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

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Medical Image Analysis
• Computational Biology

Background:

• Multi-component T1ρ mapping of the knee joint is crucial for diagnosing conditions like osteoarthritis.
• Traditional nonlinear least squares (NLS) methods are computationally intensive, limiting their clinical applicability.
• Deep learning (DL) offers speed but typically requires extensive training datasets.

Purpose of the Study:

• To develop an efficient and accurate method for multi-component T1ρ mapping of the knee joint.
• To overcome the limitations of computational intensity in NLS methods and data requirements in DL.
• To introduce a hybrid deep learning approach for accelerated and precise T1ρ quantification.

Main Methods:

• Proposed Synthetic data-Guided supervised Deep Learning Network (SGDNet) using synthetically generated data for training.
• Integrated residual connections and a self-attention module into SGDNet for improved gradient flow and accuracy.
• Developed a hybrid approach (HSGDNet) combining SGDNet outputs with NLS for enhanced precision and speed.
• Employed a customized loss function to ensure parameter fidelity and data consistency.

Main Results:

• HSGDNet achieved significant average error reductions: 91.4% (ME), 31.5% (SE), and 36.0% (BE).
• HSGDNet accelerated T1ρ fitting by approximately 67.4× (ME), 53.9× (SE), and 42.3× (BE) compared to NLS.
• Validated HSGDNet on an early osteoarthritis (EOA) dataset, demonstrating robustness under pathological and protocol variations.

Conclusions:

• HSGDNet provides a rapid, precise, and robust solution for multi-component T1ρ mapping in the knee joint.
• The synthetic data-driven approach eliminates the need for large experimental datasets, facilitating DL model training.
• HSGDNet shows potential for improved clinical diagnosis and monitoring of knee joint pathologies.