Research project is (co) funded by the Slovenian Research Agency.
UL Member: Faculty of Mathematics and Physics
Project: Computer-aided differential diagnosis of parkinsonism based on FDG-PET imaging
Period: 1. 10. 2021 - 30. 9. 2024
Range per year: 1,4 FTE category: C
Head: Urban Simončič
Research activity: Natural sciences and mathematics
In this project we will develop two novel differential diagnostic tools for early differential diagnosis of parkinsonisms and compare them with the standard clinical diagnostics. Parkinsonian syndromes are a group of syndromes, characterized by hypokinesia, rigidity, resting tremor and abnormal gait and posture. The most common is Parkinson's disease (PD), with an estimated global prevalence of more than 10 million cases and about 5000 cases in Slovenia. Atypical form of parkinsonisms includes multiple system atrophy (MSA), progressive supranuclear palsy (PSP) and corticobasal degeneration (CBD). The usual approach to differential diagnosis among parkinsonian syndromes is clinical, but is often challenging in early disease stages, with up to 25% diagnostic failures.
Metabolic brain imaging with fluorodeoxyglucose and positron emission tomography (FDG- PET) measures cerebral glucose metabolism that is closely associated with the local neural integrity. The topographic characteristics of abnormal glucose consumption are different in various parkinsoniisms. Visual interpretation of FDG-PET images improves differential diagnosis accuracy, but requires highly trained expert. A multivariate spatial covariance technique known as scaled subprofile model (SSM) based on principal component analysis (PCA) on FDG-PET images is used to identify a disease-specific metabolic pattern. Quantitative index of pattern expression for particular syndrome can discriminate patients with this particular parkinsonism from healthy individuals. A multiple-pattern analytical technique provides probabilities for PD, MSA, PSP and CDB, based on the FDG-PET image. However, this technique relies on network imaging biomarkers that were optimized for discrimination between patients with one particular parkinsonism and healthy individuals, and does not reach the performance of the visual evaluation of FDG-PET images supported by univariate voxel-based statistical analyses.
Our first approach towards the computer-aided differential diagnosis of parkinsonism is an extension of network analysis. Patients with all types of parkinsonism will be included in the analysis, while conventional network analysis typically include healthy controls and one parkinsonism; multinomial logistic regression will be applied after SSM/PCA, instead of standard logistic regression. In the second approach will classify parkinsonian patients with deep learning-based model, consisting of multiple convolution neural network layers. We expect that we will exceed the performance of multiple-pattern imaging technique with sensitivity and specificity of 75% and 90%, respectively, and possibly reach the sensitivity and specificity of visual evaluation of FDG-PET images supported by univariate voxel-based statistical analyses (90% and 95%, respectively). Accurate early diagnosis of parkinsonism is critically important because: (1) different parkinsonisms are treated differently, (2) clinical trials of new and potentially disease-modifying drugs might fail due to the enrolment of misdiagnosed patients, and (3) prognosis of different parkinsonian syndromes differs considerably.
For this project we have assembled a unique interdisciplinary research team that includes medical physicists from the Faculty of mathematics and physics, neurologists from the Department of neurology, University Medical Centre Ljubljana as well as nuclear medicine physicians and medical physicists from the Department of nuclear medicine, University Medical Centre Ljubljana. As such, the team assures unique multidisciplinary platform for the proposed project that have expertise in medical image analysis, access to the clinical data and executive role in clinical practice. Therefore, the team can make progress in diagnosis of parkinsonism through the developement of algorithms for computer-aided differential diagnostics, verify novel algorithms on real clinical data and implement these algorithms into clinical practice.