Ceschin, Rafael
(2017)
A COMPUTATIONAL FRAMEWORK FOR NEONATAL BRAIN MRI STRUCTURE SEGMENTATION AND CLASSIFICATION.
Doctoral Dissertation, University of Pittsburgh.
(Unpublished)
Abstract
Deep Learning is increasingly being used in both supervised and unsupervised learning to derive complex patterns from data.
However, the successful implementation of deep learning using medical imaging requires careful consideration for the quality and availability of data. Infants diagnosed with CHD are at a higher risk for neurodevelopmental impairment. Many of these deficits may be attenuated by early detection and intervention. However, we currently lack effective diagnostic tools for the reliable detection of these disorders at the neonatal period. We believe that the structural correlates of the cognitive deficits associated with developmental abnormalities can be measured within the first few months of life. Based on this assumption, we hypothesize that we can use an atlas registration based structural segmentation pipeline to sufficiently reduce the search space of neonatal structural brain MRI to viably implement convolutional neural networks for dysplasia classification. Secondly, we hypothesize that convolutional neural networks can successfully identify morphological biomarkers capable of detecting structurally abnormal brain substructures.
In this study, we develop a computational framework for the automated classification of dysplastic substructures from neonatal MRI.
We validate our implementation on a dataset of neonates born with CHD, as this is a vulnerable population for structural dysmaturation. We chose the cerebellum as the initial test substructure because of its relatively simple structure and known vulnerability to structural dysplasia in infants born with CHD. We then apply the same method to the hippocampus, a more challenging substructure due to its complex morphological properties. We attempt to overcome the limited availability of clinical data in neonatal populations by first extracting each brain substructure of interest and individually registering them into a standard space. This greatly reduces the search space required to learn the subtle abnormalities associated with a given pathology, making it feasible to implement a 3-D CNN as the classification algorithm. We achieved excellent classification accuracy in detecting dysplastic cerebelli, and demonstrate a viable computational framework for search space reduction using limited clinical datasets. All methods developed in this work are designed to be extensible, reproducible, and generalizable diagnostic tools for future neuroimaging problems.
Share
| Citation/Export: |
|
| Social Networking: |
|
Details
| Item Type: |
University of Pittsburgh ETD
|
| Status: |
Unpublished |
| Creators/Authors: |
|
| ETD Committee: |
|
| Date: |
3 November 2017 |
| Date Type: |
Publication |
| Defense Date: |
26 October 2017 |
| Approval Date: |
3 November 2017 |
| Submission Date: |
3 November 2017 |
| Access Restriction: |
1 year -- Restrict access to University of Pittsburgh for a period of 1 year. |
| Number of Pages: |
128 |
| Institution: |
University of Pittsburgh |
| Schools and Programs: |
School of Medicine > Biomedical Informatics |
| Degree: |
PhD - Doctor of Philosophy |
| Thesis Type: |
Doctoral Dissertation |
| Refereed: |
Yes |
| Uncontrolled Keywords: |
Deep Learning, Neonatal MRI, Congenital Heart Disease, Structural MRI |
| Date Deposited: |
03 Nov 2017 18:13 |
| Last Modified: |
03 Nov 2018 05:15 |
| URI: |
http://d-scholarship.pitt.edu/id/eprint/33312 |
Metrics
Monthly Views for the past 3 years
Plum Analytics
Actions (login required)
 |
View Item |
![[feed]](/images/feed-icon-32x32.png){kind=icon}