Various substances in the human body, such as iron calcium, and lipids, can alter magnetic susceptibility (Deistung et al., 2017). Abnormal changes in their content are often associated with the progression of specific diseases. In recent years, Quantitative susceptibility mapping (QSM) has been widely applied in clinical disease research, including neurodegenerative diseases (Shahmaei et al., 2019; Guan et al., 2017), vascular diseases (Probst et al., 2021; Hong et al., 2023), neuroinflammation (Chen et al., 2014), and traumatic brain injury (Chary et al., 2021). For example, in Parkinson’s disease, QSM values in the substantia nigra can be used to differentiate between patients and controls (Shahmaei et al., 2019; Guan et al., 2017), and QSM images can assist in the preoperative localization of deep brain nuclei (Cong et al., 2020). In Alzheimer’s disease, QSM can reflect pathological iron deposition associated with amyloid protein (Gong et al., 2019).

In conventional and clinical research applications, susceptibility-weighted imaging mainly uses three-dimensional GRE sequences. To achieve a strong susceptibility effect, a relatively long echo time is required to establish phase contrast (usually >30 ms for a 3 T MRI scanner) in QSM imaging, and multiple echoes are typically acquired to compute susceptibility (Shahmaei et al., 2019; Guan et al., 2017; QSM Consensus Organization Committee et al., 2024). For visualizing fine anatomical structures (such as small venules and cortical structures) and small lesions (such as microbleeds) (Rotta et al., 2021), a high spatial resolution is also required (Shen et al., 2020). A recent consensus paper suggests using isotropic acquisition protocols to avoid partial volume effect (QSM Consensus Organization Committee et al., 2024). The combination of these various demands can significantly extend acquisition times, thereby reducing patient compliance.

In recent years, deep learning (DL) methods have gained widespread attention and application in the field of magnetic resonance imaging (MRI), being utilized in various directions such as accelerating image acquisition, reducing gadolinium dose (Gong et al., 2018), and improving image resolution (Avanzo et al., 2020; Johnson et al., 2023; Almansour et al., 2023). In terms of accelerating MRI acquisition, DL methods are primarily used to recover k-space information from undersampled k-space data or to directly reconstruct images (Han et al., 2020; Wang et al., 2021). They can also enhance image quality based on low-quality images reconstructed from undersampled data (Bash et al., 2021). Previously, Gao (Gao et al., 2021) et al. trained a Deep Complex Residual Network (DCRNet) using 7 T MRI data. They tested the model on several retrospectively under-sampled datasets and one prospectively under-sampled dataset, demonstrating substantially reduced artifacts and blurring compared to two iterative methods and one deep learning method. Similarly, Zhang (Zhang et al., 2023) et al. developed a framework, Learned Acquisition and Reconstruction Optimization (LARO), designed to accelerate multi-echo GRE pulse sequences for QSM. The authors optimized the k-space sampling pattern and employed a recurrent temporal feature fusion model to capture signal redundancies along echo times. Their methods were tested on prospectively under-sampled k-space datasets from 10 healthy subjects, achieving an acceleration factor of 8 while maintaining QSM quality. While these studies demonstrated the feasibility of applying DL methods to QSM acquisition, they primarily focused on methodology development. The generalizability and stability of DL-based QSM remain untested in larger sample sizes.

In this study, we aimed to test whether a DL-based reconstruction method could be used for accelerating QSM acquisition without significant impact on the measurement of susceptibility in deep brain nuclei. Specifically, we investigated the overall image similarity between PI and DL-QSM images using the structural similarity (SSIM) and peak signal-to-noise ratio (PSNR) indices. We tested the correlation and differences between different sets of QSM images. At last, we investigated whether the acceleration would affect the correlation analyses between age and nuclei QSM values.

A total of 59 participants were enrolled in this study (Table 1, mean age: 44.0 y/o; range: 18–75 y/o; male/female: 22/37). Using PI-QSM images as reference, the mean SSIM of DL-QSM images were 0.87 (min-max: 0.82–0.89), 0.86 (0.77–0.89), and 0.85 (0.81–0.89) for acceleration factors of 3, 4, and 5, and respective mean PSNR were 44.56 (37.31–52.12), 44.53 (37.42–53.10), and 44.23 (37.35–52.91).

As shown in Figure 2, under different acceleration factors, DL-QSM exhibits a similar overall appearance compared to PI images. Figure 3 shows imaging details in the basal ganglia and brainstem regions. In general, the shape and local patterns of deep brain nuclei were mostly preserved, with a tendency toward smoother images under higher acceleration conditions.

As shown in Table 2, the susceptibility values of the seven deep brain structures in each DL-QSM group have a very high correlation with the PI-QSM group (r > 0.95, p < 0.001). In the PI-QSM and DL-QSM-AF3 groups, the GPe showed the highest correlation (r = 0.994, p < 0.001), while the RN showed the lowest correlation (r = 0.963, p < 0.001). In the PI-QSM and DL-QSM-AF4 groups, the GPe also showed the highest correlation (r = 0.993, p < 0.001), while the SNr showed the lowest correlation (r = 0.950, p < 0.001). In the PI-QSM and DL-QSM-AF5 groups, the GPe again showed the highest correlation (r = 0.993, p < 0.001), while the Ca showed the lowest correlation (r = 0.950, p < 0.001). Overall, among the PI-QSM group and all DL-QSM groups, the GPe showed the highest susceptibility value correlation.

On the group level (Table 3), the region with the greatest susceptibility difference between the PI-QSM and DL-QSM-AF3 images is the SNr, with a mean difference of 4.19%; followed by the Ca, with a mean difference of 3.31%. The greatest difference between PI-QSM and DL-QSM-AF4 is in the Ca, at 2.98%; while the greatest difference between PI-QSM and DL-QSM-AF5 is in the Pu, at 3.15%. There was no systematic bias between the values from different image groups.

As shown on the Bland–Altman plot (Figure 4), the differences between PI and DL-QSM images distributed evenly across different susceptibility levels. For putamen and caudate, differences between the PI and DL images were mostly less than 5 ppb. For other nuclei, the difference was higher, reaching over 10 ppb. When looking into the source of these discrepancies, we noticed the influence of smoothness on local patterns in the QSM images (Figure 5). In some cases, mild head motion and imaging artifacts might also contribute to the differences (Table 4).

In the PI-QSM images, the susceptibility values of the Pu, Ca, SNc, and RN are positively correlated with age and have statistical significance, with the Pu showing the strongest correlation with age (r = 0.631, p < 0.001). In the other three DL-QSM image groups, the susceptibility values of the Pu, SNc, and RN also show significant positive correlations with age, with the Pu showing the strongest correlation with age (AF-3: r = 0.644, p < 0.001; AF-4: r = 0.653, p < 0.001; AF-5: r = 0.658, p < 0.001). The Ca shows a significant correlation with age in the DL-QSM-AF3 images (r = 0.284, p = 0.029) and a marginally significant positive correlation with age in the DL-QSM-AF4 (r = 0.224, p = 0.088) and DL-QSM-AF5 images (r = 0.247, p = 0.060).

This study showed that DL-QSM images with acceleration factors of 3, 4, and 5 had good similarity to PI-QSM images, with high correlation in the measured susceptibility values of nuclei (r > 0.95). When the acceleration factor was 5, the required scanning time was only one-third of the PI-QSM acquisition time, greatly improving scanning efficiency. Although there were differences in some nuclei between different QSM images, these differences were usually less than 5% on the group level. Such small variations do not significantly affect the correlation analyses between age and QSM values.

Under the condition of 5-fold acceleration, we did not observe significant differences in the overall image appearance and nuclei structures. It should be noted that as the k-space signal sampling decreased, the images appeared slightly smoother, which might lead to slight confusion of signals in adjacent nuclei or white matter areas. The smoothing might have affected the voxel values at the nuclei boundaries and causing fluctuations in the susceptibility values of some nuclei regions. However, the differences in susceptibility values of these nuclei are general small on the group level.

On the individual level, larger differences might occur in some participants. Such variations could be due to more complex reasons besides the abovementioned smoothing effect. Head motion between scans could change the angle between the acquisition plane and the main magnetic field direction, which might alter the susceptibility effect of the dipoles. Small head motion during each QSM acquisition could cause artifacts and affect QSM estimation. There might be other factors such as magnetic field shimming, temperature changes, and various imaging artifacts. One study scanned phantoms containing five different concentrations of gadolinium solutions using 12 clinical and 3 preclinical scanners (Deh et al., 2019). At the level of 0.26 ppm, the standard deviation of the measurements was 32 ppb. In a replication study that scanned 14 participant four times using an identical scanner and the same protocol, the standard deviation of QSM values in deep brain nuclei across different imaging sessions can reach 19 ppb (Santin et al., 2017). In our study, differences between PI and DL images were at a comparable level (mostly less than 20 ppb), suggesting the influence of undersampling and DL reconstruction could be tolerated.

The basal ganglia and brainstem nuclei are key regions of the brain responsible for processing motor control, emotions, cognition, and various other functions (Groenewegen, 2003; Seger, 2006). Due to metabolic changes, vascular damages, and other factors, significant iron deposition can occur in these nuclei during brain aging (Ashraf et al., 2018). Some researchers have used QSM to study normal populations, showing that the susceptibility values of multiple nuclei in normal people tend to gradually increase with age (Acosta-Cabronero et al., 2016; Li et al., 2023). Consistent with previous results (Acosta-Cabronero et al., 2016), our study also confirmed that the Pu region shows the most significant increase in susceptibility, followed by the RN, SNc, and Ca regions. Furthermore, the correlations between age and DL-QSM values were similar to those with PI-QSM values, suggesting that the acceleration did not induce unexpected impacts to the analyses.

Although the current study preliminarily demonstrates the potential of DL-QSM sequences in magnetic resonance imaging, the robustness and general applicability of the conclusions need to be validated in larger-scale cohort studies due to the limited sample size. Particularly, expanding the sample size for different groups of patients with neurological disorders will help comprehensively evaluate the stability and diagnostic accuracy of DL-QSM sequences under various pathological conditions. Secondly, under the current experimental conditions, using an MR scanner equipped with an NVIDIA GeForce RTX 2080 Ti graphics card, image reconstruction takes several minutes to generate QSM images. Upgrading to newer, more powerful graphics cards could significantly improve image processing speed. Thirdly, while using fully sampled data as a reference is ideal, the long acquisition time may result in a high failure rate in clinical settings. Therefore, we used parallel imaging data as the reference. Since the model performance has already been assessed, and our primary goal was to evaluate the practicality of this method, comparing the image quality of DL-QSM with clinically established methods is considered acceptable. Furthermore, three-dimensional high-resolution QSM images have important value in preoperative nuclei localization, our future research will further investigate these related topics.

In summary, DL-QSM could be used for measuring susceptibility values of deep brain nuclei. An AF up to 5 did not significantly impact the accuracy of QSM results, nor the correlation between age and susceptibility in deep brain nuclei.