Tran, Honson: A Deep Learning Approach For Marker-less Cerebral Palsy Classification Using Gait Video
Title: A Deep Learning Approach For Marker-less Cerebral Palsy Classification Using Gait Video
Name: Honson Tran
Major: Computer Science
School affiliation: School of Arts and Sciences
Programs: Aresty – RA Program
Other contributors: Jason Phu, Kang Li
Abstract: Cerebral Palsy is the most common of all childhood disabilities, affecting approximately three live births out of every thousand in the United States. About 764,000 children and adults currently have Cerebral Palsy. To accurately diagnose CP, accurate identification is a must. Often, a popular method for identifying CP is through gait analysis. By studying how patients walk, we can determine the severity of CP for each patient. However, assessing this severity has is quite inconvenient. For starters, state-of-the-art systems usually are very involved in setup and costs, making locations for testing to be limited. This is hugely inconvenient for patients who require frequent assessments but live farther away from these systems, causing both transportation and financial issues.
With the help of artificial intelligence and computer vision, we can tackle the accessibility issue of these systems, and also create a standardized approach for assessing CP. Although we have these systems in place, most of the diagnosis for severity is based on the doctor. Wouldn’t it be more useful if these ratings for severity were standardized? How would we be able to do it? Well, if we trained a program to rank severity for CP patients, we can have a more quantitative judgment instead of qualitative. Also, using artificial intelligence would help us find more relationships between and correlations that are not as noticeable to the naked eye with just the simple use of a couple of phone cameras.
So, how do we accomplish this? We collect a bunch of data, and we train it through a specific AI network called an hourglass network. Our dataset includes walking pattern records of 6 healthy children and 6 CP children. Subjects were asked to walk on a treadmill for about one minute with a digital camera recording their gait pattern and a synchronized motion capture system directly measuring their body movement. These video clips were then fed into the network for training and processing. Overall, the goal of this research is to explore the possibilities of attempting to decrease the use of expensive equipment and to make an assessment as simple as using a mobile application. With enough data, we hope to provide a basis for creating an objective rating system to calculate CP severity quantitatively from anywhere in the world.