Fall Research Expo 2020

Artificial Intelligence in Radiology Across Four Modalities

I worked together with Alisha Agarwal, Anjali Gupta, Abigail Manion, Lauren Rodio, and Samantha Turner.

PURM gave me the unique opportunity to collaborate on multiple projects over the summer with Dr. Chamith Rajapakse and his team remotely. Despite the limitations due to COVID-19 and remote work, I am entirely grateful that I joined a lab with a passionate team that fostered a welcoming environment and went above and beyond to help me understand every aspect of their projects. Before working in this lab, I had no experience doing any research, let alone medical research. This fact was short-lived as I was introduced into the world of radiology.

My main project was to create 3D segmentations of the cervical, lumbar, and thoracic spine based on CT and MR images. Currently, determining the anterior (Ha), middle (Hm), and posterior (Hp) heights of vertebral bodies is time-consuming and resource-intensive. However, artificial intelligence algorithms offer the ability to automatically determine Ha, Hm, and Hp and segment vertebral bodies for 3D bone quality assessment. Overall, the team aimed to develop and test a deep learning neural network for analyzing sagittal spine CT and MR images. Through this project, I learned how to properly segment the spinal cord's vertebrae and discs using the software application ITK-SNAP. I now understand how to analyze medical images and how they are used to achieve further medical breakthroughs.

Furthermore, I worked on another similar project that sought to develop a convolutional neural network to reduce the planning time for oral and maxillofacial surgery. Using DICOM series of head CT scans, I created a 3D model of the mandible in the medical imaging software platform 3D Slicer. The mandible would be digitally isolated by erasing portions of the upper teeth, the cervical spine, and most of the skull. The 3D model would then be smoothed, masked, and stored in a central database to await further analysis.

Lastly, I joined a project team developing xRAD, a cloud-based electronic medical records system (EMRS) that aims to empower physicians of developing nations in providing online access to patient data and communication. The modular system can create patient documenting charts, view scans, send out prescription orders, and determine the hospital's patient population statistics. I was tasked with contacting rural hospitals in the Philippines to implement the system. This project has taught me professional interpersonal skills as I facilitated video conferences between hospital stakeholders and my team.

My PURM summer has been an incredibly humbling experience that confirmed my educational interests in studying medicine. The research projects I worked on reminded me of how innovative medicine can positively impact human life. Moreover, working alongside such an intelligent and giving lab team cannot be beaten, and for that, I am genuinely thankful. I am excited for my future work in this lab.

PURM - Penn Undergraduate Research Mentoring Program
College of Arts & Sciences 2023
Join James Bradley for a virtual discussion
PURM - Penn Undergraduate Research Mentoring Program
College of Arts & Sciences 2023

Comments

Thank you, James, for such a clear explanation of your research project!  I have two questions: (1) how many scans do you have to segment to generate enough data to feed into your model to ensure your AI is as accurate as possible? And (2) Can you foresee any issues with training on images from injured or otherwise unhealthy spines? 

Congratulations on a great job! I just have a couple of questions... when preparing this dataset to be the input of a NN did you have communication with the deep learning team? And if so, how did that influence the decisions you made while segmenting the images? 

Awesome project! How much more efficient is designing an AI to program scans as opposed to doing it manually? Is this software close to becoming available for doctors and scientists, or has it been tested in actual patients?

I enjoyed reading the poster about the different projects you and your team worked on, very exciting especially with how certain regions of the spine can automatically be determined through AI. I want to ask about xRAD, where it stores patient data; what makes this a user friendly alternative to other patient records systems?