Using Prealignment in Global Volume Registration to Reduce Motion in rs-fMRIs (Jenna) and Machine Learning: The Big Picture and My Experience with Predicting Readmissions After Ileostomy Creation (Rob)

Seminar Date: 
Seminar Time: 
11am - 12pm
Seminar Location: 
5607 Baum Boulevard, Room 407A
Jenna Schabdach,MS & Rob Handzel, MD

Jenna's Abstract: Resting-state functional magnetic resonance images (rs-fMRI) are invaluable tools for evaluating the neurodevelopmental status of infants and neonates. Unfortunately, rs-fMRI sequences are highly susceptible to motion. Many post-acquisition motion mitigation techniques have been developed to attempt to remove the effects of motion from rs-fMRI sequences. Each technique addresses some effects of motion, but at a cost: signal is often removed along with noise, and one category of techniques eliminates entire frames from the rs-fMRI sequence. Many techniques use some form of global volume registration to align the frames of a sequence in the same space. Traditional global volume registration minimizes the differences between each frame in the sequence and the reference volume, but may not minimize the differences between other pairs of volumes. Herein, we develop the framework for an alternate approach to global volume registration. We suggest that an rs-fMRI sequence can be viewed as a directed acyclic graph (DAG), which models the relationships between neighboring frames. The relationship between a pair of neighboring frames can be used to initialize pairwise volume registrations. Prealigning volumes before the registration minimizes both the global and local differences between frames in the registered sequence. We applied both the DAG-based and the traditional global registration methods to a simulated rs-fMRI sequence and to a set of healthy neonatal rs-fMR images with significant motion artifacts (N=17). Both registration frameworks showed comparable corrective ability in terms of reducing the mean and standard deviation of the correlation ratio in most images; however, the DAG-based framework was more effective in reducing motion artifacts to levels meeting Power et al.'s framewise displacement and derivative of intensity variance criteria. The DAG-based registration framework shows great potential for reducing the effects of motion without removing signal, although larger studies are needed to confirm the generalizability of this technique to other patient populations as well as its compatibility with existing motion mitigation pipelines.


Rob's Abstract:  The development and utilization of machine learning in the tech industry has led to an increasing interest in the use of machine learning in medicine. Robert Handzel will discuss his experience with developing predictive models using data from a surgical registry to predict readmissions after an ileostomy creation. He will also discuss both the barriers and his vision of how machine learning will fit within the medical workflow.