HSGDNet: Hybrid Synthetic-Data-Guided Deep Learning With NLS Refinement for Fast Multi-Component T1ρ Knee Mapping

Multi-component T1ρ mapping of the knee joint using nonlinear least squares (NLS)-based methods is usually a computationally intensive task, limiting its application to only a few voxels in the knee joint. Deep learning (DL) is a computationally fast alternative, but requires a large amount of training data. We propose the Synthetic data-Guided supervised DL Network (SGDNet) that utilizes synthetically generated data for training, eliminating the need for large datasets of T1ρ maps. Initially, residual connections are added to improve gradient flow and stabilize training. A self-attention module is also integrated into the SGDNet to obtain more accurate estimated relaxation maps. Additionally, to ensure both parameter fidelity and data consistency, we employ a customized loss function that penalizes discrepancies between actual and predicted T1ρ values as well as between measured and simulated MR signals. To combine speed and precision, we further introduce HSGDNet, a hybrid approach that uses SGDNet's outputs as initialization for a few NLS iterations. Extensive experimental analysis reveals that HSGDNet outperforms the competing methods by achieving average error reductions of 91.4%, 31.5%, and 36.0% for mono-exponential (ME), stretched-exponential (SE), and bi-exponential (BE) components, respectively. HSGDNet accelerates whole-knee T1ρ fitting over NLS by approximately 67.4 × for ME, 53.9 × for SE, and 42.3 × for BE. Finally, to evaluate robustness under pathological and protocol variations, we validate HSGDNet on an early osteoarthritis (EOA) dataset acquired with distinct spin-lock times (TSLs) values. Overall, HSGDNet establishes itself as an efficient method for rapid, precise, and robust multi-component T1ρ mapping in the knee joint.