Towards convergence of AI and blockchain for personalized medicine in pharmacogenomics.

Focal demyelinating lesions in the brain and spinal cord constitute the pathological hallmark of multiple sclerosis (MS) [, ]. While all lesions appear hyperintense on T2-weighted MR images, only some are visible as “black holes” (BHs) on T1-weighted images, appearing hypointense compared to surrounding white matter (WM) []. BHs correlate with clinical disability at least as well as T2 lesions [] and mostly represent neuronal tissue destruction and axonal loss []. BH lesion counts and volumes are used as secondary endpoints in MS clinical trials [, ]. Being able to semi-automatically analyze BHs on 2D T1-weighted images is useful clinically, and there is a lot of existing data (e.g., legacy data from large clinical trials, and current and legacy data from clinical practice) that could provide useful insights and new information about the disease. Several automated methods exist that can detect and/or quantify BHs [–]. However, as detailed in Supplementary Table , these methods were validated on small groups of subjects from a single center [–, ] or only indirectly without direct comparison to manual labels [, ], and the image types necessary for some of these methods are not always available (e.g., (3D-)FLAIR [, ] or 3D-T1 [, ]).