Convolutional Neural Network-Based Automatic Segmentation of Substantia Nigra on Nigrosome and Neuromelanin Sensitive MR Images

Automated volumetric determination of high R2* regions in substantia nigra: A feasibility study of quantifying substantia nigra atrophy in progressive supranuclear palsy

“…Although previous studies with convolutional neural networks have been proposed for automatic segmentation of SN, most of them focused on qualitative MRI imaging. 7,8,25 In this study, customized U-net++ was implemented and a comparative study was performed with other U-net types to assess the segmentation performance using both qualitative SWI and quantitative R 2 * map. The 3D SN volume and surface were subsequently reconstructed using the customized network.…”

“…10 3D SN segmentation on neuromelanin-sensitive MRI and SWI using a Vnet convolutional neural network yielded an average DSC in five-fold cross-validation of 0.7. 8 Further, SN segmentation was also performed using an ensemble of five state-of-the-art convolutional neural networks that achieved average DSC in five-fold cross-validation of 0.87-0.93 using SWI images collected only from healthy individuals, which may be highly computationally intensive. 7 These comparisons validated that our customized U-net++ network consistently produced efficient SN segmentation results for HC, PD, and PSP compared with other methods.…”

“…An automatic system that provides volume, surface area, and other shape properties of 3D iron‐rich regions in SN from raw MRI data using deep learning is highly valuable because it contributes to computer‐aided diagnosis that can reduce subjectivity and increase overall sensitivity in stratifying patients. Although previous studies with convolutional neural networks have been proposed for automatic segmentation of SN, most of them focused on qualitative MRI imaging 7,8,25 . In this study, customized U‐net++ was implemented and a comparative study was performed with other U‐net types to assess the segmentation performance using both qualitative SWI and quantitative R 2 * map.…”

“…Different neuromelanin‐ and iron‐containing midbrain nuclei were also segmented using a template of the brain from QSM data and achieved a DSC of 0.85 and 0.87 for neuromelanin‐ and iron‐containing SN, respectively 10 . 3D SN segmentation on neuromelanin‐sensitive MRI and SWI using a V‐net convolutional neural network yielded an average DSC in five‐fold cross‐validation of 0.7 8 . Further, SN segmentation was also performed using an ensemble of five state‐of‐the‐art convolutional neural networks that achieved average DSC in five‐fold cross‐validation of 0.87–0.93 using SWI images collected only from healthy individuals, which may be highly computationally intensive 7 .…”

“…Because unbiased and reproducible methods involving deep learning can potentially contribute to accurate segmentation, previous studies have investigated the automatic segmentation of SN using a convolutional neural network 7–10 . The segmentation of the SN has been mostly performed on qualitative MRI, such as neuromelanin‐sensitive MRI and susceptibility‐weighted imaging (SWI), whose contrasts may vary depending on the subjects, scanner, and scanning parameters 7–9 . In recent years, for reasons of necessity, segmentation of SN was also performed using one of quantitative MRI, quantitative susceptibility mapping (QSM) 10,11 .…”