Amal Farag PhD, Presenter: Nothing to Disclose

Evrim Bengi Turkbey MD, Abstract Co-Author: Nothing to Disclose

Le Lu PhD, Abstract Co-Author: Nothing to Disclose

Jiamin Liu PhD, Abstract Co-Author: Nothing to Disclose

Ronald M. Summers MD, PhD, Abstract Co-Author: Royalties, iCAD, Inc Research funded, iCAD, Inc Stockholder, Johnson & Johnson Grant, Viatronix, Inc

To develop an automated pancreas segmentation method using abdominal CT examinations.

60 subjects (mean age= 47±16 yrs, 37 % women) who were either a healthy kidney donor candidate (n=17) or had no major abdominal abnormality in the consecutive retrospective search of PACS in a month (n=43) were included in this study. Images were acquired in the portal venous phase with a slice thickness of 1.5-2.5 mm on MDCT scanners. The computationally efficient method is based on a hierarchical, three-tiered information propagation by supervised training and classifying of image patches, superpixels and 3D connected components. First, over-segmentation is obtained by employing the Simple Linear Iterative Clustering (SLIC) method. Second, a multi-level, multi-process feature extraction and classification framework is implemented that allows superpixel label maps to be projected back into the 3D volumetric space to obtain 3D segmentation. Numerous statistical and texture information features are utilized to describe pancreas or not on a patch level (i.e. 25x25 pixels) and superpixels-level (i.e. SLIC). The multi-phase feature extraction is coupled with random forest classifiers that are trained once on the patch level and in a two-level cascade on the superpixels level. Experiments were conducted using six-fold cross-validation. The pancreas was manually segmented by a radiologist for the reference standard.

The mean pancreatic volume was 59.2±30.1 cm3 for the reference standard. The total automated segmented pancreas volume was 66.1±43.9 cm3 (Figure), of which an average 41.6±25.4 cm3 corresponds to the true pancreas tissue when compared to the reference standard. The correlation coefficient between the automated pancreas segmentation and reference standard was 0.83. Dice (similarity) coefficient of 64.9%±22.6 was obtained in comparison to the state-of-the-art results of 58.2% ±20.0.

The proposed method shows promising automated segmentation results on one of the most challenging and unsolved radiology image processing problems. The highest similarity index was obtained compared to prior studies.

Automated pancreas segmentation is challenging due to high variation in pancreas anatomy and volume. An important potential clinical application is pancreatic volume measurement in diabetic patients.

Farag, A, Turkbey, E, Lu, L, Liu, J, Summers, R, Automated Pancreas Segmentation Using a Multi-level Information Propagation Approach in Abdominal Computed Tomography.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14013108.html