Please note: Links to movies, ppt slideshows and any other multimedia files are ... 2.jpg. For automatic 3D registration of kidney, several training data sets were ...
A New 3D Hybrid Algorithm for Automatic Segmentation of Visceral Adipose Tissue e-Poster: 37 Congress: esmrmb2015 2015 Type: Scientific Paper Topic: Preclinical Studies and Basic Science / Processing and quantification Authors: F. Fallah1, F. Schick2, B. Yang1; 1Stuttgart/DE, 2Tuebingen/DE MeSH: Abdomen [A01.047] Body Weights and Measures [E05.118] Diagnosis, Computer-Assisted [E01.158] Computing Methodologies [L01.224] Keywords: 3D Registration of Kidney, Hybrid Algorithm, 3D Automatic Segmentation, Visceral Adipose Tissue
Any information contained in this pdf file is automatically generated from digital material submitted to e-Poster by third parties in the form of scientific presentations. References to any names, marks, products, or services of third parties or hypertext links to third-party sites or information are provided solely as a convenience to you and do not in any way constitute or imply ESMRMB’s endorsement, sponsorship or recommendation of the third party, information, product, or service. ESMRMB is not responsible for the content of these pages and does not make any representations regarding the content or accuracy of material in this file. As per copyright regulations, any unauthorised use of the material or parts thereof as well as commercial reproduction or multiple distribution by any traditional or electronically based reproduction/publication method is strictly prohibited. You agree to defend, indemnify, and hold ESMRMB harmless from and against any and all claims, damages, costs, and expenses, including attorneys’ fees, arising from or related to your use of these pages. Please note: Links to movies, ppt slideshows and any other multimedia files are not available in the pdf version of presentations. www.esmrmb.org
1. Purpose For investigation of obesity related disorders, accurate volumetry of visceral adipose tissues (VAT) is of great significance. Differentiation of VAT from other fat compartments using subtle differences in chemical composition, demands long echo acquisition. This hinders breath-hold imaging and is prone to variations among subjects. Therefore, we aimed to establish a fully automatic method for accurate VAT volumetry, based on known spatial distribution of VAT. This extends the previous segmentation of T1-weighted 2D axial images [1, 2] to full 3D volumetry, applicable to both T1-weighted and adipose only MRI data. The latter could be generated by chemical shift decomposition (e.g. Dixon or IDEAL methods). In this work, for automatic segmentation of VAT, we also aimed to localize and segment abdominal region without need to whole body or whole trunk scans as what was done in previous methods [3, 4, 5]. 2. Material and Methods 3D T1-weighted and adipose-only data were acquired in axial direction from abdomen of 15 healthy volunteers with BMI of 18-42 kg/m2, by a 3D Dixon VIBE sequence, on a 3T Magnetom Skyra (Siemens Healthcare). In a previous method [3], the abdominal region was supposed to be between heads of femurs and diaphragm; where diaphragm was localized by a fraction of the calculated distance between heads of femurs and humeri. This approach needed coverage of the entire trunk in the acquired MRI data, and also additional steps for exclusion of subcutaneous adipose tissue (SAT) using body mask and Snake algorithm. Hence it was restricted to 2D axial processing and suffered from poor capture range of Snake. Furthermore, bone marrows of spine and femurs got included in the calculated volume of VAT. In another method [5], localization of abdomen was done by deformation of a statistical shape model (SSM). Initial location of the deforming shape relied on the approximate location of the diaphragm, estimated as a fraction of the whole body height. Although this method provided an elegant means for abdomen localization by exclusion of bone marrows, it demanded a whole body scan.
1.jpg
Localization of abdomen in left-right and anterior-posterior directions, i.e. axial plane, has been achieved by different methods, e.g. 2D Snake [3] or surface fitting [4]. However, as mentioned above, this localization in superior-inferior direction is still inefficient and inaccurate. In this work, the starting 3D contours for abdomen localization were found by 3D registration of kidney. For this purpose, isotropic voxel size of 2Χ2Χ2mm3 were acquired. After edge-preserved denoising [6], RF-induced biases in both T1-weighted and adipose-only data were corrected using Bias Correction by Fitting of Adipose Tissue (BC-FAT) [7]. Then Otsu's thresholds or Fuzzy-C-means (FCM) classification and morphological filtering gave the total adipose tissue (TAT) mask. This mask was processed for automatic segmentation of abdominal adipose compartments.
2.jpg
For automatic 3D registration of kidney, several training data sets were segmented manually by MITK® software. The polygon representation of these reference contours were found and the corresponding vertices between them were associated by an B-spline transformation [8]. Alignment of the associated polygonal contours were done by minimization of the Procrustes distance between them. Final refinements were achieved by use of gray level profiles of active appearance model (AAM) [9]. After extraction of 3D contours of kidney, the steps of abdomen localization were as follows: 1. Calculation of gradient magnitude of TAT mask. 2. Watershed transform of the gradient magnitude of TAT mask. 3. Merging of all contacting watershed contours for finding the salient contour surrounding each AT compartment. 4. Exclusion of SAT by removal of voxels connected to the previously found body contour. 5. Exclusion of bone marrows of spine by removal of closed and separated contours located between kidneys in posterior direction. 6. Inclusion of other closed and separated contours located between kidneys in inferior and anterior direction.
3.jpg
It should be noted that in lean subjects, whose VAT compartments were separated, inclusion of 3D contour of kidney in the TAT mask could greatly enhance the speed of processing. In such data sets, the 3D watershed contours were used as the initial contours of a 3D Balloon Snake deformation algorithm converging to the boundaries of the abdomen in a 3D total lean tissue (TLT) mask. The TLT mask were generated in a similar way to TAT mask. The only difference was use of water-only data (instead of adipose-only data) obtained from CSD.
4.jpg
3. Results The automatic 3D algorithm was applied to both T1-weighted and adipose-only data of 15 subjects with BMI of 18-42 kg/m2. For each subject, the volumes of VAT quantified using proposed method and a previous 2D method [3] were compared. For one subject with BMI of 37.3 kg/m2, the entire VAT volume got segmented manually by a
radiologist, providing a ground truth for evaluation of the performance of the new and previous [3] proposals. Comparison with this ground truth, showed that the new algorithm could outperform the previous proposal. The superiority of the new algorithm in performance, could also be observed by inspection of the segmented areas of SAT and VAT in one coronal slice of T1-weighted or adipose-only data of aforementioned subject. The improvement in accuracy of VAT segmentation was achieved by enhancement of capture range and concave object handling via combination of 3D watershed transform and 3D Balloon Snake deformation.
5.jpg
For all 15 subjects, manual segmentation of VAT area in one axial slice located in L4 vertebral region, provided the ground truth. With respect to this ground truth, the relative differences between manually and automatically segmented VAT areas were used as an objective criterion for comparison of the new and previous proposals. This objective criterion is defined as:
6.jpg
Above criterion also showed superiority of the proposed method to previous proposal in accuracy of VAT segmentation.
7.jpg
4. Conclusion We proposed a new 3D algorithm for automatic segmentation of VAT. The novelty of this method was localization of the abdomen region by automatic 3D registration of kidney contours and combination of the 3D watershed transform with the 3D Balloon Snake deformation. Extraction of 3D contours of kidney could have a two fold application. One is automatic localization of abdomen without need to whole trunk coverage. The second is separation of adipose compartments inside kidney from segmented VAT region. This could support further studies about correlation between VAT volume, kidney's AT volume and different metabolic disorders. In combination of 3D watershed transform and 3D Balloon Snake deformation, watershed contours provided the initial contours for the Balloon Snake with the possibility of inflation and deflation of contours via internal and external forces of Balloon. This facilitated the placement of initial contours inside the abdominal region and also ruled out the need to contain the medial axes of abdomen in the initial deforming contours. In addition to automatic exclusion of bone marrows from VAT volume, above combination enhanced the capture range of Balloon, noise suppression capability, and concave object handling of the automatic segmentation algorithm in comparison with the previous proposal [3].
References: 1. 2. 3. 4. 5. 6. 7. 8. 9. 10.
Kuk et al., Obesity 2006, 14:336–341. Despres et al., Nature 2006, 444(7121):881–887. Wuerslin et al., JMRI 2010, 31:430–439. Addeman et al., JMRI 2015, 41:233–241. Wald et al., JMRI 2012, 36:1421–1434. Whitaker, et al., Proceedings International Conference on Image Processing, 2001, vol.3, no., pp.142,145 vol.3, 2001. Wuerslin et al., JMRI 2011, 34:716–726. Slagmolen et al., MICCAI Workshop on 3D Segmentation 2007, pp 197-206. Wald et al., JMRI 2012, 36:1421–1434. Dagher et al., Image and Vision Computing 2008, 26:905–912.
5. Mediafiles
1.jpg
2.jpg
3.jpg
4.jpg
5.jpg
6.jpg
7.jpg
