracy of 93% (when compared to expert segmentations) is achieved in approx. 1 sec. on a 512 x 512 x 400 scan, on a standard quad-core desktop machine. Comparison with state of the art SVM, GMM and atlas-based algorithms show superior performance for our technique.
A new algorithm for the automatic segmentation of anatomy in CT scans is presented. Shape and context modelling, together with the use of multiple trees yield high accuracy on previously unseen datasets. High computational efficiency is achieved thanks to the simplicity of the visual processing and the algorithm's parallel nature. No manual intervention is ever required.
Applications include: 1) Automatic estimation of organs’ volume and density; 2) Semantic image registration; 3) Organ-driven image navigation (see fig.); 4) Automatic tagging of patients’ scans to aid their search and retrieval.
Automated anatomy segmentation unlocks quantitative information hidden within CT images and speeds-up radiology workflow.
Automatic Semantic Segmentation of Anatomical Structures in CT Scans. Radiological Society of North America 2010 Scientific Assembly and Annual Meeting, November 28 - December 3, 2010 ,Chicago IL. http://archive.rsna.org/2010/9003197.html