Mining databases to better understand neurodegenerative diseases
Automatic search and retrieval methods applied to a large brain dataset can extract valuable additional information.
04 May 2007
Ahmet Ekin, Radu Jasinschi, Jeroen van der Grond, and Mark van Buchem
Neurodegeneration correlates positively with age, and disorders such as Parkinson's and Alzheimer's are becoming more prevalent. An estimated 4.5 million people in the United States alone suffer from the latter, a number that could triple by 2050.1 Diagnosis and treatment require improved understanding of the causes and pathophysiological progression of these disorders. One source of new insights will be large medical image databases.
Two major obstacles to large database analysis are the absence of robust and automatic tools to objectively quantify imagery data, and a paucity of data that has been consistently acquired. Overcoming these problems will require concerted effort on the part of researchers who develop technology and the clinical technicians who use it. To that end, a collaboration between Philips Research and Leiden University Medical Center has brought about validated, fully automated, and fast magnetic resonance (MR) image-processing methods together with a large MR database. One of our main interests in this dataset has been to investigate whether the shape of an abnormality, such as a tumor or structural deformation, provides additional information about neurodegeneration and the disease process.
Figure 1. An axial slice of brain T2 image (left) with the red blocks showing the region of interest for shape features (middle), and the detected hypointense voxels in basal ganglia in red (right).
Some abnormalities, common and observed in multiple neurodegenerative diseases, have the potential to clarify the underlying mechanisms. Iron accumulation in the basal ganglia is one example.2 Tissues with high iron concentration appear hypointense in the T2 contrast of MR that makes possible the automatic noninvasive detection of iron-related deformations. MR is the preferred modality for the brain because of its high specificity. By changing device settings, the resulting images emphasize different physiological features. These images, such as T2, are called contrasts.
Figure 2. A 26-dimensional shape feature for a patient in our database (left) and a 13-dimensional signed hemispheric asymmetry feature for the same patient.
We have developed automatic methods to locate the basal ganglia and detect T2 hypointensities indicative of iron accumulation.3 To describe the shape of the these regions, we tessellate the detected region of interest with a 13-block grid in each brain hemisphere (see Figure 1).4,5 Based on the hypointense voxel (volume pixel) percentage in these blocks, we propose nine different descriptors: three for shape, four for accumulation, and two for asymmetrical distribution in the left and right hemispheres. The 26-dimensional (26D) block hypointensity percentage feature is our main approximate shape feature. The first 13 dimensions form the left hemisphere, and the remaining 13, the right. We obtain the four accumulation features by averaging selected entries of 26D shape features: total brain hypointensity load (average of the 26D vector); hypointensity load in the left hemisphere (average of the first 13 entries); the right hemisphere (average of the last 13 entries); and left/right (2D feature by concatenating left and right load). Both asymmetry features are 13D, with absolute and signed difference of the left and right shape features. Figure 2 shows two of these features.
Comparison of shape and accumulation features should show whether shape provides additional information to the cumulative features. We use search and retrieval followed by a nonparametric correlation analysis of the rankings to compare the two. A patient is selected as a query patient. The other patients in the database are then ranked by the similarity of each of the above nine features to those of the query patient. This results in nine different rankings (one ranking per feature). When two features are alike, the rankings should also be similar. Kendall's rank correlation coefficient6 computes the similarity of two rankings. This is done for all distinct pairings. To achieve robust statistics, we selected each patient in the database (n=550) as a query patient and performed the above steps separately for all. Averaging the results delivers the final correlation values.
Our experiments demonstrate a low correlation of 0.31 iron accumulation in the left hemisphere, showing that it is not predictive for the right hemisphere, and vise versa. Cumulative features have medium-strength correlation (around 0.50) to the corresponding shape features that describe the spatial distribution. This indicates that shape analysis can provide additional information. The asymmetry features did not show dependence on the sign (with a high correlation value of 0.78).
In sum, the shape of an abnormality such as iron-related hypointensities possesses additional information that can be extracted from the cumulative features commonly employed in clinical practice. We hope to further the analysis of large medical databases to validate the clinical relevance of shape features and to extend and refine the block-based shape descriptors and make them more voxel-accurate.
Ahmet Ekin, Radu Jasinschi
Video Processing and Analysis,
Philips Research
Eindhoven, The Netherlands
Ahmet Ekin is a senior scientist in the Video Processing Group at Philips Research. He received his PhD from the Department of Electrical and Computer Engineering, University of Rochester, NY, in 2003. He worked at Eastman Kodak, also in Rochester, as a consultant on MPEG-7. He served as a summer intern at AT&T Labs, Middletown, NJ, in 2001, and as a technical co-op at the Thomas J. Watson Research Center in 2003. His research interests include brain image analysis, search and retrieval, visualization, data mining, and pattern recognition.
Radu Jasinschi is a senior scientist in the Video Processing Group at Philips Research. He obtained his PhD in physics in 1983 at the University of São Paulo, Brazil, did postdoctoral studies in theoretical physics at Harvard University, and obtained a second PhD in 1995 from the Electrical and Computer Engineering Department at Carnegie Mellon University, where he was a research scientist (1986–1988) before moving to the University of Maryland (1988–1992). He worked at Tektronix (1996–1999) and Philips Research, USA (1999–2002). He holds seven US patents and has over 40 publications.
Jeroen van der Grond, Mark van Buchem
Department of Radiology
Leiden University Medical Center (LUMC)
Leiden, The Netherlands
Jeroen van der Grond graduated in 1987 with a degree in biochemistry and received his PhD in 1991 in radiology. From 1992 he was assistant professor, and from 1995, associate professor, in the Department of Radiology at University Medical Center Utrecht, with a focus on cerebrovascular imaging. In 2004 he moved to the LUMC. He specializes in anatomical and functional visualization of the human brain using magnetic resonance imaging and computed tomography. His areas of interest include aging, Alzheimer's disease, cerebral amyloid angiopathy, CVA (cerebrovascular accident), Huntington's disease, CADASIL (cerebral autosomal dominant arteriopathy with subcortical infarcts and leukoencephalopathy) and lupus.
Mark van Buchem studied medicine at Leiden University, receiving his MD in 1987. He trained in radiology at LUMC, and obtained his doctorate in 1992. In 1994–1995 he was a research fellow at the Hospital of the University of Pennsylvania. He was section chief of neuroradiology at LUMC in 1998–2004, and was appointed as professor of neuroradiology in 2002. In 2005–2006 he was a visiting professor at Harvard Medical School. Currently, he directs the neuroimaging research program and is an attending neuroradiologist at LUMC.
References:
Recent News
