CINet: Communication-aware Inter-modality Network for Distributed Multi-Modal MRI Reconstruction

Abstract AI-enabled medical imaging is increasingly moving toward distributed acquisition, computing, and inference, where information required by a downstream task may reside at different nodes. In multi-modal magnetic resonance imaging (MRI) reconstruction, this situation arises when a fully sampled auxiliary contrast is not locally available but can provide complementary information for reconstructing an undersampled target contrast. Existing methods generally assume direct access to the auxiliary modality and therefore do not consider the communication constraints involved in transferring such information between distributed nodes. To address this problem, we propose the Communication-aware Inter-modality Network (CINet) for distributed multi-modal MRI reconstruction, which jointly considers auxiliary representation transmission and cross-modal reconstruction under communication constraints. Specifically, a Task-Oriented Communication Encoder reweights multi-scale auxiliary features and generates compact representations under a prescribed communication budget. The resulting representations are transmitted through a noisy channel, and a Target-Conditioned Selective Fusion module selectively incorporates the received auxiliary information according to the current target representation. CINet thus coordinates compact auxiliary information transmission with its downstream utilization for target reconstruction. Experiments on multi-modal prostate MRI demonstrate that CINet achieves competitive or superior reconstruction performance under different undersampling patterns and acceleration factors while preserving local structural details. Further evaluations under varying communication budgets and channel signal-to-noise ratios characterize the communication--reconstruction trade-off under different transmission conditions.