Radiomics analysis of substantia nigra on multi-echo susceptibility map-weighted imaging for differentiating Parkinson's disease from atypical parkinsonian syndromes.

Abstract

Background: While the "swallow tail" sign observed in the substantia nigra (SN) on susceptibility map-weighted imaging (SMWI) serves as an effective marker for differentiating patients with Parkinson's disease (PD) from healthy individuals, its visual assessment proves inadequate in differentiating PD from atypical Parkinson syndromes (APS).

Purpose: To employ radiomic features extracted from multi-echo SMWI of the SN to distinguish between PD and APS.

Material and methods: SMWI data were acquired from 63 PD patients, 38 APS patients, and 89 healthy controls. The participants were randomly assigned to either training or test groups in a 7:3 proportion. Utilizing the PyRadiomics software, a set of radiomic features were extracted from SN for analysis. Features underwent standardization via the maximum-minimum method, with 166 statistically significant features identified through independent t-tests. To minimize the risk of overfitting, the least absolute shrinkage and selection operator (LASSO) algorithm was implemented to identify and select the five most significant features from the radiomic dataset. Five distinct machine-learning classifiers were developed to distinguish between PD, APS, and healthy controls. The SHapley Additive Explanations was employed to gain insights into and visualize the relative importance of each feature within these models.

Results: Morphological, first-order, texture, and wavelet transform features of the SN emerged as the most crucial determinants. The light gradient-boosting machine model demonstrated superior performance in distinguishing between PD, APS, and healthy controls.

Conclusion: Radiomic features of the SN derived from SMWI show promise in differentiating PD from APS, potentially enhancing diagnostic accuracy in clinical settings.