MRI/PET image segmentation for different tissues (Research Assistant, Dept. of Radiology, University of Wisconsin, Madison)
This project aims at developing a fully automated algorithm to classify tissues in a MRI image into bone, fat, water and other special tissues like liver, lungs, white/gray matter present in the MRI image. It is meant to be a generalized algorithm which can be used for any part of the body. To enable imaging of bone and lung, an ultra-short TE (UTE) radial acquisition was utilized in combination with an IDEAL-like reconstruction for quantitative water-fat separation. Preliminary aim is to make improvements in the existing segmentation based Attenuation Correction (AC) maps obtained from MRI images. Substantial improvements have been made by us in producing reliable AC-maps for chest. Chest is considered as one of the difficult body parts for MRI, as MRI is less sensitive to lung and spongeous bone tissues, both of which are present in abundance in chest region. This is a work in progress where we are working on segmentation based AC-map creation for other parts of body like brain, limbs etc from MRI based images. We will further investigate performances based on Atlas based registration methods too.
Joint Channel and Noise Variance Estimation in a MIMO Setup (Independent Study, Dept. of ECE, University of Wisconsin, Madison)
Analysis of non-stationary channels is a field of interest given how important mobile environments are today. This work aims at joint detection of parameters of channel and noise distributions in a MIMO setup by principle of Maximum Likelihood Estimation (MLE). MIMO channel is studied under statistical modelling.
Vocal Tract Modeling using FEM (Independent Study, Supervisor: Paul Milenkovic, Dept. of ECE, University of Wisconsin-Madison)
Studied the waveform generated by an acoustic model. Wave propagation is modeled as per the Helmholtz's differential equation model with forced boundary conditions. Aim was to build a model representative of the human vocal tract. In human vocal tract, the vibration of pharynx generates (sound) waves which stall at the mouth-air interface, after which sound propagates into the surrounding (outside mouth). In this study, wave transmission properties were studied using multiple boundary conditions, representative of an actual mouth-air interface, which is practically not completely elastic in nature.
Chirplet analysis of current signature in welding process to predict weld joint strength (Research Assistant, Supervisor: Surjya K. Pal, Dept. of Mechanical Engineering, Indian Institute of Technology, Kharagpur)
Adaptive chirplet transform was used to predict the weld strength joint of the final welded parts obtained, each having its unique process parameters. Wavelet transform method couldn't differentiate in certain cases, between two weld pieces of highly varying weld joint strength. So certain sensor parameters of these cases were analyzed using their chirplet coefficients. An algorithm was developed to extract certain statistical features from these coefficients. These parameters were found to be extremely helpful in not only distinguishing weld joints with different weld joint strengths but also well differentiated the cases were the wavelet analysis failed.
Multiresolution Analysis as a Tool for detection of fault in power transformer during Impulse Test (UGC Summer Research Intern, Supervisor: L. Satish, Dept. of Electrical Engineering, Indian Institute of Science, Bangalore)
A transformer’s high frequency equivalent circuit was simulated using PSpice version 9.2. In simulation, a switching fault was introduced in any of the sections for a very small duration of time, such that its effect on neutral current response of the transformer circuit couldn't be detected by naked eye. Multiresolution signal decomposition was then applied to the neutral current response, which showed the time of occurrence of the fault with high accuracy. The codes for performing the Multiresolution signal decomposition of the signal was self-written rather than using any MATLAB toolbox directly.
Extraction of transmission zeros for highly selective filters (Undergardauate Final Thesis, Supervisor: Srikanta Pal, Dept. of Electrical \& Communication Engineering, Birla Institute of Technology, Mesra.)
Determination of order and finding the location of the transmission zeroes of a filter for meeting the given specifications as the passband ripple, and the loss (in dB) beyond a given stopband frequency point, must always be greater than a specified value. An indigenous algorithm is developed for solving the highly nonlinear Chebyshev equations simultaneous to yield the order of the filter, location of the tx zeros and extreme frequencies. This method would help in designing a filter from the given specifications for various practical purposes. This algorithm has been successfully tested over a wide range of specifications.
Neurowavelet Packet Analysis based on current signature for weld joint strength prediction in pulsed metal inert gas welding Process (Research Assistant, Supervisor: Srikanta Pal, Dept. of Electrical \& Communication Engineering, Birla Institute of Technology, Mesra.)
Analysis of weld strength joints from entropy data of a pulsed metal inert gas welding process by application of Wavelet Analysis. RMS values Wavelet coefficients were fed into a neural network model to predict weld joint strength for future processes.
Other Small Projects: