Large-scale neuroimaging datasets acquired from multiple sites often use diverse scanners, formats, and acquisition protocols, resulting in differences in the acquired data. Manually reconciling these variations can be an arduous task that involves writing specific esoteric scripts for reading and re-formatting a maze of data formats such as DICOM, BIDS, and more, which increases the burden on researchers and reduces the reproducibility of analyses.

We present MRdataset, a unified interface that aims to simplify the traversal of neuroimaging datasets by adopting modular classes for each dataset format as well as different elements in the hierarchy (e.g., Subject, Session). Regardless of the format, MRdataset provides a consistent set of methods for data access without having to reconfigure the script every time for a new format or an extremely specific minor local variation. MRdataset also accommodates variations across scanners through custom classes for acquisition parameters (e.g., TR, TE) that encapsulate contextual information, including their physical units, value range, and their level of criticality to diverse types of analyses, to enable easy conversion across vendors that use different units (seconds vs. milliseconds).

Transitioning away from multiple formats/vendors is not possible. Therefore, tools such as MRdataset are required to foster consistency by accounting for these subtle differences in various formats and scanners.