Deep-learning based fully automatic segmentation of the globus pallidus interna and externa using ultra-high 7 Tesla MRI

Abstract

Deep brain stimulation (DBS) surgery has been shown to dramatically improve the quality of life for patients with various motor dysfunctions, such as those afflicted with Parkinson's disease (PD), dystonia and essential tremor (ET), by relieving motor symptoms associated with such pathologies. The success of DBS procedures is directly related to the proper placement of the electrodes, which requires the ability to accurately detect and identify relevant target structures within the subcortical basal ganglia region. In particular, accurate and reliable segmentation of the globus pallidus (GP) interna is of great interest for DBS surgery for PD and dystonia. In this work, we present a deep-learning based neural network, which we term GP-net, for the automatic segmentation of both the external and internal segments of the globus pallidus. High resolution 7 Tesla images from 101 subjects were used in this study; GP-net is trained on a cohort of 58 subjects, containing patients with movement disorders as well as healthy control subjects. GP-net performs 3D inference in a patient-specific manner, alleviating the need for atlas-based segmentation. GP-net was extensively validated, both quantitatively and qualitatively over 43 test subjects including patients with movement disorders and healthy control and is shown to consistently produce improved segmentation results compared with state-of-the-art atlas-based segmentations. We also demonstrate a post-operative lead location assessment with respect to a segmented globus pallidus obtained by GP-net.

Mapping the electrophysiological structure of dystonic Globus Pallidus pars interna through intraoperative microelectrode recordings

“…Even if recent advances have been made in GPi bordering [40,41] and functional segregation of GPi [37], these approaches are not adequate to provide real-time feedback to neurosurgeon teams during DBS procedures for target localization. We used the definition of candidate neural biomarkers to locate and characterize neural behaviour inside different areas of GPi and to distinguish GPi borders for patients with primary dystonia.…”

Machine learning in deep brain stimulation: A systematic review

“…The planning phase comes third, the clinical problem being how to select the stimulation targets and determine the electrode trajectories prior to the operation. A first strategy is to assist the surgeon by automatically segmenting, or localising the subcortical structures of interest from pre-operative images ( [22,26,32,48,51,64,79,82]). Such works are important as the surgical targets are often small with low contrast in clinical images.…”

“…More specialized ANN structures, such as CNNs and RNNs were also used. CNNs were used 13 times in total, notably six times for image analysis: three times for subcortical structures segmentation or localisation with the VGG model [51], a modified ResNet structure [58], a custom structure called Hough-CNN [32] based on Hough voting, and U-net-based structures [64,79,82].…”

“…The majority of the papers did not employ a patient-wise validation method [10,13,14,15,16,17,18,19,20,23,24,25,30,31,28,34,35,37,38,39,40,41,42,43,44,45,46,49 [20,47], that further validation has to be done on other recording conditions or other centers [33,65,70,77,82] or that inter-patient variability was a limiting problem [46,68].…”

“…However, there were some papers that simply did not separate training and testing patients, could lead to biased measurements of their performance. The results reported by these contributions are interesting as preliminary work, but cannot safely be considered as representative of a real, prospective usage.Others measured their performance by employing a patient-wise validation method[11,12,21,22,26,27,32,29,33,36,47,48,51,54,55,57,58,61,62,63,64,65,66,69,70,72,73,74,75,76,77,78,79,80,81,82]. While it does not ensure the complete absence of data leakage (that could occur, for example, by mixing the validation and testing sets, or by selecting or normalize features with the whole database), these results can be considered as more reliable, and more representative of prospective-usage performance.…”