Introduction

Quantitative T1/T2/ADC mapping allow characterization of subtle changes in tissue biomarkers, and are promising for early detection, discrimination of benign and malignant tumors, and prediction of treatment response in clinical breast MRI^1-5. However, quantitative multi-parametric MRI is typically performed in separate 2D acquisitions, which is time-consuming and may suffer from inter-scan motion and image distortion. We previously developed a whole-brain simultaneous T1/T2/ADC quantification approach using the Multitasking framework⁶. In breast MRI, however, the significant nonuniformity of the B1+ field can result in inaccurate T1 measurements^7-8. We extended the previously proposed Multitasking framework to achieve rapid, simultaneous whole-breast T1/T2/ADC quantification, incorporating a separately acquired B1+ map to compensate for the nonuniform B1+ variation. The agreement with reference approaches/literature values and measurement repeatability are evaluated in healthy subjects.

Methods

Pulse Sequence and Data Acquisition: The pulse sequence and data acquisition are similar to what we previously proposed for whole-brain T1/T2/ADC quantification⁶. Specifically, T1/T2/diffusion contrasts were generated by cycling through a series of 4 T2-preparations (durations $$$\tau$$$=14, 31, 50, 68ms) and 3 diffusion-preparations ($$$b$$$=800s/mm², diffusion directions $$$d$$$=[1 1 -1],[1 -1 1],[-1 1 1]) (Fig. 1A). The T2-preparation with $$$\tau$$$=31ms had the same duration as diffusion-preparations, acting as $$$b$$$=0 module. Data acquisition was performed using 3D segmented FLASH readouts (flip angle=5$$$^{\circ}$$$, TR/TE=11.1/5.8ms), resulting in the signal equation:$$S_{n}=\frac{1}{2}\cdot A\cdot(e^{-\frac{TR}{T1}}\cos(\alpha))^{n}\cdot e^{-\frac{\tau }{T2}}\cdot e^{-bD}\cdot\sin(\alpha),$$where $$$A$$$ absorbs proton density, overall B1 receive sensitivity, and T2* weighting, $$$n$$$ is the readout index, $$$\alpha$$$ denotes FLASH flip angle, and $$$D$$$ represents the diffusion coefficient associated with $$$d$$$. Imaging data ($$$\mathbf{d}_{\text{img}}$$$) were sampled with Gaussian variable density along $$$\mathbf{k}_{y}$$$ and $$$\mathbf{k}_{z}$$$, while training data ($$$\mathbf{d}_{\text{tr}}$$$) were sampled every 8 readouts at k-space center (Fig. 1B).

Image Reconstruction: We represent the underlying multidimensional image as a 5-way low-rank tensor $$$\mathcal{X}$$$ with the first dimension concatenating all voxel locations $$$\mathbf{r}=(x,y,z)$$$ and four dimensions indexing four time/parameter variables $$$n,\tau, b, d$$$ respectively. The image model is expressed as $$$\mathbf{X}_{(1)}=\mathbf{U}\mathbf{\Phi}$$$ where $$$\mathbf{X}_{(1)}$$$ denotes mode-1 matricization of $$$\mathcal{X}$$$, $$$\mathbf{U}$$$ contains spatial coefficient maps, and $$$\mathbf{\Phi}$$$ contains separable multidimensional temporal basis functions describing the dynamic processes of the four time/parameter dimensions⁹. A time-resolved phase map $$$\mathbf{P}$$$ is used to compensate for inter- and intra-shot phase inconsistencies⁶.
We first estimate $$$\mathbf{P}$$$ from a “real-time” reconstruction with only one time dimension (elapsed time $$$t$$$) mixing T1 decay, T2 relaxation, and diffusion contrasts:$$\mathbf{P}=\angle\mathbf{I}_{0},\text{with}\ \mathbf{I}_{0}=\mathbf{U}_{\text{rt,0}}\mathbf{\Phi}_{\text{rt,0}},\text{where}\ \mathbf{U}_{\text{rt,0}}=\arg\min_{\mathbf{U}_{\text{rt,0}}}\left\|\mathbf{d}_{\text{img}}-E(\mathbf{U}_{\text{rt,0}}\mathbf{\Phi}_{\text{rt,0}})\right\|^{2},$$where the real-time temporal subspace $$$\mathbf{\Phi}_{\text{rt,0}}$$$ is extracted from SVD of $$$\mathbf{d}_{\text{tr}}$$$, and $$$E(\cdot)$$$ combines spatial encoding and sampling.

Secondly, $$$\mathbf{\Phi}$$$ is determined via Bloch-constrained low-rank tensor completion of the features extracted from the phase-free $$$\left | \mathbf{I}_{0} \right |$$$ and reorganized as a training tensor⁶, followed by high-order SVD⁹ of the completed training tensor.

Lastly, $$$\mathbf{U}$$$ is recovered by incorporating $$$\mathbf{P}$$$ and $$$\mathbf{\Phi}$$$ into the forward model¹⁰ and solving:$$\mathbf{U}=\arg\min_{\mathbf{U}}\left\|\mathbf{d}_{\text{img}}-E(\mathbf{P}\circ(\mathbf{U}\mathbf{\Phi}))\right\|^{2}+R_{s}(\mathbf{U}),$$where $$$R_{s}(\cdot)$$$ performs spatial total variation regularization.

Experiment Design: Data were collected on a 3T scanner (Vida, Siemens) on a homemade T1 phantom¹¹ and n=13 healthy subjects. In phantom, inversion-recovery-TSE (IR-TSE) images were collected for reference T1 mapping to validate our B1+-compensated Multitasking T1 mapping. In vivo protocols included fat-saturated T1-weighted and T2-STIR imaging, along with multi-echo-spin-echo for reference T2 mapping, and diffusion-weighted readout-segmented-EPI for reference ADC mapping. There is no widely accepted T1 mapping method in breast MRI. Reference B1+ maps were separately acquired for both in vitro and in vivo using a turbo-FLASH sequence preceded by a pre-saturation pulse with prescribed flip angle 80$$$^{\circ}$$$¹² (scan time=25s). The Multitasking sequence (scan time=8min) was run twice to test the repeatability for in vivo study. Imaging parameters were: FOV=240x240x140mm³, resolution=1.5x1.5x5mm³.

Image Analysis: Multi-parametric fitting of $$$A$$$, $$$\alpha$$$, T1, T2, and diffusion coefficients of 3 directions was performed. The B1+ map was incorporated as the initial guess of $$$\alpha$$$ to compensate for B1+ field variation. Intra-class correlation coefficients (ICC) were compared between Multitasking and references for both phantom and in vivo studies.

Results

With B1+ compensation, phantom T1 measurements agrees substantially better ($$$R^{2}$$$=0.973, ICC=0.99) with reference T1 measurements than without B1+ compensation ($$$R^{2}$$$=0.503, ICC=0.72) (Fig.2). High quality, co-registered breast T1/T2/ADC maps are presented, where T2/ADC are consistent with the references and uniform T1 is generated with B1+ compensation (Fig.3). T1/T2/ADC measurements show good repeatability between the first and second Multitasking experiment (Fig.4). Multitasking T1/T2/ADC population measurements are shown in Table 1, where T2/ADC show excellent agreement with reference values (ICC>0.94), and all three measurements fall within the literature range^13-17.

Discussion

High quality, co-registered quantitative T1/T2/ADC maps were generated with the Multitasking framework. The substantial T1 variation caused by the nonuniformity of the B1+ field in the breast coil was compensated, resulting in uniform T1 mapping of the breast. T1/T2/ADC measurements agreed with literature values and reference methods where available.

Conclusion

We presented simultaneous quantification of whole-breast T1/T2/ADC with the MR Multitasking framework in 8.5min, incorporating a separate B1+ map to compensate for the nonuniform B1+ field commonly occurred in breast MRI scans. Co-registered, distortion-free T1/T2/ADC maps are provided, and the values are repeatable and agree well with the references and literature values. Clinical validations will be conducted in future work. The proposed method has the potential to provide imaging biomarkers for comprehensive tumor characterization, early detection, staging, and treatment monitoring of breast cancer.

Acknowledgements

This work was supported by NIH 1R01EB028146.