I chose to research this topic because I am convinced that neurodegenerative disorders are associated with systematic structural and functional modifications in the brain; however, this might not be evident to the naked eye, particularly at early stages of Parkinson disease. Computer-aided pattern recognition approaches might detect such subtle abnormalities associated with neurodegenerative diseases at the individual level and may therefore aid in the clinical diagnosis of individual patients.
Although the field of pattern recognition of brain imaging is progressing, the application to clinical routine is still limited, notably due to medicolegal issues. Each diagnostic test and technique cannot be 100% accurate. It must, therefore, be clarified how to handle false-positive and false-negative results. This medicolegal and ethical issue is similar to the discussion of self-driving cars. Who is responsible if a computer-driven car causes an accident?
Feedback regarding this research has been positive, and the fact that the field of pattern recognition of neurodegenerative diseases is rapidly expanding demonstrates the interest in this domain. Our own group published related papers for the diagnosis of mild cognitive impairment, including one in AJNR.1
Except for the multiple studies using the ADNI dataset, most existing scientific studies in the domain of pattern recognition in neurodegenerative diseases use a single imaging parameter, with relatively small sample sizes and 1 or a few scanners after careful harmonization.
However, in a clinical setting, there are different scanners with different protocols; typical MRI protocols include several sequences (eg, T1 for gray matter, T2/FLAIR/DWI/DTI for white matter, T2*/SWI for microbleeds, etc), clinical data, and sometimes follow-up imaging studies.
In order to bring pattern recognition to the next and clinically relevant level, it would be important to integrate all of this information into the classification analysis in order to improve the accuracy and robustness of classification. This is conceptually simple and straightforward yet very difficult to implement and optimize practically.