-Medical Imaging Researcher -
David S. Wack, PhD
Recent Publications
Below are a few recent publications representing the algorithm and auditory imaging work of our lab.
MACC - Minimum Area Contour Change
​
The MACC algorythm was recently published through BMC Medical Imaging
http://www.biomedcentral.com/1471-2342/13/29 . The article shows that MACC can be used to dramtically improve agreement between expert raters drawing lesions. However, MACC is also an excellent choice for lesion ROI propogation, and can provide a dramatic time savings for an analysis center while also improving agreement between operators. MACC has already been used for other applications such as CSF flow. See "Cine cerebrospinal fluid imaging in multiple sclerosis", , Journal of Magnetic Resonance Imaging.
DOEE -- Detection and Outline Error Estimate
​
Work on DOEE was a result of earlier pattern recognition work. A problem that we faced was that algorithm testing was very dependent on the data set. What we observed is that agreement measured using Similarity Index was higher when assessing raters on high lesions images rather than low lesion load images. Our work on DOEE breaks agreement measures into two parts: 1) ability of raters to detect the same lesion; 2) ability of the raters to outline a lesion the same. Our two measures are stable for measuring rater agreement across a large range of lesion loads and can be used to estimate SI's dependence on lesion load. See http://www.biomedcentral.com/1471-2342/12/17 .
Complex Singular Value Decomposition
Much of our lab's work focuses on physiological parameter estimation from dynamic images. We have adopted a method that we developed for assessment of evoked response EEG, for dynamic imaging PET, MRI, and CT.
The two images on the left show a 1 minute frame at 31 minutes post injection from a dynamic Raclopride PET image, without CSVD (left) and with CSVD (right). Our paper also showed that the noise reduction improved the variance of physiologic parameters, such as binding potential. See: http://www.wjh.harvard.edu/~rajendra/wack_cmir_11.pdf .
Segmented Smoothing
​CT images have incredible contrast differences between Skull, Tissue, and vascular regions. This is great for segmenting imagings, however it is problematic when smoothing an image to reduce noise.
A voxel representing an artery can have values 10x higher than surrounding voxels representing tissue. When these voxels are smoothed, the arterial voxel causes an increase in tissue voxels, and the tissue voxels cause a decrease in the arterial voxel. It is straightforward to smooth using only voxels of the same type in the smoothing neighborhood. In fact methods such as bilateral smoothing weight the smoothing based on the underlying voxels. However, these typically lose the property of being based on seperable kernels. Seperable kernels allowing implementations literally 1000s of times fatser. We present a method that maintain seperable kernels, and seperates the smoothing of voxels based on a segmentation classification.
See our recent abstract: http://stroke.ahajournals.org/cgi/content/meeting_abstract/44/2_MeetingAbstracts/ATMP14