Kanber, B;
Ruffle, J;
Cardoso, J;
Ourselin, S;
Ciccarelli, O;
(2020)
Neurosense: deep sensing of full or near-full coverage head/brain scans in human magnetic resonance imaging.
Neuroinformatics
, 18
pp. 333-336.
10.1007/s12021-019-09442-x.
Preview |
Text
Kanber_Neurosense. Deep sensing of full or near-full coverage head brain scans in human magnetic resonance imaging_AAM.pdf - Accepted Version Download (750kB) | Preview |
Abstract
The application of automated algorithms to imaging requires knowledge of its content, a curatorial task, for which we ordinarily rely on the Digital Imaging and Communications in Medicine (DICOM) header as the only source of image meta-data. However, identifying brain MRI scans that have full or near-full coverage among a large number (e.g. >5000) of scans comprising both head/brain and other body parts is a time-consuming task that cannot be automated with the use of the information stored in the DICOM header attributes alone. Depending on the clinical scenario, an entire set of scans acquired in a single visit may often be labelled “BRAIN” in the DICOM field 0018,0015 (Body Part Examined), while the individual scans will often not only include brain scans with full coverage, but also others with partial brain coverage, scans of the spinal cord, and in some cases other body parts. DICOM field 0018,1250 (Receive Coil Name) can be used to determine the type of receiver coil used, however, the use of a head coil does not guarantee a full-brain MRI scan. At other times another type of receiver coil such as a ‘dual coil’ may have been used, or the 0018,1250 (Receive Coil Name) attribute may be blank. Similarly, determining the field-of-view from the DICOM attributes is not sufficiently informative to detect full or near-full coverage brain scans, even if a brain scan is guaranteed. Furthermore, some datasets may have been converted to other formats such as the Neuroimaging Informatics Technology Initiative (NIFTI) format, with no access to the source DICOM files, and the associated tags. In other cases, DICOM files may be available but with no useful information in their headers, for example, if the DICOM files have been produced by the conversion of old plain-films, or as in legacy datasets. The imaging catalogue of the average hospital is too large to be manually checked, let alone labelled: a comprehensive framework for rendering imaging robustly auditable requires a fully automated process.
Archive Staff Only
View Item |