Radiomics Analysis of Longitudinal DaTscan Images for Improved Progression Tracking in Parkinson’s Disease

Objectives: DaTscan is increasingly utilized for diagnostic purposes in suspected parkinsonian syndromes. We aim to enable additional clinical utility of this modality for prognostication of Parkinson’s disease (PD). We investigated whether radiomics analysis (e.g. quantization of heterogeneity in uptake) as applied to longitudinal DaTscan images could predict UPDRS III (motor) for PD subjects.

Methods: We selected n=69 PD subjects from the Parkinson’s Progressive Marker Initiative (PPMI) database, who had at least 2 SPECT scans at baseline (year 0) and year 2 acquired on similar scanners (Siemens, 2-headed ECAM or Symbia systems), who underwent a high-resolution 3 T MRI scan, and for whom UPDRS assessment was available in years 0 (baseline), 2, and 4. Since routine quantitative analysis in DaTscan imaging does not assess spatial heterogeneity in uptake, we applied a framework wherein textural features were extracted from the images. This method does not require registration to a common template, and works in the subject-native space. Image analysis includes registration of SPECT images onto corresponding MRI images, automatic region-of-interest (ROI) extraction on the MRI images, and the extraction of various radiomic features. The outcome to predict is Y=year 4 motor score (UPDRS III). Non-imaging predictors included baseline measures of disease duration since symptom detection (DD-sympt.) and disease diagnosis (DD-diag.), as well as motor score, and non-motor (e.g. MoCA) clinical measures in years 0 and 2. The image predictors included 92 radiomic features extracted from the caudate, putamen, and ventral striatum of DaTscan images at years 0 and 2 to quantify heterogeneity and texture in uptake. Spearman correlation between each predictor and Y was calculated. Random forest (RF) with 5000 trees was used to combine both non-imaging and imaging variables to predict Y. The RF prediction was further evaluated using leave-one-out cross-validation. The analysis was repeated without radiomic features.

Results: Although 49 predictors were correlated with Y (unadjusted p-value<0.05), only motor scores at years 0 and 2, plus the radiomic feature eccentricity from the more affected ventral striatum, were significant with Benjamini & Hochberg adjusted p-value<0.05. The absolute difference between observed and predicted Y was D=2.74+/-2.78 (median 1.91, range 0.06~16.27) from RF and 6.75+/-6.39 (median 4.81, range 0.14~35.28) from leave-one-out cross-validation of RF. When excluding radiomics features, the D was increased to 3.26+/-3.04 (median 2.32, range 0.10~18.2) from RF, and 8.23+/-7.17 (median 6.03, range 0.31~41.4) from leave-one-out cross-validation of RF. Not making use of any imaging information, we obtain a D of 3.69+/-3.23 (median 3.02, range 0.07 ~ 17.2) from RF and 7.93 +/- 7.01 (median 6.13, range 0.0183 ~ 40.18) from leave-one-out cross-validation.

Conclusion: Our results demonstrated the ability to capture valuable information using quantitative striatal DAT SPECT imaging, enabling improved prediction of outcome when making use of radiomic features. Radiomic analysis of DAT SPECT has significant potential towards development of effective biomarkers for PD outcome. Research Support: The project was supported by the Michael J. Fox Foundation, including use of data available from the PPMI—a public-private partnership—funded by The Michael J. Fox Foundation for Parkinson's Research and funding partners (listed at www.ppmi-info.org/fundingpartners). This work was also supported by the Natural Sciences and Engineering Research Council of Canada. Table 1. Absolute difference (=D) between observed year 4 UPDRS III (Y) and prediction from random forest (RF)

• Download figure
• Open in new tab
• Download powerpoint