(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose:
The goal of this study is to use integrated Fractal Analysis and Uniform Manifold
Approximation and Projection (UMAP) methods to identify neurological disorder diseases automatically based on 4D rs-fMRI data.
Methods:
A lot of studies had demonstrated that fMRI BOLD is a complex signal, whose fractality can reveal fundamental properties of the brain functions. However, most existing fractal analysis methods have limitations and will cause information lost. In addition, small fluctuations cannot be resolved perfectly due to the discretized nature of the signal. In this paper, an innovation method was proposed that rs-fMRI data are de-noised and normalized using the well-known rs-fMRI processing toolbox DPAPI. Then, the Higuchi’s Fractal Dimension (HFD) D for each voxel is calculated.Each subject is represented by a let of HFDs, which describes the temporal characteristics of resting-state fMRI at the voxel level. UMAP was used for dimension reduction and visualized in a 2D space using the obtained HFDs. Automatic clustering and classification was obtained for Parkinson’s disease. Only fMRI data from slices 41- 47 after images registered to the Montreal Neurological Institute (MNI) common space are used for classification as it is assumed that the signals in motor areas are sensitive for the Parkinson’s disease.
Results:
Two preliminary studies using experimental fMRI data at resting state were applied to test the proposed method for automatic neurological disorder diagnosis. The first experimental rs-fMRI data is from an open database (https://openneuro.org/datasets/ds001354/versions/1.0.0). It included 14 patients with de novo Parkinsonian syndrome and 14 age- and gender- matched healthy subjects. Another experimental data is from the same open database (https://openneuro.org/datasets/ds000245/versions/00001) including 15 patients with Parkinson’s disease and 15 health control subjects are from Noritaka Yoneyama. It demonstrated the clear clusters of two groups of subjects.
Conclusion:
Most of subjects with Parkinson’s disease are separated from the healthy group. 5 healthy subjects in the first dataset and 2 healthy subjects in the second datasets are clustered with the Parkinson’s group and the reason is not clear at this stage. The preliminary results show the great potential using the proposed algorithms to perform neurological disorder diagnosis automatically. Additional datasets with other neurological disorders are being collected and tested using the proposed method.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
