The MitoCellPhe program, and some cell images, on my computer.

My name is Carson Molder, and I am a senior computer engineering student at the University of Arkansas. As an undergraduate, I have performed extensive research with Dr. Justin Zhan of the Computer Science and Computer Engineering (CSCE) Department on artificial intelligence and computer vision. My research at the University of Arkansas has focused on leveraging artificial intelligence to enhance medical and biological image analysis.

I began working with Dr. Zhan in fall 2019, shortly after he was hired at the University of Arkansas. I met him through an e-mail he sent to students in the CSCE department seeking new student researchers. Because I was interested in doing artificial intelligence research, I sent him my resume. Soon after, he offered me a position in his lab, and I began working on creating a neural network model that learns textures in medical radiography images. In my previous blog post (link), I discussed this model in detail. Since then, I have submitted a first-author research paper on the model to Frontiers in Artificial Intelligence.

In the summer and fall of 2020, I have worked with Dr. Zhan on two additional research projects. In my first project, Dr. Zhan and I collaborated with Dr. Shilpa Iyer of the Biological Sciences department on creating an image analysis tool that analyzes microscope images to evaluate the networks mitochondria form in human cells. The tool calculates over 20 parameters that describe the networks these mitochondria form.

Before this project, Dr. Iyer used a basic tool [1] that analyzed mitochondrial networks in cell images. However, this tool could only analyze single images at a time, and cells had to be manually separated within an image (if an image contained multiple cells). These steps were time-consuming, so I wanted to automate them using computer vision algorithms.

The image analysis pipeline we created, MitoCellPhe (“Mitochondrial Cellular Phenotype”), performs its analysis in two steps. First, it converts each image into a skeleton that captures the shape of the networks found in each cell. During this step, the image can optionally be split into its individual cells using an object recognition algorithm. Second, it analyzes each skeleton to calculate many parameters that describe the underlying networks.

The MitoCellPhe pipeline takes a cell image stained for mitochondria (top) and converts it into a skeleton that represents the networks within the cell. Then, it evaluates these skeletons to determine the status of the cell’s mitochondrial networks.

Many of these parameters are novel—in addition to tracking networks, our pipeline calculates statistics on individual, round mitochondria (“punctates”) and long, straight mitochondria (“rods”). These parameters can indicate a cell’s respiratory health and could help doctors diagnose respiratory illnesses like COVID-19. Currently, I am contributing to a biology paper on the pipeline that describes its methodology and performance on two sets of cell lines.

In my second project, Dr. Zhan and I have begun working with Dr. Xiawei Ouyang at UAMS to create a neural network that could discover how a mother’s factors during pregnancy affect her children’s brain development. Since this August, Dr. Zhan and I have been exploring potential models for such a network, including EmbraceNet [2].

EmbraceNet is a powerful neural network model that can dynamically leverage multiple modes of data to make predictions, even when some modes of data are missing. Dr. Ouyang has provided an anonymized dataset of over 80 children subjects that contains many modes of data, like IQ test results and brain MRI imaging data. EmbraceNet can combine these sources, even when some modes are missing. In spring 2021, we plan to train EmbraceNet or another on this dataset to predict a child’s neurological development. Such a model could help doctors give targeted medical advice to expecting mothers in the future that increases their children’s brain development.

In the short term, I plan for my work on this network to serve as my honors thesis. In the long term, I plan to pursue a Ph.D., studying ways to adapt artificial intelligence to real-world problems. In addition to my work in medicine, I am interested in using artificial intelligence to accelerate computer systems like data centers, high-performance computing clusters, and storage systems. Such models could accelerate computing in every field, offering better performance for medical computation, scientific computation, and in industry. The Honors College Team Research Grant has been significant in bolstering my research career, studying ways to leverages artificial intelligence, and solidifying my decision to pursue a Ph.D.

References

1. Valente, A.J and Maddalena, L.A. and Robb, E.L. and Moradi, F. and Stuart, J.A. “A simple ImageJ macro tool for analyzing mitochondrial network morphology in mammalian cell culture.” Acta Histochemica, vol. 119, pp. 315-326, 2017.
2. Choi, J.H. and Lee, J.S. “EmbraceNet: A robust deep learning architecture for multimodal classification.” Information Fusion, vol. 51, pp. 259-270, 2019.