An interactive cascaded deep learning framework with expert refinement for accurate striatal subregion segmentation

Accurate delineation of striatal subregions in brain MRI is essential for reliable dopaminergic PET quantification and the assessment of neurodegenerative disorders such as Parkinson’s disease. In this study, we introduce StriaSeg-iARM, a cascaded deep learning framework designed for high-precision segmentation of 12 anatomically defined striatal subregions in native space. The model employs a two-stage 3D residual U-Net architecture, where the first stage localizes the striatum and the second performs voxel-wise subregional classification. To improve segmentation accuracy in anatomically altered cases, the framework incorporates an intermediate refinement step, enabling expert correction of the striatal mask prior to final inference. The model was trained on multi-scanner T1-weighted MRI data and evaluated on both internal and external datasets, including real-world cases with striatal atrophy. Compared to atlas-based normalization, a conventional single-stage model, and a cascaded model without refinement, StriaSeg-iARM achieved superior segmentation performance, with an average Dice coefficient of 0.91. Furthermore, segmentation accuracy was enhanced through expert intervention without requiring model retraining. Quantitative PET analyses based on StriaSeg-iARM also exhibited improved agreement with expert-defined segmentations. Overall, these results demonstrate the practical applicability of StriaSeg-iARM as a flexible and extensible framework for subregional segmentation and PET quantification.