Accelerated Convolutional Neural Network (CNN) for Autosegmentation of White Matter Tracts
The UPMC Department of Neurosurgery in collaboration with the Schneider Lab developed DSI Studio to perform fiber tracking on brain images to ease the process of localizing brain damage to white matter tracts during traumatic brain injury (TBI) diagnosis. Segmentation of these tracts is the process by which fiber bundles in the generated whole brain tractogram are assigned to specific tracts. This typically requires someone experienced with white matter tracts in order to complete and is very intensive.
Recently, research has evolved to develop methods to automatically segment whole brain tractograms using machine learning, but the training of these models must be done on each of the 28 white matter tracts, takes 22 hours per tract, and cannot be done simultaneously.
This project focused on devising a novel method, using a convolutional neural network, to speed up this training time for auto-segmentation methods and reduce the current 1 month training time for machine learning models to as little time as possible. As a result of this project, I learned R and Python as well as gained a conceptual understanding of concepts such as machine learning, deep learning, and convolutional neural networks. The final report can be downloaded below.