Transformer networks enable fast and robust dictionary generation for multiparametric cardiac mapping with variable timing.
Authors
Affiliations (3)
Affiliations (3)
- Department of Radiology, Lausanne University Hospital (CHUV) and University of Lausanne (UNIL), Lausanne, Switzerland.
- Cardiology Service, Cardiovascular Department, Lausanne University Hospital (CHUV) and University of Lausanne (UNIL), Lausanne, Switzerland.
- Department of Radiology, Lausanne University Hospital (CHUV) and University of Lausanne (UNIL), Lausanne, Switzerland. Electronic address: [email protected].
Abstract
Cardiac multiparametric mapping often relies on dictionary matching. When the acquisition timing varies due to ECG or navigator gating, dictionaries must be simulated for each acquisition, which is time-consuming and limits their in-line clinical utility. Deep learning can be used to accelerate this dictionary generation. In this work, we implement and evaluate a transformer network that leverages self-attention to capture long-range dependencies and compare its performance against a previously proposed fully connected multi-layer perceptron (FC-MLP). Transformer (6 layers, 700,674 trainable parameters) and FC-MLP (3 hidden layers, 300-300-75 units) networks were implemented, optimized, and trained on synthetic timing intervals generated through extended phase graph (EPG) simulation. Network performance was quantified as mean average percent error (MAPE) versus EPG simulations in silico, as a correlation in a mapping phantom and through Bland-Atman analysis in 138 patients (age 67±13y, 50F) at 3T. Inference time was recorded. Finally, the full mapping pipeline was integrated directly on the MR (magnetic resonance) scanner. EPG dictionary simulations required 30±2minutes per subject, whereas the transformer completed inference in 13.2±0.4s on CPU and 89.2±0.4ms on GPU, representing a >100-fold acceleration. In the phantom experiments, the transformer achieved near-ideal correlation with the EPG dictionary (slope 1.00 vs 1.02 for FC-MLP) and demonstrated superior accuracy in Bland-Altman analysis, with smaller bias (1.8ms vs 13.6ms) and narrower limits of agreement. Across patient datasets, the transformer exhibited tighter error distributions for T<sub>1</sub> and T<sub>2</sub> (median MAPE [IQR]: 0.37%[0.09-0.76%] and 0.50%[0.25-0.91%] compared to FC-MLP (2.04%[1.61-2.55%] and 3.03%[2.40-3.92%], p<3×10⁻¹⁷² for both), respectively. Correlation between navigator skips and MAPE was also lower for the transformer, indicating greater robustness. In vivo transformer-dictionary maps were visually indistinguishable from EPG-dictionary maps. Bland-Altman analysis in the myocardium confirmed very small biases for the transformer-dictionary maps (-1.70ms for T<sub>1</sub>, 0.23ms for T<sub>2</sub>) versus substantially larger errors for FC-MLP-dictionary maps (18.41ms for T<sub>1</sub>, 1.16ms for T<sub>2</sub>). The transformer network significantly accelerates cardiac multiparametric map reconstruction while maintaining accuracy comparable to EPG simulations and superior to the FC‑MLP. Its robustness to variable acquisition timing and minimal bias support clinical feasibility, and successful on‑scanner integration demonstrates the potential for routine in‑line deployment of quantitative T<sub>1</sub>-T<sub>2</sub> mapping.