I chose this topic because I am convinced that many neurodegenerative diseases have specific patterns of brain alterations that are too subtle to be detected by visual inspection alone, particularly at early stages of the disease. Early diagnosis is essential in order to initiate treatments to stop or at least slow down future neurodegeneration. This requires early detection at the individual level, which can be obtained using pattern recognition approaches known from machine learning.

Note that such pattern recognition analyses represent a fundamental shift in paradigm. Classic advanced neuroimaging studies compare a group of patients with a group of controls to identify brain areas involved in a given disease. Such results are fascinating from a research perspective, yet the results cannot be transferred to the diagnosis of individual patients. Pattern recognition analyses, in contrast, allow for detection of individual patients. It is fundamental to highlight that the clinically relevant question of our manuscript is the discrimination of typical PD versus atypical PD, which might be difficult on clinical grounds, in particular at early stages of the disease process. This in turn implies that pattern recognition studies should not compare patients versus controls, as this is not a clinically relevant question and is already possible based on clinical grounds. AJNR was visionary enough to detect this shift in paradigm and accept our study, which compares typical PD versus atypical PD, ie, the clinically relevant distinction, despite the fact that there is no healthy control group as would be required in “classic” group-level comparison studies.

There are, in my opinion, 2 main issues that remain to be solved prior to application of pattern recognition approaches in clinical practice. The first issue is methodologic. Pattern recognition approaches may detect very subtle brain alterations, which are well in the range of inter-scanner variability. This requires strict standardization of MR acquisitions—for example, using dedicated phantoms such that the dataset of an individual patient acquired on one MRI machine can be analyzed by a classifier trained and validated in a dataset of patients acquired on another (or ideally many other) MRI machine. The second issue is medico-legal. It is obvious that no technique may be 100% correct. Alterations detected using pattern recognition are complex and subtle, and even an experienced radiologist may not necessarily be able to validate these alterations. Therefore, the question arises how false-positive and false-negative results by the automatic classifier should be treated from a medico-legal perspective.

I have received excellent feedback so far. An increasing number of clinicians and researchers recognize that individual-level pattern recognition analyses will eventually allow early and specific diagnosis of individual patients. A related study in the domain of mild cognitive impairment¹ won the Lucien Appel Prize by the European Society of Neuroradiology in 2011.

Pattern recognition approaches may be used in many domains of neurodegeneration. In addition to Parkinson disease (PD),^2,3 such pattern recognition approaches also provide very promising results in neurodegeneration in the domain of dementia.^1,4,5 Most current pattern recognition studies use only one parameter, such as DTI-derived white matter alterations. As MRI typically includes multiple sequences, the next logical step is to integrate multiple parameters like gray matter derived from T1, white matter derived from DTI, resting-state fMRI, and arterial spin-labeling to further improve accuracy and robustness of classification. While this is conceptually straightforward, the optimal methodologic implementation of multiparametric pattern recognition is methodologically very complex and a topic of current research.⁶

We have ongoing projects of classification of multiple parameters including gray matter (T1), white matter (DTI), resting-state fMRI, perfusion (arterial spin-labeling), and iron deposition (susceptibility-weighted imaging), notably in the domains of mild cognitive impairment/Alzheimer disease and PD. We hope to present new results at the next ASNR meeting, in 2015.

References

1. Haller S, Nguyen D, Rodriguez C, et al. Individual prediction of cognitive decline in mild cognitive impairment using support vector machine-based analysis of diffusion tensor imaging data. J Alzheimers Dis 2010;22:315–27, 10.3233/JAD-2010-100840
2. Haller S, Badoud S, Nguyen D, et al. Differentiation between Parkinson disease and other forms of Parkinsonism using support vector machine analysis of susceptibility-weighted imaging (SWI): initial results. Eur Radiol 2013;23:12–19, 10.1007/s00330-012-2579-y
3. Haller S, Badoud S, Nguyen D, et al. Individual detection of patients with Parkinson disease using support vector machine analysis of diffusion tensor imaging data: initial results AJNR Am J Neuroradiol 2012;33:2123–28, 10.3174/ajnr.A3126
4. Haller S, Missonnier P, Herrmann FR, et al. Individual classification of mild cognitive impairment subtypes by support vector machine analysis of white matter DTI. AJNR Am J Neuroradiol 2013;34:283–91, 10.3174/ajnr.A3223
5. Haller S, Bartsch A, Nguyen D, et al. Cerebral microhemorrhage and iron deposition in mild cognitive impairment: susceptibility-weighted MR imaging assessment. Radiology 2010;257:764–73, 10.1148/radiol.10100612
6. Haller S, Lovblad KO, Giannakopoulos P, et al. Multivariate pattern recognition for diagnosis and prognosis in clinical neuroimaging: state of the art, current challenges and future trends. Brain Topogr 2014;27:329–37, 10.1007/s10548-014-0360-z

 

Read this article at AJNR.org …