Optimized Haralick Texture Quantification to Track Parkinson’s Disease Progression from DAT SPECT Images

Objectives To optimize Haralick texture analysis for enhanced quantification of DAT SPECT images in Parkinson’s disease (PD) patients, aiming to investigate the potential of textural features as biomarkers of PD progression.

Methods For our proposed analysis, we performed Haralick analysis, which has found increasing utility in the field of radiomics and heterogeneity quantification. This is given the fact that Haralick analysis captures valuable local information, and at the same time, some of its metrics have been shown to depict very good robustness to segmentation and overall test-retest reproducibility, which we presumed would be advantageous for tracking of disease progression. 116 subjects from the PPMI database were included (72 PD; 44 healthy controls (HC); 3T MR images; age<70). For each SPECT-to-MRI registered/segmented image, we extracted the gray-level co-occurrence matrix, including 13 spatial directions in 3D, and different parameters: (i) gray-level quantization (4,8,16,32,64,128 gray levels), (ii) distance (1,2,3,4,5,6,8 voxels), and (iii) individual vs. accumulated distance (i.e. whether at increasing distances, information from lower distances is/is-not also included). We computed 13 Haralick measures, namely: (1) energy, (2) entropy, (3) correlation, (4) contrast, (5) variance, (6) sum mean, (7) agreement, (8) cluster shade, (9) cluster tendency, (10) homogeneity, (11) max probability, (12) inverse variance, and (13) dissimilarity, along with conventional mean uptake. We performed Pearson correlation analysis between our image-based metrics and four clinical measures: (i) The unified Parkinson's disease rating scale (UPDRS) - part III (motor). Disease duration (DD), taken with respect to (ii) time of diagnosis (DD-diag.) and (iii) time of appearance of symptoms (DD-sympt.). (iv) We also included a non-motor, cognitive outcome, specifically the Montreal Cognitive Assessment (MoCA). For systematic quantification of the correlation significance, we utilized bootstrapping with replacement (R=1000 sample sets) for each heterogeneity metric vs. clinical measure pair, followed by computation of correlation significance and averaging of p-values.

Results For conventional mean uptake analysis in the putamen, significant correlations appeared only when both HC and PD were included (Pearson correlation ρ=-0.76, p-value<0.05), not when applied to PD subjects only (ρ=-0.18, p-value=0.13), and no such correlations were seen in the caudate. By contrast, our texture analysis revealed significant correlations for PD subjects in the caudate. Increasing individual distances beyond 2 voxels, the majority of measures did not show improvements, and also stayed nearly the same if distance accumulation was utilized. The Haralick measure ‘inverse variance’ was seen to be overly sensitive to specific quantization parameters, including number of gray levels, distance and distance accumulation. The measure ‘homogeneity’ performed most favorably and consistently for the three clinical measures UPDRS (p-value<0.05), DD-diag (p-value<0.01) and DD-sympt (p-value<0.05), at gray levels 32 and 64, across a range of distances (16,32,64,128). For the cognitive MoCA scale, best performance was obtained for ‘variance’ and ‘correlation’ (p-value <0.05) for the abovementioned range of gray levels, and distance < 4.

Conclusions Optimized Haralick texture metrics applied to striatal DAT SPECT showed greater sensitivity to PD symptoms, beyond conventional mean uptake analysis, suggesting that textural features hold significant potential as biomarkers of PD progression.