Following probe-handling agent. Some companies have tackled similar problems

Following recent advances in machine learning and computer vision, studies across various
disciplines are now applying powerful hammers from these fields to solve problems that were
thought to be impossible only several years ago. However, although at times fruitful, naïve
applications of previously successful algorithms without deep understanding of the underlying
principles may lead to minimal results, especially in biomedical context.
I had a taste of this medicine when I was working on a recent project for developing automation
solutions to the ultrasound-guided IV injection procedure. In ultrasound images, reliable
discriminatory feature extraction is difficult due to striking similarities between veins and
arteries. Common methodologies to identify veins used by human experts are to physically push
on the skin to check for compressibility and look for pulsations to rule out arteries, both of which
require image processing in time domain along with coordination with the probe-handling agent.
Some companies have tackled similar problems using large robotic arms to allow for precise
control of movements and easier needle pose estimation, but patients may reject the idea of a
large machine operating on them, especially in pediatrics where the procedure is used most
frequently. Given these unique characteristics in imaging modality, imaging conditions, and end
user experience, standard robotics and computer vision techniques without proper integration
with biomedical knowledge will most likely fail to provide satisfactory computerized guidance.
For other existing biomedical problems, such a phenomenon is further amplified with more
complex imaging techniques and domain-specific imaging results. With added complexity in
both realms, collaboration between pure computer scientists and clinicians may be ineffective
due to the disjoint knowledge sets.
I believe that interdisciplinary researchers with expertise in both data analytics and medicine are
the key to resolving this issue, and the Yale BME PhD program provides unique opportunities
required for passionate students to become the future leaders in this front. Within the intersection
of engineering and medicine, brain image analysis is the field that most intrigues me. Consistent
with this claim and my long history of affection for engineering with biomedical applications, I
pursued a robotics and bioengineering dual master’s degree at the University of Pennsylvania,
while conducting research in computer vision and neuroscience. I believe that my active pursuit
of interdisciplinary knowledge has prepared me as an ideal candidate for the program, and
ultimately as a future scholar in the field of medical image analysis.
As a robotics student, I focused primarily on machine learning and computer vision. Through
courses that emphasized both derivations and implementations, I solidified my understanding in
core machine learning concepts and other commonly deployed mathematical tools. Some
examples of projects include building a discriminative model for inferring tweet sentiment and
completing various state-estimation tasks in robotics. As for computer vision, I learned how to
formulate problems using concepts from linear algebra and optimization, and completed projects
for various image transformations, feature tracking, and object recognition.
Furthermore, I worked as a computer vision engineer for UPenn’s RoboCup team, where I
designed a new pipeline for robust real-time ball detection for autonomous soccer-playing
humanoid robots. An especially notable aspect of this project was that I combined elements from
standard computer vision techniques and deep learning to develop an algorithm tailored to the computationally limited on-board processor. Instead of running expensive convolutional neural
networks (CNN) on the entire image, I used a feature map in YCbCr color space that utilizes
special assumptions about the game play as well as the colors of the ball and the field to create
bounding boxes for region proposals, which were then classified with CNN. From this project, I
acquired deeper knowledge of techniques related to feature extraction and object detection, in
addition to practical coding skills (C++, Lua, MATLAB) and software development experience
in a team setting. This experience also cultivated intuition about coding for image analysis and
further established my passion for computer vision.
As a bioengineering student, I am working with Dr. Alan Stocker on a master’s thesis project on
developing a computational model for human visual perception. Human perception is vastly
complex, but until recently, a one-step Bayesian inference model was used to explain most
observed data. To combat various limitations of this simplistic approach, our lab has been
collectively investigating a new model involving an efficient encoder and a Bayesian decoder.
To explain briefly, the hypothesis is that the low-level neurons in the visual system are tuned to
maximize the mutual information between incoming sensory data and the encoded
representation, in which case we can deduce the non-linear mapping relative to the prior on the
orientation distribution. Using this model, we are able to explain existing published data by
simply changing the magnitude of noise at either the original stimulus space or the transformed
sensory space.
My initial contribution to this work was the analysis of the propagation of information at various
stages of the encoding-decoding process. Then, I designed and conducted a behavioral
experiment that could more explicitly support the model predictions, using the staircase 2-
Alternative-Force-Choice (2AFC) task and a visual stimulus created via spatial filtering of white
noise. In the process of fitting experimental data with the augmented model, I discovered a few
unexpected behaviors of the full decision network and interesting discrepancies in modeling
results based on the choice of loss functions and/or the inference method, and I am continuing to
search for an explanation for such behaviors. Naturally, this research experience helped
tremendously in enhancing my understanding of probabilistic modeling and computational
analysis. I believe that my acquired knowledge has prepared me to tackle challenges in other
computational fields, in terms of both aptitude and tenacity.
For my PhD studies, I want to combine my interests in computer vision and neuroscience to
solve various problems in medical image analysis. I believe that these interests align with those
of many faculty members in the Image Proessing and Analysis group, including Dr. James
Duncan,  whom I would be honored to have as my PhD thesis
advisor.