3D Freehand Ultrasound for Neurological Diagnosis
Final Report Abstract
Background: Substantia nigra (SN) hyper-echogenicities measured with transcranial sonography (TCS) are a clinical imaging marker for the diagnosis of Parkinson’s disease (PD). As such, TCS is a complementary imaging modality to MRI, CT and SPECT in neurology, with an A level recommendation by the European guidelines for the differential diagnosis of PD. Research has further shown a large potential of TCS for early diagnosis of PD at an early stage, before the onset of motor symptoms. However, TCS of the midbrain area requires special training and expertise of the examiner, especially in 2D, due to e.g. ultrasound image artifacts, a small field of view and difficulty of the anatomy. The exact location and spatial distribution of SN hyperechogenicities within the midbrain is still unknown. Prior to this study, our group demonstrated the feasibility of 3D acquisition and analysis of TCS for detection of PD in a pilot study. 3D-TCS is a gateway technology for more objective SN imaging and a detailed analysis of its spatial distribution. Objectives: In this study, we followed two major goals. First, we aimed at a better understanding of the neurophysiological nature of 3D-TCS imaging of the midbrain, especially its connection to PD-related iron accumulation and nigro-striatal neuronal loss. Second, we aimed at further investigating the potential of 3D-TCS for the diagnosis of PD in clinical neurology, in particular regarding the objectivity and diagnostic accuracy of the method given less experienced sonographers. Methods: A large dataset of patients and healthy controls was scanned with 3D-TCS (N=62) and iron-sensitive quantitative susceptibility mapping (QSM) MRI (N=55), resulting in a joint multi-modal TCS-QSM dataset of 50 subject (N=23 PD, N=27 healthy controls). 3D-TCS was acquired bi-laterally through both bone windows and reconstructed uni-laterally. Towards the first objective, a multi-modal TCS-QSM template and atlas was built, based on several steps of image-based multi-modal registration. In template space, locations of iron accumulation were identified with QSM-MRI, and SN hyper-echogenicities were localized based on manual expert segmentations. Towards the second objective, the midbrain and SN hyper-echogenicities were manually segmented by two raters with different expert levels. Several machine learning based methods for automatic segmentation were proposed, including Random Forest based segmentation of SN hyper-echogenicities, and a midbrain region-of-interest (ROI) segmentation based on Deep Learning. Results: Our multi-modal analysis regarding the co-localization of SN hyper-echogenicities (TCS+) and QSM- measured iron accumulation (QSM+) yielded a high degree of overlap in and around the SN, but nowhere else in the midbrain. The main areas of overlap are the SN pars compacta (SNc) and the ventral tegmental area (VTA), confirming data from several patho-anatomical studies on iron-related alterations in SNc. On our dataset, the amount of SN hyper-echogenicities measured with 3D-TCS has a high diagnostic value for detection of PD, with sensitivity and specificity of 84.6% and 88.9% respectively. The inter-rater agreement in 3D was high (ICC(A,1)=0.777, p<10^-3), the classification performance of sonographers with different experience level was statistically not significantly different. Segmentation of the midbrain ROI with (fully) convolutional neural networks, and of SN hyper-echogenicities with a Random-Forest based approach is accurate within the degree of inter-rater variability for human segmenters. Unforeseen project results: The accumulation of iron in the SNc has been well described in literature. However, an accumulation in the VTA, confirmed with co-localized SN hyper-echogenicities in this area, is a surprising and novel result in our study, implying an involvement of the VTA and thereby the mesolimbic system in Parkinson’s disease. We further demonstrated for the first time that electrode tips in Deep Brain Stimulation (DBS) neurosurgery for the treatment of PD can be visualized and accurately localized to within 4.8mm accuracy with 3D-TCS. Success in dissemination: Our research has been published in more than a dozen conference and journal papers, along with further abstracts and invited talks in medical conferences. We were awarded as best poster at the DGKN conference 2015 and as runner-up for best paper at the IPCAI conference 2015. Overall, at the time of submitting this report, our publications have accumulated more than 1000 citations, more than 800 of which for our paper “V-Net: Fully convolutional neural networks for volumetric medical image segmentation”.
Publications
- 3D transcranial ultrasound as a novel intra-operative imaging technique for DBS surgery: a feasibility study. Int J Comput Assist Radiol Surg 2015;10:891–900
Ahmadi S-A, Milletari F, Nassir N, Schuberth M, Plate A, Bötzel K
(See online at https://doi.org/10.1007/s11548-015-1191-4) - Patient-specific 3D ultrasound simulation based on convolutional ray-tracing and appearance optimization. LNCS, vol. 9350, pp 510-518. 2015
Salehi M, Ahmadi S-A, Prevost R, Navab N, Wein W
(See online at https://doi.org/10.1007/978-3-319-24571-3_61) - Robust segmentation of various anatomies in 3D ultrasound using hough forests and learned data representations. Lect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics), vol. 9350, 2015, p. 111–8
Milletari F, Ahmadi SA, Kroll C, Hennersperger C, Tombari F, Shah A, et al.
(See online at https://doi.org/10.1007/978-3-319-24571-3_14) - Coupling Convolutional Neural Networks and Hough Voting for Robust Segmentation of Ultrasound Volumes. Proc 38th Ger Conf Pattern Recognit 2016
Kroll C, Ahmadi S-A, Navab N, Milletari F
(See online at https://doi.org/10.1007/978-3-319-45886-1_36) - V-Net: Fully convolutional neural networks for volumetric medical image segmentation. Proc. - 2016 4th Int. Conf. 3D Vision, 3DV 2016
Milletari F, Navab N, Ahmadi S-A
(See online at https://doi.org/10.1109/3DV.2016.79) - Hough-CNN: Deep learning for segmentation of deep brain regions in MRI and ultrasound. Comput Vis Image Underst 2017;164:92–102
Milletari F, Ahmadi SA, Kroll C, Plate A, Rozanski V, Maiostre J, et al.
(See online at https://doi.org/10.1016/j.cviu.2017.04.002) - MR imaging differentiation of Fe2+ and Fe3+ based on relaxation and magnetic susceptibility properties. Neuroradiology 2017;59:1–7
Dietrich O, Levin J, Ahmadi S-A, Plate A, Reiser MF, Bötzel K, et al.
(See online at https://doi.org/10.1007/s00234-017-1813-3) - A baseline study for detection of Parkinson’s disease with 3D-transcranial sonography and uni-lateral reconstruction. J Neurol Sci 2019;397
Plate A, Maiostre J, Levin J, Bötzel K, Ahmadi S-A
(See online at https://doi.org/10.1016/j.jns.2018.12.001)
