Team 1 Project: Deep Learning Approaches for Cloud Property Retrieval: Comparing Fine-Tuning with Domain-Specific Architectures
Team members:
REU Student: Danielle Murphy, Department of Mathematics, University of California, Berkeley
REU Student: Caleb E. Parten, Department of Mathematical Sciences, Eastern New Mexico University
REU Student: Autumn Sterling, Department of Computer Science, George Mason University
REU Student: Haoxiang Zhang, Fairfax Christian School, Herndon, VA
REU Student: Kevin Zhang, Department of Computer Science, University of Maryland, College Park
RA Student: Xingyan Li, PhD student, Department of Information Systems, UMBC
Research Mentor: Dr. Jianwu Wang, Professor, Department of Information Systems, UMBC
Abstract: Accurate and timely retrieval of cloud properties is essential for near real-time weather forecasting. The GOES-R satellites are equipped with the ABI imager which is newer, detects 16 spectral bands, and offers higher temporal resolution than MODIS. SatVision-TOA is a foundation model trained on data from 14 MODIS spectral bands. This study aims to leverage both ABI’s enhanced capabilities and the SatVision-TOA foundation model. Two different approaches were explored: fine-tuning SatVision-TOA on ABI data and training models from scratch. Models were trained to predict four cloud properties: cloud mask, phase, optical depth, and particle size. For each approach, we developed both single-task and multitask models while employing various deep learning frameworks. Finally, we evaluated model performances to assess trade-offs between foundation model adaptation and domain-specific architectures.
Deliverables:
- Implementation Source Code at Github Repository
- Presentation Slides
- Presentation Recording
- SURF Poster
- Technical Report
- Peer reviewed publication
Team 2 Project: Improving Proton Beam Radiotherapy by Sequencing Simulated Patient Data in Compton Camera Real-Time Imaging with Neural Networks
Team members:
REU Student: Angelo Calingo, Department of Computer Science & Engineering, University of Nevada, Reno, USA
REU Student: Bikash Gautam, Department of Computer Science, Alabama Agricultural and Mechanical University, USA
REU Student: Peter L. Jin, James M. Bennett High School, Salisbury, MD, USA
REU Student: Sidhya Pathak, Department of Computer Science, University of Virginia, USA
REU Student: Michelle Zhao, Department of Computer Science, Cornell University, USA
RA Student: Harrison Lewis, PhD student, Department of Mathematics and Statistics, UMBC
RA Student: Hussam Fateen, PhD student, Department of Mathematics and Statistics, UMBC
Collaborator: Dr. Vijay R. Sharma, Department of Radiation Oncology, University of Maryland School of Medicine
Collaborator: Dr. Lei Ren, Department of Radiation Oncology, University of Maryland School of Medicine
Collaborator: Dr. Ananta Chalise, 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, M3D, 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 radiotherapy is an advanced cancer treatment technique utilizing high-energy protons to destroy tumor matter. When the proton beam interacts with the patient’s body, it emits prompt gamma rays, which can be detected by a Compton camera. However, image reconstruction of the beam path from these scatterings is plagued by mischaracterized scattering sequences and excessive image noise. To address this, machine learning models were developed to order the scattering events properly. Multiple novel datasets simulating particle interactions with human tissue were generated using Duke University CT scans and GEANT4 and Monte-Carlo Detector Effects (MCDE) software. An automated hyperparameter tuning framework was also built into the Big-Data REU Integrated Development and Experimentation (BRIDE) pipeline. This work implemented a novel event-classifier transformer and a 1D Convolutional Neural Network (CNN) to better understand spatial relationships in the data.
Deliverables:
- Implementation Source Code at Github Repository
- Presentation Slides
- Presentation Recording
- SURF Poster
- Technical Report
- Peer reviewed publication