Analyzing intra-cranial volume from brain CT

Problem: collaborators at the Helen Wills Institute collects human CT scans as part of their experimental records. Currently, these collaborators are unable to analyze data because they lack schemes for fast visual exploration and quantitative assessment of inner skull volume;

Approach: we designed computer vision algorithms to segment the inner skull and extract features, including an easy to use interface. The current algorithm consists of 3 main parts: (a) anisotropic filtering of image slices to deal with spurious intensity variations and the fact that the CT scans present higher resolution in xy than in z; (b) connected-component analysis, an algorithmic application of graph theory, where subsets of connected components are labeled based on a given heuristic; (c) anatomical landmarks, acquired through quick user-interaction at the beginning of the analysis pipeline, plus prior information about the CT scan exam, for example, the scans contains head and neck regions only; the inner skull volume will present slices with the largest area perpendicularly and superior to the temporal lobes.

Results and Impact: our preliminary results show that we can calculate the intra-cranial volume in near real-time from CT images using multiplatform software, including automated image segmentation, and measurements that were not possible before. MORE


Discovering textural patterns from MRI using statistical descriptors

Abstract: We investigate datasets of T1-weighted MRI brain scans, aiming at discriminating normal from cognitive impaired patients, by describing the white matter (WM) image intensity variation in terms of textural descriptors. First, the WM is isolated using spatial proximity of voxels to constrain the probability with which voxels of a given intensity are assigned to WM. Second, we determine sub-regions in each slice that contains only WM, which are input to a statistical method for extracting the spatial organization of gray tones in MRI scans. Our study considers five textural descriptors derived from the isotropic gray level co-occurrence matrix, as angular second moment, contrast, correlation, inverse different moment and entropy. These descriptors form the feature vectors considered for later classification algorithm. We use MRI data sets from Open Access Series of Imaging Studies, made available by the Washington University Alzheimer's Disease Research Center, HHMI, NRG and BIRN. The image processing and analysis uses ImageJ tools, including an original plugin for calculating GLCM from specified regions of interest. Experimental results indicate that textural descriptors have potential to differentiate normal WM images from dementia-related WM images.


  • SAMSI 2012: Program on Statistical and Computational Methodology for Massive Datasets

  • International Symposium on Biomedical Imaging 2013 [MORE]

  • CT intracranial volume

    MRI slicing