Ayush Goyal

Research interests in philosophy of physics and science:
Hilbert Spaces, Schroedinger’s equation, wave functions, quantum physics, Einstein’s special and general theory of relativity, the God Particle, etc.

Research interests in bioethics and bionics:
Human versus Computer Cognition
Ethics of Computer Chip Implantation in Human Subjects
Social Repercussions of Genomic Engineering
Distinguishing Humans from Computers in the Bionic Era

Research interests in biophysics and computer science:
Cardiac modelling.
Medical image analysis.
Image segmentation, registration, tracking, and 3D visualization.
Intelligent / computer vision.
Pattern / object recognition.

Transforming research into innovations via collaboration with experienced scientists in avant-garde projects is the perfect way for a young, motivated researcher like me to participate in creating novel products that will further enhance medical and robotic technologies.

Current PhD Project

I am pursuing a PhD at the University of Oxford Computational Biology Group. My project is based on applying medical image analysis techniques in blood vessel segmentation of 50umx50umx50um voxel resolution cryomicrotome image data of ex-vivo sliced frozen porcine hearts. I am developing an image processing algorithm to filter out noise and extraneous edges from the pig heart images. Using an assortment of steering, filtering, smoothing, edge detection, thresholding, active contours, level sets, and various novel and assimilated image analysis techniques I am increasing the contrast between vessels and their surrounding halo, in order to subtract out the halo for correct vessel radius detection, while also preserving the smaller and less brighter vessels, which may have intensity values equal to or less than the halo. At the same time, the algorithm will intelligently discern whether a group of bright pixels is a blood vessel, image noise, or light reflection from underlying frozen tissue or vessels behind the current image slice. In this way, vessel radii and centreline information can be extracted and used to create a 3D model reconstruction of the porcine coronary vasculature. This model can then be used for 3D visualization of the branching pattern, morphology, and topology of the coronary vasculature down to the scale of micro-vessels, which would facilitate blood flow simulations providing insight into the causes of ischemia, stenosis, and other cardiac diseases.

Academic Qualifications

Experienced in biomedical robotics and informatics, artificial intelligence, medical image processing, and development of applicable software algorithms for these areas. I am an Eta Kappa Nu recipient of the 2005 Outstanding Electrical and Computer Engineering Student Award, which was received by only the top four electrical and computer engineering graduates in the entire United States of America for the year 2005. I obtained my B.S. in Electrical Engineering with a perfect 4.0/4.0 GPA at age 17 from Boise State University, Idaho.

One of my primary research interests is the development of software for enhancing medical images, i.e. the extraction of useful data from medical images for assisting doctors and surgeons in the analysis and treatment of cancers, tumors, and other malignant cystic diseases. Of all the elements of medical image analysis, image segmentation has been and will remain a crucial constituent. Image segmentation is the predominating procedure required for automated post-processing in biomedical imaging. Feature extraction, object recognition, and three-dimensional visualization of body organs, bones, and tissues, and many other applications are dependant upon image segmentation.

I am interested in developing novel image segmentation techniques by exploring which of several techniques for segmenting objects from images is best suited for drawing information from biomedical images, such as magnetic resonance and computed tomography. Threshold techniques, edge-detection, similar texture procedures, the active contour model, scale-space approach, and semi-automatic implementations are some of the many different practices for image segmentation:

The threshold technique analyzes local pixel data to segment image objects, and it is efficient for locating objects whose pixels’ intensity levels are distinctly apart from the intensity level of the background. This technique disregards the location / position of the object with reference to other objects in the image, and thus a fuzzy image’s indistinct region margins can result in flawed segmentation.
Edge-detection or contour detection depicts the edges / contours in an image at which the pixel intensity changes acutely. However, the defect of this technique is its frailty in joining distinct contour lines that together would constitute a segmenting boundary.
The similar texture method divides the image into associated zones by clustering bordering pixels of comparable intensity levels. Neighboring zones are combined based on the level of uniformity or dissimilarity of intensity levels and/or sharpness of zone margins. This technique needs to balance between rigorous divisions of the zones, which could result in disintegration of the image, versus moderate divisions of the zones, which could result in fusing of distinct zones with obscure margins.
The active contour model is a connection conserving reposed segmentation method, which begins with preliminary margin outlines symbolized as continuous curve-fitting lines, which it recursively alters with energy-minimizing splines directed by restrictive influences from image intensities that draw it towards lines and edges. I would be interested in combining this approach with “elastic” contours, which can dynamically fit to preferred object lines / edges in the image. Of particular importance in this approach is critical avoidance of local minima traps.
The scale space approach segments image objects at multiple scales, which are promulgated from undeveloped to fastidious scales. The decisive factor for segmentation can be randomly intricate, accounting for global and local factors. Scale space segmentation entails the concept that a linear signal could be explicitly divided into zones, with the scale of division controlled by a single scale parameter. The minima and maxima of the slope, i.e., the x-axis crossings of the second derivatives of multi-scale-curved forms of a signal, develop definitions of classifying associations among zones / segments at varying scales. Slope minima or maxima at undeveloped scales can be understood as following from tallying features at finer scales. A hierarchy of zones / segments is created when three segments divided by a slope maximum and slope minimum are merged when the slope extrema annul each other at a bigger scale.
Semi automatic segmentation is possible by using a wavelet-based algorithm to extract an object from an image by a human’s manual selection of a margin section that distinguishes an object from the image background. The wavelet algorithm scrutinizes the characteristic frequencies to detect a line / edge around the object and therefore functions well even if the margin section selected manually is fuzzy.

During my undergraduate education, I won a $5000 National Institute of Health (USA) sponsored Idaho Biomedical Research Infrastructure Network (BRIN) project (Grant # P20RR16454) grant award for assisting Boise State engineers in producing 3-D models of the pediatric knee by segmenting and region-growing magnetic resonance images (MRI) in biomedical virtualization software. I gained invaluable teamwork experience and problem solving skills when working to solve the problem of segmenting bone tissue from growth plate regions in MRI scans of the pediatric knee. My MATLAB image processing routine applied contour functionality and edge detection algorithms on magnetic resonance images to increase the contrast between growth plate and bone tissue in the endeavor to automate the segmentation process. The magnetic resonance images were processed in MATLAB as DICOM files. DICOM or Digital Imaging and Communications in Medicine is a medical imaging standard for repositing and communicating patient information. DICOM encompasses not only a file format, but also a TCP/IP based communication protocol. The standard enables two separate software applications to exchange patient data related to the medical image stored in a DICOM file.

The BRIN project's objective was to aid Anterior Cruciate Ligament (ACL) reconstruction, a common surgical procedure for adults, but complicated for children. ACL surgery in a growing pediatric knee involves drilling through the femoral and tibial physes (growth plates). Damage to the physes can be minimized using 3-D virtual and physical models for pre-operative planning and spatial reference. Virtual models of the physes can be developed using MRI scans of the knee. Segmentation, which is critical for virtual model creation, differentiates between the physes and surrounding tissue. This process involves rigorous human judgment to distinguish between the physes and surrounding tissue. For this project, I successfully developed and tested a segmentation routine that had great value as a "pre-processing" tool applied to MRI images before the creation of virtual and physical 3-D models.

My MATLAB routine was developed and tested to automatically enhance the MRI images by increasing the contrast and creating boundaries between the physes and surrounding tissue. However my algorithm would be greatly augmented if doctors trained my algorithm by manually segmenting the physes in several MRI scans of a pediatric knee. This information would be useful to train my algorithm in recognizing and detecting the difference in texture and pixel intensity level of the physes region as distinct from the surrounding bone tissue. The idea of humans manually segmenting an image for training computer software is intrinsic to Peekaboom, an online networked game developed by computer scientists at Carnegie Melon University to collect information about where in a given image a particular labeled object is located. This game augments the information gleaned by the ESP, another game that induces humans (web game players) to label images collected from the web. Peekaboom provides an exquisite interface for humans to interact with images, so that specific object locations can be identified in the images. It is an exemplar of fun human computation on the web that provides the collection of useful data. Moreover, other than simply improving computer vision and pattern recognition systems, Peekaboom provides a platform for addressing extensive issues that can be solved by human interaction with computers for augmenting the computers’ calculations. I would love to explore the possibility of using an interface similar to Peekaboom, using which medical experts could define the pixel area locations of particular bodily organs or tumors/cysts to train a segmentation algorithm in detecting the organs’ or tumors’ locations and boundaries.

Work Experience

Above, I have described my research background, experience, and interests in biomedical image applications. I also have experience in several aspects of designing computer systems:

Integrated circuit design: I have experience in simulating VLSI circuit models for double data rate DRAM technology using the DF2 layout schematic in HSpice software.
Hardware embedded systems design: I designed a robot car by programming with C language a PIC 18F4520 microcontroller for my embedded systems undergraduate project using an H-bridge for isolating the two DC motors (running the front wheels) from the microcontroller itself.
Firmware programming: I have programmed PIC 18F4620 microcontrollers to be used as base nodes in airplane cabin environmental monitoring systems. This ongoing research at Boise State University is sponsored by the USA’s Federal Aviation Administration to create wireless sensor networks to detect and record contamination in airplane cabins.
Software applications: At Hewlett-Packard's Storage Works Division, I designing and programmed in Python language a user interface for network storage engineers to analyze test results logged by servers and arrays.
Algorithm development: As I mentioned above, I developed a MATLAB image processing routine that applied contour functionality and edge detection algorithms on MRI scans of the pediatric knee to increase the contrast and create boundaries between growth plate and bone tissue in the endeavor to automate the segmentation, which is critical for virtual model creation.

As a software network test engineer at Hewlett-Packard Company, I designed and programmed in Python language a user interface (LogLyzerd, which is derived from Log Analyzer) for network storage engineers to analyze test results logged by servers and arrays. The project focused on creating an algorithm for time-correlating log files generated by servers and arrays that are configured in a huge storage area network. The software was built to adapt to new test ring configurations and learn about new test events (errors, check conditions, fail conditions, etc.) automatically. This project focused on adaptive software, but I would also be interested in working on also building adaptive machines, which can learn from interaction with the user and recognize the user's needs.

Delineated above are my interests in biomedical image segmentation, however I also appreciate research in the development of software programs, algorithms, and/or embedded circuit hardware for any field of image processing, pattern recognition, object recognition, and even other fields of robotics.