- Cancer research
- Computational imaging
- Optics and image science
- Computational biophysics
- Cancer systems biology
- Image processing, analysis, and diagnostics
- Radar signal processing
- Bio-statistics and quantum information theory
Image processing and analysis: I have been involved in developing various image processing tools within larger frameworks of
imaging, informatics, and color science.
- Nanoscale nuclear architecture mapping (nanoNAM): The development of nanoNAM included developing nanoNAM data work flow, GUI, and image registration,
segmentation, and distortion correction algorithms. As an example, nanoNAM requires hyperspectral image acquisition. The resulting
image stack has a wavelength-dependent and spatially-varying distortion that has to be corrected for
extracting structural information from the data. Consequently, I developed a spline-based approach for estimating
the spatially-varying distortion field and correcting for it.
S. Uttam, H. V. Pham, J. LaFace, B. Leibowitz, J. Yu , R. E. Brand, D. J. Hartman, and Y. Liu,
Early prediction of cancer progression by depth-resolved nanoscale mapping of nuclear architecture
from unstained tissue specimens,
Cancer Res.; 75: 4718-4727 (2015)
- Mapping RGB color space to a dominant wavelength within a hyper-spectral wavelength
range:: Computing the wavelengths associated with the incoming light illuminating a color sensor is
generally an ill-posed inverse problem that lacks a unique solution. However, by restricting the calculation
to a single dimension, a dominant wavelength can be associated with the sensor color. This restriction is
achieved by mapping the RGB color sensor data to the two-dimensional chromatic CIE XYZ space with
a fixed value of light intensity Y. In this chromatic space the hyperspectral wavelength range is given by
a locus of points. By measuring the illuminant white point and noting that the sensor color is a convex
combination of a wavelength on the hyperspectral locus and the white point, I developed a ray-based
inversion that associated a color with the dominant wavelength residing on the hyperspectral locus. I used
this approach in our work on spectral encoding of spatial frequency, to characterize the dominant
axial spatial period of a 3D object in real-time.
S. Alexandrov, S. Uttam, R. Bista, R Zhao, and Y. Liu, Real-time quantitative visualization
of 3D structural information,
Opt. Express ; 20: 923-9214 (2012).
- Nuclear segmentation for stochastic optical reconstruction microscopy (STORM) super-resolution imaging:
I have developed a nuclei segmentation method for stochastic optical reconstruction
microscopy that avoids segmentation getting stuck in local minima due to numerous and distinct chromatin
clusters within the nuclei reconstructed using localization microscopy.
In STORM imaging of chromatin, the nucleus comprises of chromatin clusters that when taken in aggregate define the nucleus. Our nucleus segmentation algorithm exploits this idea of aggregation by performing elliptical parameterization of the nucleus through the eigendecompostion of the coordinates of the chromatin clusters. The resulting estimates of the major and minor axes is transformed to polar coordinates to compute the boundary of the nucleus. We empirically found that this eigen-based aggregation approach was able to correctly identify the majority of the chromatin clusters within the nucleus, except those clusters right on or near the boundary, where background staining becomes a confounding factor. To consistently exclude background staining, we applied a center-of-mass based local criterion on our initial boundary estimate, allowing us to robustly estimate it in a single boundary traversal. Specifically, we started with the initial boundary estimate and performed dilation and erosion on it to respectively get larger and smaller nuclei. The region of the nucleus between the two defined an annulus with the initial boundary estimate defining the principal curve of the chromatin clusters that lay in this annulus. We performed a single traversal on the principal curve, computing at each of its points the local estimate of the center-of-mass of the clusters. If the center-of-mass lay beyond the corresponding point on the principal curve, the nucleus boundary was locally expanded to the outer ring of the annulus, and if the opposite was true it was shrunk to inner ring. The expansion and shrinkage were performed smoothly using spline interpolation. The resulting boundary defined the segmented nucleus. An additional layer of robustness was provided through a graphical user interface for the expert to quickly visualize the segmentation results and adjust them, if necessary. We found that on most occasions the segmentation was accurate
J. Xu, H. Ma, J. Jin, S. Uttam, R. Fu, Y. Huang, and Y. Liu, “Super-resolution imaging
of higher-order chromatin structures at different epigenetic states in single mammalian cells,” Cell Reports; 24(4): 873-882 (2018).
Correcting stain variations in nuclear refractive index of histology specimens: This work was motivated by a desire to
account for the effect of variations in Hematoxylin
and Eosin (H & E) staining in histology specimens on measurements of refractive index. Theoretically, the well-established
Kramers-Kronig relations provide the ideal basis for computing variations in refractive index of H & E
stained samples. However, their accuracy is limited by the narrow bandwidth of the measured absorption spectrum, that
either requires assuming a valid extrapolation model to extrapolate absorption data beyond the measured spectral range,
or putting strict and possibly unrealistic restrictions on the spectrum outised the known range to justify the sufficiency of the available narrow bandwidth
data. These strategies are not practical for clinical samples. This work, therefore, focused on developing a regression model that derives from well-established
empirical relationship between the change in refractive index of a protein or nucleic acid solution and their concentration or mass density. The regression
model is then used to correct for variations in nuclear refractive index.
S. Uttam, R. Bista, D. Hartman, R. Brand, and Y. Liu,
Correction of stain variations in nuclear refractive index of
clinical histology specimens,
J. Biomed. Opt.;16: 116013 (2011).
Computer-aided diagnostics: As a part of my Master's thesis I developed a three-stage diagnostic
algorithm for classifying benign/malignant microcalcifications in digital mammograms. The first stage
extracted scale-based features. Specifically, wavelet texture features based on redundant multiresolution
analysis and contrast features based on a local band-limited contrast technique were extracted. The second
stage involved optimizing feature information by reducing the extracted texture and contrast features to
their respective independent feature sets using principal component based dimensionality reduction, followed by
independent component analysis. Feature fusion was then performed so that both (wavelet and contrast-based) sets of independent
features were represented by a single feature set. The third stage involved classification of these features as
malignant or benign using trained (using 10-fold cross validation) support vector machine. The algorithm
was applied to two publicly available mammogram databases: DDSM and MIAS. The resulting area under
the ROC curves respectively for the two bases were 0:98 ± 0:01 and 0:97 ± 0:02, and the sensitivity was
95:5% ± 0:5% and 95:1% ± 0:4%.
Back to top of page