2023 Research

Team 1 Project: Predicting Ice-bed Topography using Predictive Modeling

Team members:

  • REU Student: Katherine Yi, Department of Computer Science, Purdue University, West Lafayette, IN
  • REU Student: Angelina Dewar, Department of Physics, University of Oregon, Eugene, OR
  • REU Student: Tartela Tabassum, Department of Information Systems, University of Maryland, Baltimore County, Baltimore, MD
  • REU Student: Jason Lu, College of Information Studies, University of Maryland, College Park, MD
  • REU Student: Ray Chen, Marriotts Ridge High School, Ellicott City, MD
  • RA Student: Homayra Alam, PhD student, Department of Information Systems, UMBC
  • Collaborator: Omar Faruque, PhD student, Department of Information Systems, UMBC
  • Collaborator: Sikan Li, Texas Advanced Computing Center, University of Texas at Austin
  • Collaborator: Dr. Mathieu Morlighem, Professor, Department of Earth Sciences, Dartmouth College
  • Research Mentor: Dr. Jianwu Wang, Associate Professor, Department of Information Systems, UMBC

Abstract: The purpose of this research is to study how different machine learning and statistical models can be used to predict bed topography in Greenland using ice-penetrating radar and satellite imagery data. Accurate bed topography representations are crucial for understanding ice sheet stability, melt, and vulnerability to climate change. We explored nine predictive models including dense neural network, LSTM, variational auto-encoder (VAE), extreme gradient boosting (XGBoost), gaussian process regression, and kriging based residual learning. Model performance was evaluated with mean absolute error (MAE), root mean squared error (RMSE), coefficient of determination (R2), and terrain ruggedness index (TRI). In addition to testing various predictive models, different interpolation methods, including Nearest Neighbor interpolation, Bilinear Interpolation, and Universal Kriging were used to obtain estimates the values of ice surface features at the ice bed observation locations. The XGBoost model with Universal Kriging interpolation exhibited strong predictive capabilities but demands extensive resources. Alternatively, the XGBoost model with bilinear interpolation showed robust predictive capabilities and required fewer resources. These models effectively captured the complexity of the Greenland ice sheet terrain with precision and efficiency, making them valuable tools for representing spatial patterns in diverse landscapes.

Deliverables:

Team 2 Project: Enhancing Real-Time Imaging for Radiotherapy: Leveraging Hyperparameter Tuning with PyTorch

Team members:

  • REU Student: Kaelen Baird, Departments of Computer Science and of Mathematics, Skidmore College, Saratoga Springs, NY
  • REU Student: Sam Kadel, Departments of Computer Science and of Psychology, Mount Holyoke College, South Hadley, MA
  • REU Student: Brandt Kaufmann, Department of Mathematics and Statistics, University of San Francisco San Francisco, CA
  • REU Student: Ruth Obe, Department of Computer Science, University of Houston-Clear Lake, Houston, TX
  • REU Student: Yasmin Soltani, Department of Biomedical Engineering, University of Houston, Houston, TX
  • RA Student: Mostafa Cham, PhD student, Department of Information Systems, UMBC
  • Collaborator: Zhuoran Jiang, Medical Physics Graduate Program, Duke University
  • Collaborator: Dr. Vijay R. Sharma, Department of Radiation Oncology, University of Maryland School of Medicine
  • Collaborator: Dr. Lei Ren, Associate Professor, Department of Radiation Oncology, University of Maryland School of Medicine
  • Collaborator: Dr. Stephen W. Peterson, Department of Physics, University of Cape Town, South Africa
  • Collaborator: Dr. Jerimy C. Polf, H3D, Inc.
  • Collaborator: Dr. Carlos A. Barajas, Department of Mathematics and Statistics, UMBC
  • Research Mentor: Dr. Matthias K. Gobbert, Professor of Mathematics, Department of Mathematics and Statistics, UMBC

Abstract: Proton beam therapy is an advanced form of cancer radiotherapy that uses high-energy pro-ton beams to deliver precise and targeted radiation to tumors, mitigating unnecessary radiation exposure to surrounding healthy tissues. Utilizing real-time imaging of prompt gamma rays can enhance the effectiveness of this therapy. Compton cameras are proposed for this purpose, capturing prompt gamma rays emitted by proton beams as they traverse a patient’s body. However, the Compton camera’s non-zero time resolution results in simultaneous recording of interactions, causing reconstructed images to be noisy and lacking the necessary level of detail to effectively assess proton delivery for the patient. In an effort to address the challenges posed by the Compton camera’s resolution and its impact on image quality, machine learning techniques, such as recurrent neural networks, are employed to classify and refine the generated data. These advanced algorithms can effectively distinguish various interaction types and enhance the captured information, leading to more precise evaluations of proton delivery during the patient’s treatment. To achieve the objectives of enhancing data captured by the Compton camera, a PyTorch model was specifically designed. This decision was driven by PyTorch’s flexibility, powerful capabilities in handling sequential data, and enhanced GPU usage, accelerating the model’s computations and further optimizing the processing of large-scale data. The model successfully demonstrated faster training performance compared to previous approaches and achieves an overall fair accuracy with so far limited hyperparameter tuning, highlighting its effectiveness in advancing real-time imaging of prompt gamma rays for enhanced evaluation of proton delivery in cancer therapy.

Deliverables: