A physics-guided and learning-based estimation method for segmenting 3D DaT-Scan SPECT images

Objectives: Quantitative uptakes in the caudate, putamen, and globus pallidus (GP) as measured from dopamine transporter (DaT)-SPECT images are being explored as biomarkers for diagnosis and measuring severity of Parkinson’s disease [1]. Reliable quantification of uptake requires reliable segmentation of these regions, which is challenging due to the low resolution of SPECT images and resultant partial volume effects, high system noise, physiological variability, and small sizes of these regions, in particular the GP [2, 3]. As an example, the GP is visually almost impossible to demarcate from SPECT images (Fig. 1a). However, MR images acquired from previously acquired patient populations can provide prior distributions of the shapes of striatal and pallidial regions. We integrate this knowledge with a physics and estimation-based approach to develop a method for reliable 3D segmentation of DaT SPECT images.

Methods: We designed a Bayesian approach that yields the posterior-mean estimate of the fraction volume that a particular striatal or pallidial region occupies within each voxel of a SPECT image. The estimate is obtained by minimizing a binary cross entropy loss function between the estimated and true fractional volumes over a population of images, where the prior distribution of the true fractional volumes is obtained from existing populations of MR images. The approach was implemented using an auto-encoder architecture. The approach was evaluated using realistic clinically guided simulation studies. Clinical T1-weighted MR images obtained from the OASIS 3 and ADNI databases were delineated using Freesurfer into grey matter, white matter, caudate, putamen, and globus pallidus for both hemispheres and registered in MNI-152 space with voxel size of 1 mm. For each MR image, clinically derived distribution of DaT uptakes [5] was sampled to determine the DaT activity ratios in these regions, resulting in unique 600 DaT activity distribution maps. Through realistic SPECT simulation and 3D OSEM reconstruction, SPECT images corresponding to these distribution maps with voxel sizes of both 2mm or 4mm were created. The network was trained on 500 images and tested on 100 images. The segmentation accuracy was evaluated using the metrics of fuzzy dice similarity coefficient (fDSC), fuzzy Jacard similarity coefficient (fJSC), and Hausdorff distance (HD) of caudate, putamen and globus pallidus for both hemispheres. The robustness of the proposed method to misalignment of SPECT image with the MR population was also evaluated.

Results: The proposed method significantly outperforms existing segmentation methods for both the left and right caudate, putamen, and globus pallidus regions for both voxel sizes (Fig. 1b). For e.g., the method yielded an fDSC of 0.863 (95% CI: 0.857, 0.869) for left putamen with 4 mm voxel size. Qualitatively, the boundaries yielded by the proposed method matched the ground-truth boundaries (Fig. 1a). The method was relatively insensitive to up to ± 4 degree misalignment between SPECT image and MR populations for both 2 mm and 4 mm voxel size (Fig. 1c).

Conclusions: A MR-guided fully automated segmentation method significantly outperformed other existing segmentation of caudate, putamen and globus pallidus for DaT SPECT images as evaluated using computational studies. The results motivate further validation of these methods using physical-phantom data and clinical studies.