Ensemble Averaging to Improve the Precision and Repeatability of Deep Image Prior Lung MR Fingerprinting at 0.55T

Background: Prior work has demonstrated the ability to perform simultaneous lung T₁, T₂, and M₀ mapping with MR Fingerprinting at 0.55T using a Deep Image Prior (DIP) reconstruction to suppress noise.^1,2,3 However, the stochastic nature of network initialization and training can produce inconsistent tissue property values when the same data are reconstructed multiple times. This study proposes using ensemble averaging to improve the precision and repeatability of DIP for 0.55T lung MRF.

Methods: MRF data were acquired using a 2D spiral FISP sequence⁴ during a 15s breathhold.² Maps were reconstructed using a DIP method that combines a low-rank signal approximation with zero-shot deep learning.³ A U-Net, which takes a fixed tensor initialized with random values as input, generates subspace images that are passed to a fully-connected network to estimate quantitative maps. Both networks are trained de novo using a loss function to enforce consistency with acquired k-space data. To reduce variability arising from random network initialization and stochastic optimization during training, an ensemble averaging strategy is proposed, whereby reconstructions are repeated multiple times, and the final maps are obtained by averaging the results from each run.

This approach was evaluated in 15 healthy subjects at 0.55T (MAGNETOM Free.Max, Siemens) with MRF performed during inspiratory and expiratory breathholds. Repeatability was assessed in one subject by reconstructing the same dataset 10 times and assessing variability in mean parenchymal T₁, T₂, and M₀ values. Next, a leave-one-out scheme was used to assess the impact of ensemble averaging: (1) maps from 9 runs were averaged and T₁, T₂, and M₀ values were recorded, (2) this process was repeated 10 times, each time leaving out a different run, and (3) variability in the resulting measurements was calculated. Finally, the effect of ensemble averaging on the distribution of tissue property measurements across all 15 subjects was evaluated.

Results: Repeated reconstructions of the same dataset are shown in Fig. 1A, where lung tissue property measurements varied from 868-972 ms for T₁ and 43-50 ms for T₂. Ensemble averaging improved precision, as indicated by the lower standard deviations (SDs) (Fig. 1B). Ensemble averaging improved the repeatability of reconstructions, with yielding SDs lower by nearly an order of magnitude (Fig. 2) and led to lower variability in mean T₁ and T₂ values across all subjects (Fig. 3). Significantly higher T₁ and M₀ were observed in expiration vs. inspiration both with and without ensemble averaging. This approach comes at the expense of increased computation time (40 min vs. 6.7 hours).

Conclusion: Ensemble averaging is an effective strategy to reduce variability in 0.55T lung MRF measurements reconstructed using DIP. While computationally expensive, ongoing work is being conducted to accelerate the reconstruction using transfer learning.

Figure 1. Representative expiratory-phase 0.55T lung tissue property maps (1.8x1.8x8 mm3) are shown in one healthy subject from (A) 3 repetitions of the DIP reconstruction (B) averaging the maps from the 10 repeated reconstructions. The mean ± standard deviation of T1, T2, and M0 within the lung parenchyma is indicated on each map.
Figure 2. The variability of repeated DIP reconstructions with and without ensemble averaging was assessed in one subject for MRF scans collected during inspiratory and expiratory breathholds. Mean T1, T2, and M0 values are shown from 10 repeated reconstructions of the same data both (A) without and (B) with ensemble averaging. Mean ± standard deviation for each group is labeled.
Figure 3. Distribution of T1, T2, and M0 values over 15 subjects using (A) a single repetition of DIP and (B) with ensemble averaging of 10 DIP runs. T1 was significantly higher in expiration compared to inspiration using a paired t-test, both with a single DIP run (p=3.1×10^-4) and with ensemble averaging (p=9.9×10^-6). Similarly, M0 was significantly higher in expiration with a single DIP run (p=3.2×10^-7) and with ensemble averaging (p=1.0×10^-6).