Classification of Amyloid status in cognitively normal subjects using NeuroRadiomics

Introduction: Classification of the amyloid status of [¹⁸F]Florbetapir positron emission tomography scans is normally performed by visual assessment. In the earliest, preclinical stages of Alzheimer’s disease, visual reading can be particularly challenging due to relatively lower amyloid load levels in the brain. Moreover, visual interpretation requires a highly experienced Nuclear Medicine physician to perform these classifications. Therefore, the aim of this study was to identify the optimal Radiomics features for classification of amyloid status in a sample of cognitively normal individuals with Subjective Cognitive Decline (SCD).

Methods: SUVr images of [¹⁸F]Florbetapir PET scans of 155 SCD subjects from SCIENCe cohort were included. 118 were classified as Aβ- and 37 as Aβ+ by visual assessment. T1 weighted MRI scans were also acquired from the subjects. Grey and white matter segmentations from the T1 weighted MRI scans were obtained using SPM8. Radiomics features were extracted using the RaCat software¹. Three bin configurations were tested: bin size 0.05 and 0.25, and a fixed bin number of 64. Three regions (whole brain, grey and white matter) were studied. Feature elimination to remove redundant and less statistically relevant features was performed in three steps: (1) features with less significant differences, (2) features highly correlated with SUVr or volume, (3) pair-wise elimination (using mutual correlations). Features common for all nine combinations or the regions (post elimination) were considered optimal. A simple T-Test was performed to assess statistically significant differences between the Aβ+ and Aβ- subjects for each radiomics feature and region in step one of the feature elimination. Spearman’s rank correlations was used in the second step of feature elimination. In this step, a threshold of 0.7 (correlation) was used. Any feature with a correlation higher than this threshold to either SUVr or volume was removed.

Results: A total of 478 radiomics features per combination were extracted. Less than 22 radiomics features per combination survived all three steps of feature elimination (Figure 1a). PET uptake metrics Original mean and Exact Volume were the only features common among all the nine combinations post feature elimination. Optimal features per region of interest are illustrated in Figure 1b. Most of these features that are either common at a region of interest level or for all the nine combinations had significant difference between the Aβ- and Aβ+ scans but no significant differences among bin configurations were observed.

Conclusions: A significant difference between the amyloid positive and negative scans could be obtained using most of the features after redundancy feature elimination. However, none of the metrics were able to fully differentiate subjects. Yet, some features were better able to differentiate subject groups better than the commonly used SUVr. The latter suggests opportunities for radiomics to improve classification of amyloid status.

References

Pfaehler E, Zwanenburg A, de Jong JR, Boellaard R. “RaCaT: An open source and easy to use radiomics calculator tool”. PLoS One. 2019 Feb 20;14(2)

• Download figure
• Open in new tab
• Download powerpoint

• Download figure
• Open in new tab
• Download powerpoint