Research: An Introduction

Ever elementary school, I've always been fascinated by the wonders of the STEM fields. As I grew into middle and now high school, I strive to use my passion for STEM in beneficial ways by doing research and applying for patents. I used the Beijing's notoriously poor air quality as the foundation for much of my air quality research. I also investigated new technologies such as carbon nanotubes and unmanned aerial vehicles. In the summer of 2016, I researched the best methods to choose training subjects for automatic segmentation of infant brains to better understand newborn brain development.

Here is some of the work I've done, comprised of a brief overview and an abstract, excerpts of an abstract, or preview of my research. You can download these files, along with the full paper or report in the downloads section.

Optimal Atlas Selection Methods for Automatic Segmentation of Infant Brains (2016)

INTRODUCTION. Although the premise of automatic brain segmentation in adult brains has been researched extensively, similar automatic segmentation in infant brains has received relatively little attention. Such automatic segmentations in infant brains, which have greater variability, if performed well, will allow researchers and medical professionals to increase segmentation speed and efficiency while maintaining accuracy, allowing less labor to be devoted on tasks that can be handled automatically. Automatic newborn brain segmentation will additionally lead to greater advances in the study of infant brain development. This study investigates the benefits of utilizing multi-atlas segmentation (MAS) to further ensure segmentation accuracy.

METHOD. A multi-atlas selection framework to automatically segment newborn brains with Magnetic Resonance Imaging (MRI) images was proposed. Fourteen subjects were utilized, complete with their structural brain MRI images and the corresponding manual labels, of which nine are from Boston Children’s Hospital and five are from South Africa. Of the 14 subjects, seven different atlas selection methods were proposed and implemented to perform training as atlas selection is a crucial step in segmentation and it is thus necessary to test a variety of methods to obtain the best results. A comparison of results using the Dice coefficient for all seven methods was conducted. These seven methods are as follows. 1) Using all thirteen other atlases for training. 2) Using all nine BCH atlases for training (eight if test subject is BCH). 3) Using all five South African atlases for training (four if test subject is South African). 4) Atlas selection by computing the similarity between each pair of the thirteen training atlases through an average of two similarity features (age and mutual information), and using the pairwise similarity with other subjects for each subject as features for k-means clustering. Similarity between the test subject and all other subjects were also computed, and all atlases in the closest cluster to the test subject were selected for training. 5) Similar to the previous method, with the difference being that three similarity features were used, including age, mutual information, and volume, with a different weighting scheme. 6) Similar to the previous method, with the difference being that instead of selecting all atlases in the closest cluster, the closest atlas in each cluster was selected. 7) Similar to the previous method, with the difference being that instead of selecting the closest atlas in each cluster, the five closest atlases to the test subject were selected regardless of clustering. In addition, number of closest subjects from method 7) were varied between 1 and 13 to determine the optimal number of closest subjects.

RESULTS. A weighted average of the Dice coefficients of all labels was computed for each of the atlas selection methods, based on volume of the respective brain region, in addition to a simple unweighted average of the individual Dice coefficients for each label. These results show that, on average, using all subjects for training was optimal in both label averages in addition to 22 of the 30 individual labels, with Dice coefficients of 80.0% and 70.9% for the weighted and unweighted averages, respectively. This is compared to the averages across all subjects and selection methods of 80.3% and 72.1%, respectively. However, these values from using all data subjects for training were only 2.2% and 0.8% higher than the average Dice coefficients for the last-ranked training set group, respectively (method 2 and method 7, respectively). Additionally, it was found that most labels and subjects displayed an increasing trend of Dice values through an increase in number of closest subjects from atlas selection (method 7).

CONCLUSION. The results show that although the use of all subjects for training may be optimal, the results are not statistically significant enough to definitively justify the use of one training set over another. Thus, this study has provided additional guidance for the optimal method of atlas selection for automatic segmentation of newborn brains and has demonstrated the feasibility of a MAS atlas selection framework.

Distributional and Developmental Analysis of PM2.5 in Beijing, China (2016)

3 years after China declared war on pollution, I wanted to know if Beijing's air quality really has improved. I found 9 years worth of hourly data from the U.S. Embassy website in Beijing, and used these data to analyze yearly, seasonal, daily, and hourly trends and patterns, as well as to fit a probabilistic distribution to the data. This is my third research project on air quality in Beijing. The abstract was accepted for the International Conference on Air Pollution and Control, which will be held in London on May 25-26, 2016. The abstract is as follows.

PM2.5 poses a large threat to people’s health and the environment and is an issue of large concern in Beijing, brought to the attention of the government by the media. In addition, both the United States Embassy in Beijing and the government of China have increased monitoring of PM2.5 in recent years, and have made real-time data available to the public. This report utilizes hourly historical data (2008-2016) from the U.S. Embassy in Beijing for the first time. The first objective was to attempt to fit probability distributions to the data to better predict number of days exceeding the standard, and the second was to uncover any yearly, seasonal, monthly, daily, and hourly patterns and trends that may arise to better understand of air control policy.

In these data, 66,650 hours and 2687 days provided valid data. Lognormal, gamma, and Weibull distributions were fit to the data through an estimation of parameters. The Chi-squared test was employed to compare the actual data with the fitted distributions. The data were used to uncover trends, patterns, and improvements in PM2.5 concentration over the period of time with valid data in addition to specific periods of time that received large amounts of media attention, analyzed to gain a better understanding of causes of air pollution.

The data show a clear indication that Beijing’s air quality is unhealthy, with an average of 94.07µg/m3 across all 66,650 hours with valid data. It was found that no distribution fit the entire dataset of all 2687 days well, but each of the three above distribution types was optimal in at least one of the yearly data sets, with the lognormal distribution found to fit recent years better. An improvement in air quality beginning in 2014 was discovered, with the first five months of 2016 reporting an average PM2.5 concentration that is 23.8% lower than the average of the same period in all years, perhaps the result of various new pollution-control policies. It was also found that the winter and fall months contained more days in both good and extremely polluted categories, leading to a higher average but a comparable median in these months. Additionally, the evening hours, especially in the winter, reported much higher PM2.5 concentrations than the afternoon hours, possibly due to the prohibition of trucks in the city in the daytime and the increased use of coal for heating in the colder months when residents are home in the evening. Lastly, through analysis of special intervals that attracted media attention for either unnaturally good or bad air quality, the government’s temporary pollution control measures, such as more intensive road-space rationing and factory closures, are shown to be effective.

In summary, air quality in Beijing is improving steadily and do follow standard probability distributions to an extent, but still needs improvement. Analysis will be updated when new data become available.

Deep Learning for Monocular and Autonomous UAV Flight in Woods (2015)

In the summer of 2008, 32 people perished as the result of over 3,000 wildfires in Northern and Central California. Such a disaster (second-most costly forest fire in US history) came to my attention, and I was intrigued by the possibility of creating an unmanned aerial vehicle that can fly autonomously in the woods to search for people and notify relevant personnel. A forest environment is very complicated, so an effective way for guidance is needed. To avoid a crash, depth information between the vehicle and the obstacles in the surroundings should be available for the drone to navigate.

The method can be divided into two groups in overarching machine learning: training and testing. During training, I use a camera-LiDAR (light-radar) system to obtain depth information of the images, so that the images are tagged with their respective depths. During testing, a Parrot AR Drone 2.0 was connected to the computer to use the training results to avoid trees.

I am the first to propose to use CNN in UAV guidance problem, which has many advantages upon other learning methods. The input of CNN is the whole image, so no features will be lost. The feature extraction and selection progress is done by the CNN, which will automatically choose the best features in the specified problem.

When applied our methods, the drone can fly autonomously in the woods, in my tests, 73% of the time successfully. Such a vehicle would be tremendously helpful in a variety of applications.

Effective Filtration of PM 2.5 in Air Cleaners and Masks (2014)

Continuing my interest in Beijing's air quality, I embarked on this project, my second in the field of environmental science. As the Chinese public gained increasing amounts of knowledge and expressed greater concern for Beijing's air pollution, residents have flocked to the stores, buying masks and air purifiers in huge quantities. Amidst the explosion of these industries, I wanted to know if and how much these masks and air purifiers actually helped. You can download the report in the downloads section.

Stretchable and Wearable Strain Sensors Using Super-Aligned Carbon Nanotube (SACNT) Films (2014)

In the winter and spring of my freshman year at high school, I worked on this project because of my increasing interest in chemistry. It was also one of my first research experiences, and I wanted to take advantage of this new material (SACNT) in an increasingly popular field of material science (carbon nanotubes) and see what I could do. The abstract is as follows.

Due to an increasing need for devices that are able to be stretched or bended while maintaining functionality, stretchable electronics have attracted much attention in recent years. Conventional rigid materials usually do not meet requirements of such high functionality. The purpose of this project is to use a new material - super-aligned carbon nanotube (SACNT) films to make stretchable and wearable strain sensors and demonstrate the potential applications for developing human-friendly devices.

In this study, I have found that the SACNT films possess super flexibility along the direction perpendicular to the axial direction, while most people have studied the electrical and thermal properties of carbon nanotubes on the axial direction. These carbon nanotubes are able to detect strain from a measured change in resistance, and can be stretched and worn due to its flexible properties at a macroscopic scale. The experimental results indicated that such SACNT strain sensors are capable of measuring strains up to 380% (76 times as much as conventional metal strain gauges), with high durability, fast response, low creep, and a super linear relationship between strain and resistance. I have made two devices using the SACNT sensors – a data glove and a simulated step counter that can detect various forms of motion. The data glove is light, simple, and able to sense any range of motion of the hand. This device might be used as a master-hand to control a remote slave robot, and perform tasks that humans cannot perform. Through this study, I have demonstrated the functionality of the new material and created human-friendly devices with abilities that are infeasible by a mere extension of conventional technology.

Reduction in PM2.5 Levels at the International School of Beijing Due to Positive Building Pressurization and Air Filtration Upgrades (2014)

This is another project I worked on with a friend and my teachers at the International School of Beijing, analyzing our school's approach to enhancing air quality indoors. The paper was ultimately accepteed at the International Indoor Air Quality Conference in Hong Kong, 2014. It was my first research experience on air quality in Beijing. The abstract is as follows.

INTRODUCTION: PM2.5, or particulate matter smaller than 2.5 μm in diameter, is a major air pollutant in Beijing, China. While recent attention has been focused on outdoor PM2.5 levels, many people spend much of their time indoors in the city and indoor air quality in Beijing is largely unstudied. Indoor air quality at the International School of Beijing (ISB) was negatively affected by a lack of proper building pressurization and inadequate air filtration. This resulted in excessively high indoor PM2.5 concentrations in the school during times of poor outdoor air quality. This study investigated the impact of recent upgrades to the air handling system at ISB on indoor concentrations of PM2.5. Part of ISB’s air handling system was upgraded during June and July of 2013 to create positive building pressurization and improved fresh air filtration.

METHODS: Two TSI DUSTTRAK II Aerosol Monitors (models 8530 and 8532) were used to measure PM2.5 concentration during two monitoring periods. Monitoring occurred before the implementation of the upgrades (during 17 days in Feb/Mar 2013) and after implementation (during 24 days in July/Aug 2013). PM2.5 concentrations were measured in 24 indoor locations, such as classrooms and offices, and compared to outdoor readings at three daily collection times.

RESULTS: The average indoor PM2.5 concentration before implementation of the upgrades was found to be 18.2 μg/m3, while the average PM2.5 concentrations outdoors during the same period was 96.4 μg/m3, which is equivalent to an 81% reduction. The average PM2.5 concentration of indoor air after the implementation of the upgrades was 5.2 μg/m3, while the average PM2.5 concentration of outdoor air during the same period was 111.5 μg/m3, which is equivalent to a 95% reduction. Fluctuations in indoor PM2.5 concentrations were significantly reduced after the upgrades and maintained an average indoor PM2.5 level of below 12 μg/m3 even though outdoor PM2.5 values fluctuated between 4 μg/m3 and 505 μg/m3. Indoor monitoring sites that were specifically targeted by the upgrades showed even greater reductions in PM2.5 concentrations.

CONCLUSIONS: This study supports the effectiveness of air management upgrades at ISB to address the issues of negative building pressurization, inadequate air filtration and high outdoor pollution. Therefore, schools in highly polluted cities can safeguard the health of students and staff through targeted air management improvements. This study found that high school students could be trained to effectively conduct such studies.