For this project, I have trained a 2.5-D Transformer Backbone UNet to perform image segmentation on MRI images of brain lesions found in infants suffering from hypoxic ischemic encephalopathy (HIE).
Mortality rates from HIE can range from 15% to 82% in severe cases (Korf et al., 2023). An accurate, well-trained deep learning model could contribute to a reduction in this mortality rate by allowing for timely diagnosis and treatment of this condition and its related complications.
This project is inspired by a challenge listed on the Grand Challenge website.
This project was under active development until April 2026, with code and data being regularly updated as the project progressed throughout the semester. The project is now complete. A description of the project idea and final results may be obtained from the provided presentations. The entire code, all input datasets, output metrics and results for each training iteration as well as the final testing, and the weights of the final model, are all available in this repository.
The training, test and validation datasets have been derived from the work of Bao et al., 2025.