MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel
  1. Home
  2. Meeting Program
  3. All
  4. Radiomic Features Combined with Hybrid Machine Learning Robustly Identify Parkinson Disease Subtypes

Radiomic Features Combined with Hybrid Machine Learning Robustly Identify Parkinson Disease Subtypes

M Salmanpourpaeenafrakati1,2 M Shamsaei2, A Saberi3, G Hajianfar4, H Soltanian-zadeh5, A Rahmim1*, (1) Amirkabir University Of Technology,Tehran ,Iran,(2) University of British Columbia, BC cancer, Vancouver, BC, CA,(3) Islamic Azad University,Tehran ,Iran,(4) Rajaie Cardiovascular Medical And Research Center,Tehran ,Iran,(5) Henry Ford Health System,Detroit, MI, USA

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose:
Subdividing Parkinson’s disease (PD) into robust subtypes to enable more tailored treatment strategies.

Methods:
We developed advanced hybrid machine learning (HML) methods that were robust to variations in the number of subjects and features, for longitudinal datasets (years 0, 1, 2 & 4; Parkinson’s Progressive Marker Initiative). The following steps were taken: I) Segmentation of dorsal striatum on DaT SPECT images: via MRI, and directly on SPECT; II) Extraction of radiomic features (RFs) using our standardized SERA software; III) generation of 15 datasets: 5 with only non-imaging clinical information (1 timeless data; 4 cross-sectional data), and also including SPECT segmented using MRI (5 sets) or SPECT itself (5 sets); IV) applying HML constructed using 16 feature reduction algorithms, 8 clustering algorithms and 19 classifiers; V) cluster number optimization; VI) applying modified information criterion (MIC) for optimal subtypes selection; VII) Cross-linking subgroups; and VIII) confirming the findings using High-Dimensional Hotelling’s-T2 Test.

Results:
When using non-imaging clinical information only or combining it with conventional SPECT information, the clusters were not robust to variations in features, whereas utilizing RFs enabled consistent generation of clusters. We arrived at 3 clusters in each year, and on timeless data. Using training and testing process of K-means and multiple classifiers, we demonstrated 3 distinct subtypes consistently across years in cross-sectional and in timeless data. Hotelling’s-T2 test re-confirmed our findings. Heat-map analyses showed the 3 identified PD subtypes as: i) motor-dominant, ii) non-motor-dominant, ii) mixed motor/non-motor. Subtypes generated using SPECT-based segmentation or solely using conventional SPECT information remained less consistent when the number of subjects changed.

Conclusion:
Appropriate HML and independent statistical tests enabled robust identification of 3 subtypes in PD. This was achieved by combining clinical information with RFs extracted from SPECT images segmented using MRI, also demonstrating robustness to the number of subjects and features.

Keywords

Not Applicable / None Entered.

Taxonomy

IM- Multi-Modality Imaging Systems: MRI/SPECT - human

Contact Email

Additional Links