GRU-Enhanced CNN Accurately Identifies Liver MRI Sequences Without Metadata Across Centers

A deep learning model (ConvNeXt+GRU) accurately identified dynamic phases and sequences in liver MRI across multiple centers, without relying on DICOM metadata. T2-weighted imaging showed greater variability in external testing.

• Retrospective study with internal training/validation and external test sets from three institutions; 13 sequence categories (exact exam count not reported).
• Sequential modeling (GRU or transformer) substantially improved performance over a CNN-only baseline; ConvNeXt+GRU achieved highest examination-level accuracy (exact figure not reported).
• Reduced performance on external data was mainly for T2-weighted sequences due to high inter-institutional variability.

RadPigeon summaries are original and for information only. They are not clinical advice.