Metadata Supervised MRI Representations for Modelling and Controlling Acquisition Variability

Medical imaging models often struggle with 'acquisition shift,' where variations in scanner hardware and protocols change image appearance without altering the underlying anatomy. Current methods typically treat this variability as a nuisance to be suppressed or removed. This paper asks whether acquisition variability can instead be explicitly modeled and controlled by leveraging the DICOM metadata that accompanies every clinical scan.

The authors propose MaRaI (Multimodal Acquisition-aware Radiology AI), a two-part framework. First, they use MR-CLIP to align MRI volumes with natural-language prompts derived from DICOM metadata, creating a shared embedding space that captures contrast as a structured variable. Second, they use DIST-CLIP to factorize each scan into an acquisition-invariant anatomical map and a contrast-specific representation. They also introduce a 'Style Fusion Decoder' that uses cross-attention and Adaptive Instance Normalization (AdaIN) to perform anatomy-preserving image harmonisation, allowing for the synthesis of images across different protocols using either a reference scan or raw metadata.

The study demonstrates that explicit disentanglement of anatomy from contrast provides measurable clinical benefits. Anatomical maps derived from the model show significantly reduced inter-sequence volume variability compared to raw images, and they improve the performance of Alzheimer’s disease classification models when tested across different scanner vendors. Furthermore, the unified harmonisation model outperforms existing state-of-the-art methods in both pixel-level fidelity (SSIM/PSNR) and the preservation of anatomical structures, as validated by segmentation consistency. The model also successfully reduces scanner-specific 'fingerprints' in downstream tasks like brain-age regression while maintaining predictive accuracy.