Senior Design Projects
ECS193 A/B Winter & Spring 2023
Using Deep Learning of Graphs To Predict Response to Treatment in Early Psychosis from Task fMRI Medical Imaging
| ********** | |
| Ian Davidson and Jason Smucny | |
| Computer Science and Center for Imaging |
Project's details
| Using Deep Learning of Graphs To Predict Response to Treatment in Early Psychosis from Task fMRI Medical Imaging | |
| Response to treatment in early psychosis is highly variable, and no established methods yet exist that can predict whether or not people with recent onset illness will clinically improve after their first year of treatment. Using a cognitive control task, the Carter/Smucny laboratories have recently demonstrated that fMRI scans during cognitive control-associated activation can modestly (66% accuracy, with 70% positive predictive value and 60% negative predictive value; Smucny et al. (2019), Am J Psychiatry) predict if a person will develop recent onset psychosis. This preliminary study only used a few brain regions as predictors, did not examine functional connectivity, and did not incorporate all task conditions into the predictive model. | |
| The goal of this study is to improve upon these preliminary findings by taking a more comprehensive, "whole brain" connectivity approach using deep learning. Specifically, students will be provided with connectivity matrices of regions of interests (ROIs) that represent a fully connected graph between each pair of brain ROIs during various conditions of the AX-CPT task. The students will then construct a deep learning architecture in which a deep learner will assembled the connectivity matrices for each task condition and then combined to predict treatment response. This may also be combined with background knowledge such as the prior frontoparietal activation-based predictors as described above in Background to enhance prediction. The goal of the project is to improve upon the modest performance of the preliminary result described above using this more comprehensive graph-based approach to using task-associated data. | |
| • A hyperparameter-tuned deep graph learning-based architecture that improves upon the results of the Smucny et al. (2019) study in regard to predicting treatment response. • The students will be expected to deliver the source code in Github. • Comprehensive documentation of the code is required. • Preparation of appropriate tables and figures for publication in paper (students will be co-authors) if results are significant. |
|
| • Knowledge of python and python-based deep learning and data analysis tools are necessary (e.g., Pandas, scikit-learn, tensorflow, PyTorch) • Experience with data and results visualization • The ability to communicate results effectively in writing and via oral presentation. • The ability to work effectively as a team. |
|
| ********** | |
| 30-60 min weekly or more | |
| Open source project | |
| Attachment | Click here |
| No | |
| Team members | N/A |
| N/A | |
| N/A |