Porting CUDA workloads to AMD hardware
The existing TransMIL and CLAM codebases were designed around NVIDIA GPUs and CUDA. Because Frontier uses AMD GPUs, my first engineering challenge was adapting the software environment for ROCm and resolving compatibility issues across the deep-learning stack.
- Adapted CUDA-oriented research code for Frontier’s AMD GPU environment.
- Worked across PyTorch, ROCm, package compatibility, and HPC job execution.
- Established a working foundation for preprocessing, feature extraction, training, and evaluation.
Building the end-to-end pathology pipeline
After establishing a working environment, I contributed across the complete machine-learning workflow. Whole-slide images were segmented into non-overlapping 256×256 patches, converted into learned embeddings, and aggregated by a Multiple Instance Learning model for slide-level classification.
- Parallelized background removal, patch extraction, and feature generation across Frontier compute nodes.
- Processed slides averaging roughly 30,000 patches at 10× magnification.
- Integrated self-supervised ViT-MAE embeddings with the TransMIL classification architecture.
- Tested configurations using LUAD, LUSC, and DLBC histopathology data.
Learning the complete model-development process
One of the most rewarding parts of the internship was learning from experienced mentors who helped me connect each technical step to the larger research goal. It was both fun and challenging to move beyond running an existing model and understand how a machine-learning study is built from beginning to end.
- Prepared and organized datasets for preprocessing, training, validation, and testing.
- Learned how model architecture, feature extraction, and data quality affect downstream results.
- Tracked experiments, inspected training behavior, and evaluated models with multiple performance metrics.
- Discussed unexpected results, limitations, and future experiments with my mentor and research group.
- Gained confidence working through the full research cycle instead of focusing on only one isolated task.
Evaluation and research findings
I compared a conventional ImageNet-pretrained ResNet50 feature extractor with ViT-MAE models pretrained on histopathology images. In the reported experiments, the ViT-MAE configurations achieved test accuracies from 96.77% to 100%, while the ResNet50 baseline achieved 86.04%. I also used 10-fold cross-validation and bootstrap confidence intervals to evaluate variability and robustness.
- Evaluated accuracy, Cohen’s kappa, F1 score, recall, precision, and AUC.
- Generated 95% confidence intervals from approximately 1,000 bootstrap samples.
- Compared convergence behavior and training-loss curves across model configurations.
- Identified the unexpectedly strong transfer performance of a DLBC-pretrained ViT-MAE model as an area for future investigation.
| Model | Accuracy | Cohen Kappa | F1 | Recall | Precision | AUC |
|---|---|---|---|---|---|---|
| ResNet50 (ImageNet) | 0.8604 | 0.7180 | 0.8586 | 0.8605 | 0.8614 | 0.9433 |
| ViT-MAE (LUSC/LUAD) | 0.9677 | 0.9344 | 0.9672 | 0.9677 | 0.9674 | 0.9894 |
| ViT-MAE (DLBC) | 1.0000 | 1.0000 | 1.0000 | 1.0000 | 1.0000 | 1.0000 |
Communicating the research
I documented the project in a formal internship research report and was selected as one of the student poster presenters at the 2025 Smoky Mountains Computational Sciences and Engineering Conference. Presenting the work challenged me to explain the technical pipeline, results, limitations, and scientific relevance to an audience beyond the immediate project team.
Internship Research Report · July 2025
Transformer-Based Histopathology Image Classification
Project report detailing thebackground, methodology, and evaluation.
Exploring science and the ORNL community
The internship also included opportunities to explore the broader ORNL community and learn outside my immediate research project. These activities made the summer more engaging and gave me a wider view of the people, facilities, and technologies connected to the laboratory.
Serving the East Tennessee community
My ORNL experience extended beyond the laboratory. I participated in two ORNL-sponsored Saturday volunteer events, spending about five hours at each event for more than 10 total hours of community service. It was rewarding to work alongside other interns and staff while supporting organizations that serve people and animals throughout the Knoxville area.
- Young-Williams Animal Center: helped clean and organize animal areas and assisted with food and other facility supplies.
- The Love Kitchen: helped pack food and essential supplies for individuals and families in need.
- Worked alongside other ORNL interns and staff while supporting local nonprofit organizations.
What I learned
This internship showed me how rewarding it is to learn from experienced scientists and gain hands-on experience with machine learning, data preparation, and high-performance computing from start to finish. It was also exciting to use cutting-edge technology and work on a supercomputer like Frontier. I am also grateful that I had the opportunity to participate in volunteer events while interning at Oak Ridge National Laboratory.