Introduction

The cerebellum and its connections (cerebellar peduncles) have been implicated to play a significant role in the pathogenesis of essential tremor (ET) [1,2], however these abnormalities are not grossly evident on basic structural imaging. It is plausible that advanced techniques such as radiomics which possess the capability to extract subtle intensity, statistical and textural features could provide a biological signature for ET.

Methods

Data were acquired from two different scanners. The T1 Weighted inversion recovery fast gradient echo MR images were acquired on a 3T Philips Acheiva MRI scanner with a 32-channel head coil from 40 ET patients (Age 44.95 ± 12.63, 13 Female) and 38 HC (Age 46.05 ± 9.59, 10 Female) that were a part of a previous study [1]. The following acquisition parameters were used: TR=8.2ms, TE=3.8ms, flip angle 8°, field of view (FOV) = 256 X 256 X 165. Additionally, T1 weighted MPRAGE images from 17 ET patients (Age 51.59, 4 Female) and 23 HCs (Age 41.79 ± 11.98 , 7 Female) acquired using a Philips Ingenia 3T scanner with a 32 channel head coil with TR=8.05ms, TE=3.6ms, flip angle=8°, field of view (FOV) = 256 X 256 X 211 mm, voxel size = 1x1x1mm, acquisition matrix = 256 mm was also added to this cohort (total n=118; 57 ET, 61 HC). Following conversion of DICOM images to NIFTI, the standard FreeSurfer pipeline was applied to the T1 weighted images [3,4]. Seventeen regions in the cerebellar lobules defined bilaterally from the SUIT atlas [5] included lobules I-IV, V, VI, Crus I, Crus II, Vermis Crus I, Vermis Crus II, VIIb, VIIIa, VIIIb, IX, Vermis IX, X, Vermis X, Dentate nuclei, Interposed nuclei and Fastigial nuclei. The inferior, superior and middle cerebellar peduncles defined bilaterally on a probabilistic atlas of cerebellar white matter[6] at a 50% probability threshold was used to define the cerebellar peduncles. All the cerebellar ROIs defined in the MNI space were non-linearly registered to individual subject space using advanced normalization tools [7] and the registration quality was manually assessed. The open-source package pyradiomics[8] was used to extract 18 first-order statistical features, 24 GLCM (Gray Level Co-occurance Matrix) based features, 16 GLRLM (Gray level Run-Length Matrix) based features, 16 GLSZM (Gray Level Size Zone Matrix) based features, 5 NGTDM based features (Neighbouring Gray Tone Difference Matrix) and 14 GLDM (Gray Level Depenedence Matrix) based features. An 80-20 split of the data (n=118) into training and testing sets was done, with 94 subjects in the training set and 24 in the testing set. Since data from two different scanners were used, a data harmonization procedure [9] was applied to the computed radiomics features to mitigate the effect of scanner differences in these features. A z-score normalization was performed on all the harmonized radiomics features following which a two-stage feature selection method was employed to reduce the dimensionality of the data and retain an optimum feature set by eliminating redundant features. The first stage was a univariate feature selection using the top 100 best of highest scoring features with the ANOVA F-value used as the scoring parameter. This reduced the feature set to 100 features. The second stage involved a Random Forest Recursive Feature Elimination with Cross Validation (RF-RFECV), with a 5-fold CV used as a standard evaluation strategy. The 60 most discriminant features of were used as the input to the final classifier. The sklearn machine learning library was used for all the implementations [10].

Results

The radiomics based RF classifier performed with an accuracy of 83.33% (average CV accuracy 85.14%, Area under- ROC (AU-ROC) =0.92) (Figure.3 and Table 1). The 60 most important features that were retained after two-stage feature selection (Figure.1) included 22 GLCM features, 16 GLRLM, 9 first order, 6 GLDM and 7 GLSZM features. The ROIs that these belonged to were mostly right hemispheric, and consisted of the Right Dentate, Right V, Right VI, Right I-1V, bilaterally the Crus I and SCP, Vermis VI, and the Left MCP (Figure.1) of which Right Dentate and Right V alone contributed to the top ten of the most discriminative features (Figure.2).

Conclusion

The top 10 most discriminative features identified in patients with ET include the dentate nucleus and the lobule V of the cerebellum. The dentate nucleus has been frequently suggested to be involved in the pathogenesis of ET, additionally, lobule V is part of the motor cerebellum which is involved in the cerebello-thalamo-cortical network, a part of the tremor network. These results suggest that radiomics was able to accurately identify regions suggested to be involved in the pathogenesis of ET. Future studies including other tremor disorders will be key to establish the utility of radiomics in diagnosing ET.

Table 1: Classification results based on T1 radiomics

Category		Precision	Recall	F1 score	Support	Test Accuracy	Avg CV Accuracy (5-fold)
ET vs HC	ET	0.83	0.83	0.83	12	0.83	0.85
	HC	0.83	0.83	0.83	12