Cross-Modality Deep Learning-based Striatum Segmentation for Dopamine Transporter SPECT

Dopamine transporter (DaT) SPECT is a competent imaging method to diagnosis Parkinson’s disease (PD). Striatum segmentation on DaT SPECT is necessary to quantify striatal uptake, but is challenging due to the inferior resolution of SPECT images obtained from current general-purpose scanners. MRI is the preferred image reference for striatum segmentation due to its excellent soft tissue contrast. Deep learning (DL)-based segmentation method has been preliminarily applied on simulated DaT SPECT. This work proposes a cross-modality-DL-based striatum segmentation, estimating MR striatal maps from clinical SPECT images.
¹²³
I-Ioflupane DaT SPECT and T1-weighted MR images from 200 anonymized subjects with 152 PD and 48 healthy controls are analyzed from the PPMI database. DaT SPECT and MR images are registered, and 4 striatal compartments are manually segmented from MR images by a radiologist as the label. nnU-Net and standard U-Net are implemented using Pytorch with the Adam optimizer running for 1000 epochs. SPECT and MR striatal label pairs are split into train, validation, and test groups (136:24:40). DL methods are also compared to SPECT thresholding-based segmentation (THR-Seg). Dice and Hausdorff distance (HD) 95%, striatal binding ratio (SBR) and asymmetry index (ASI) are analyzed. Results show that nnU-Net achieves better Dice ([0.667, 0.689] vs [0.595, 0.615]) and HD 95% ([1.719, 1.879] vs [2.359, 2.643]) as compared to U-Net for all striatal compartments, which cannot be resolved by THR-Seg. nnU-Net is superior to U-Net and THR-Seg in SBR consistency (95% confidence interval of [-0.196, 0.203] vs [-0.275, 0.183] vs [-1.185, 0.109]) and ASI correlation (Pearson correlation coefficient, 0.80 vs 0.66 vs 0.37) for the whole striatum, as well as for individual compartments. The proposed cross-modality-DL-based striatum segmentation is feasible for DaT SPECT.