Carlos A. Parra PhD, Abstract Co-Author: Nothing to Disclose

Samuel L. Brady MS, PhD, Presenter: Nothing to Disclose

Robert A. Kaufman MD, Abstract Co-Author: Nothing to Disclose

CT image quality is typically estimated using a variety of well-established system level descriptors. As part of an ongoing study to quantify these quality descriptors, the proposed neural network-based segmentation method was used to identity a set of anatomical structures from clinical CT imaging for a variety of patient sizes and body compositions. Future studies will rely on these results to establish an image quality methodology based on anatomical structures.

Contrast in CT imaging data was enhanced via histogram equalization, before applying 3-D filters used to derive frequency and intensity information; selected filters included morphological filters (open, close, dilate, erode), FFT band-pass filters, and Hessian filters. A linear vector quantization (LVQ) neural network was used to classify each voxel. For network training, labels from a set of anatomical structures and tissues (liver, spleen, kidney, lung, blood, bone, muscle, adipose) were assigned to coordinates manually selected from each structure. Tissue classification was based on a vector of local features (mean, standard deviation, maximum, minimum voxel intensity) computed from a seed growing vicinity around each voxel from each tissue volume. Code was written in MATLAB using Image Processing and Neural Networks toolboxes. (The MathWorks, Inc., Natick, MA, USA).

Successful segmentation was attained within a three-dimensional volume for liver, spleen, kidney, lung, blood, bone, muscle, and adipose tissue using different combinations of image features. Due to voxel size anisotropy, segmentation is optimized when the size of the local feature seed growing vicinity (in mm) closely resembles a cubic region. The segmentation code functions across a range of body sizes from thin to obese pediatric patients, and for a variety of patient image contrast levels (i.e., for contrast and non-contrast studies).

The segmentation code accurately identifies and segments thoracic and abdominal organs, providing the potential to segment different combinations of internal organs for patient-level automated in-vivo quantitative image quality analysis system.

Accurate anatomical segmentation of CT will be used for automated patient image quality analysis and further dose reduction investigations. Such analysis also can be extrapolated to adult CT imaging.

Parra, C, Brady, S, Kaufman, R, Automated Segmentation of Chest and Abdominal Organs for Image Quality Assessment in Pediatric CT.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14019613.html Accessed May 9, 2025