Introduction

Susceptibility source separation methods (χ-separation,¹ DECOMPOSE-QSM,² and Emmerich et al.³) have recently been proposed for estimating the individual concentrations of positive susceptibility (χ_pos) and negative susceptibility (χ_neg), which illustrate the distribution of two dominant susceptibility contributors (i.e., iron and myelin) in the brain. Among them, χ-separation utilizes R₂’ (= R₂*-R₂) information that requires the acquisition of R₂* and R₂ maps. Despite its utility, however, this requirement for an R₂ map hampers a wide application of χ-separation due to the lack of an R₂ map in routine clinical imaging or public dataset including UK Biobank (UKB).⁴ Recently, deep learning-based relaxometry that translates contrast-weighted images to relaxometric maps has been suggested.⁵ In this study, we propose two pipelines that allow us to apply χ-separation to UKB data and a routine clinical protocol using R₂ maps from deep neural networks. The performance of the proposed pipelines is evaluated by comparing the results with those using a reference R₂ map.

Methods

[In-vivo data acquisition] UKB data include MPRAGE, T₂-FLAIR, and GRE_3mm (TR/TE1/TE2=27/9.4/19.7ms, resolution=0.8×0.8×3mm³). We scanned 20 subjects (train/validation/test: 10/1/9) using the same parameter as the UKB protocol (IRB-approved). Additionally, 2D multi-echo spin-echo (MESE; TR/TE=6530/20:20:100ms, FOV=208×256×120mm³, resolution=1×1×2mm³) and GRE_1mm (TR/TE1/TE2/TE3/TE4=27/4.3/9.4/14.6/19.7ms, resolution=0.8×0.8×1mm³) were acquired for reference R₂ and R₂* maps.⁶
[χ-separation for UKB protocol] The R₂* map of the UKB protocol datasets (2 echoes, 3mm thickness) suffered from errors and B₀-field inhomogeneity artifacts,⁷ which resulted in inaccurate χ-separation results. Hence, we developed an additional network (R₂* denoising network) that infers R₂* values of GRE_1mm (4 echoes, 1mm thickness) from GRE_3mm data (2 echoes, 3mm thickness). A local field map is generated from GRE_3mm data using the same protocol as the UKB QSM protocol.⁸ For R₂ mapping, two neural networks were trained to generate R₂ maps for two different input types (Fig. 1). The first network (UKB_T1w+FLAIR) was trained using the input of MPRAGE+T₂-FLAIR. The second network (UKB_FLAIR) utilized only T₂-FLAIR as an input. Lastly, the R₂ map can be assumed to have a constant value of mean R₂⁹ (15.5 Hz; UKB_Constant). Using local field and R₂’ maps, iLSQR¹⁰-style χ-separation¹ was conducted to reconstruct χ_pos and χ_neg. For a reference χ-separation map (Reference_UKB), the same χ-separation processing is conducted with the reference R₂ from MESE, R₂* from GRE_1mm, and local frequency maps from GRE_3mm (resampled to 1.05×1×3mm³ resolution if necessary).
[χ-separation for routine clinical protocol] To demonstrate the feasibility of χ-separation application to a routine clinical scan that contains T₂-weighted images and multi-echo GRE images, we utilized MESE and GRE_1mm from our scan. Using the MESE data, T₂-weighted images were synthesized by considering the turbo spin-echo phase encoding. GRE_1mm was resampled to 2mm thickness and used for the generation of local field and R₂* maps. R₂* denoising network was not used. R₂ mapping network with the input of a T₂-weighted image (CL_T2w) was trained to produce a reference R₂ map. Similar to the UKB application, an additional reconstruction with constant R₂ (15.5 Hz; CL_Constant) was performed. The χ-separation with reference R₂ from MESE was also conducted to generate reference χ-separation maps (Reference_CL).
[Deep neural network] All deep neural networks share the same structure of Unet⁵ with fully connected layers on the backend. Training was conducted with an L1 loss between output and reference.
[Evaluation] Nine subjects were utilized for the validation of the proposed methods. The χ_pos and χ_neg maps from UKB_T1w+FLAIR, UKB_FLAIR, and UKB_Constant were compared to those of Reference_UKB, and normalized root-mean-squared error (NRMSE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM) were calculated within a brain mask while masking out CSF. Similarly, the χ-separation results from CL_T2w and CL_Constant are compared to Reference_CL. We also generated group-averaged χ-separation maps from UKB_T1w+FLAIR and Reference_UKB by normalizing them to the MNI space.¹¹

Results

The χ-separation maps of UKB_T1w+FLAIR, UKB_FLAIR, and UKB_Constant (Fig. 2) demonstrate consistent results when compared with those of Reference_UKB. As expected, UKB_T1w+FLAIR shows the best performance (NRMSE = 54.4±4.7% vs 56.3±5.7% for UKB_FLAIR and 60.6±4.3% for UKB_Constant; see Fig. 2b for PSNR and SSIM, same trend) in the reconstruction of χ_pos and χ_neg maps. When the χ-separation results from the clinical routine are evaluated (Fig. 3), CL_T2w shows great reconstruction accuracy when compared to that of Reference_CL (NRMSE = 25.5±3.0% vs 36.0±1.9% for CL_Constant; see Fig. 3b for PSNR and SSIM, same trend). The zoomed-in χ_pos and χ_neg maps (Fig. 4) confirm that UKB_T1w+FLAIR and CL_T2w results show similar structures to those of the references, revealing details of iron and myelin distribution in the brain. Figure 5 demonstrates the χ-separation-averaged templates of Reference_UKB and UKB_T1w+FLAIR, revealing similar susceptibility distribution.

Conclusion and Discussion

In this study, we developed a pipeline for reconstructing χ-separation maps from the UK Biobank protocol dataset and a routine clinical scan dataset by leveraging a deep neural network. For the UKB application, a network using both MPRAGE and T₂-FLAIR demonstrated the best correspondence to the reference data. Moreover, even data with the constant R₂ reported a good match to the reference χ-separation maps. Our results suggest that χ-separation can be applied to existing large-scale data including UKB dataset. Using this method, we will explore full UKB data, potentially benefiting from the large population dataset.