The Medical University of South Carolina
MUSC Broadcast Messages
Find
Broadcast Messages
by email
To stop getting
Broadcast Messages
by email:

Using Deep Learning to Predict Autism in Infants Prior to 24 Months of Age

(Education/Training/Seminars)

January 31
11:00 pm
110, BEB
2014-01-23
Sandy Bird birdsm@musc.edu

The Center for Biomedical Imaging-Guest Lecture

Brent C. Munsell, Ph.D.
Visiting Assistant Professor
Department of Computer Science
College of Charleston

"Using Deep Learning to Predict Autism in Infants Prior to 24 Months of Age"

Brain enlargement has been observed in infants with autism prior to 24 months of age; specifically infants with autism brains are up to 10 percent larger than infants of the same age without autism, thus leading to larger cortical volume, cortical surface area, and cortical thickness measures. In general, symptoms of Autism spectral disorders are diagnosed at 24 months using the Autism Diagnostic Observation Schedule, however if a method existed that could accurately predict the onset of autism prior to 24 months, it be a great benefit for physicians and health care professionals that provide treatments for infants with autism. In this research a novel classification method is developed to predict the onset of autism in infants prior to 24 months using cortical features extracted from 6 month and 12 month high resolution neuroimages. The high dimension cortical features are reduced to low dimension codes using deep-learning, which are then used to train linear-SVM classifiers able to recognize infants at high-risk for autism. Experiments are performed that show the trained classifiers can recognize high-risk infants from low-risk ones with 88% accuracy, and can recognize high-risk-positive infants from high-risk-negative ones with 79% accuracy. Additionally, a new approach is developed to understand which cortical features in the high dimension cortical feature space impact classification in the low dimension code space.

171 Ashley Avenue · Charleston SC 29425 · (843) 792-2300