Evaluation of an automated atlas-based classification method for DaTscanTM SPECT using an age-matched normal database

Objectives To evaluate an automated atlas-based method for the differentiation of Parkinson's Disease (PD) from healthy controls (HC) using DaTscan SPECT.

Methods 408 subjects were selected from the Parkinson’s Progression Markers Initiative (PPMI) database (www.ppmi-info.org/data). This group was divided into a “training dataset” (71 HC, 133 PD) and a “testing dataset” (70 HC, 134 PD). DaTscan SPECT images from each subject were automatically registered to a template using a two step affine registration, with the 1st step focused on the whole brain and the 2nd focused on independent hemispheres. Predefined atlas volumes of interest (VOIs) were then transferred to the SPECT image. Uptake for each atlas region (Striatum, Caudate, Putamen, Anterior Putamen, Posterior Putamen) was compared to an age-matched normal database (+/- 5 years from subject age), and z-scores were recorded for each atlas region. Kappa values from the “training dataset” were used to determine best cutoffs for differentiating PD from HC using z-scores from each region. These cutoffs were then applied to the “testing dataset”, and accuracy was recorded.

Results The z-score value of the Posterior Putamen was found to be the most accurate single differentiator with 95.5% accuracy in the training dataset and 92.7% accuracy in the testing dataset (both L and R evaluated). High accuracy was also observed using the Putamen with 92.1% accuracy in the training dataset and 90.2% accuracy in the testing dataset.

Conclusions An automated atlas-based method for DaTscan classification demonstrated >90% accuracy using a z-score cutoff for a single atlas region. We plan to perform future studies using multi-step classification methods to see if even greater accuracy can be achieved.