Exploring linearity of deep neural network trained QSM: QSMnet+

Abstract

Recently, deep neural network-powered quantitative susceptibility mapping (QSM), QSMnet, successfully performed ill-conditioned dipole inversion in QSM and generated highquality susceptibility maps. In this paper, the network, which was trained by healthy volunteer data, is evaluated for hemorrhagic lesions that have substantially higher susceptibility than healthy tissues in order to test "linearity" of QSMnet for susceptibility. The results show that QSMnet underestimates susceptibility in hemorrhagic lesions, revealing degraded linearity of the network for the untrained susceptibility range. To overcome this limitation, a data augmentation method is proposed to generalize the network for a wider range of susceptibility.The newly trained network, which is referred to as QSMnet + , is assessed in computer-simulated lesions with an extended susceptibility range (-1.4 ppm to +1.4 ppm) and also in twelve hemorrhagic patients. The simulation results demonstrate improved linearity of QSMnet + over QSMnet (root mean square error of QSMnet + : 0.04 ppm vs. QSMnet: 0.36 ppm). When applied to patient data, QSMnet + maps show less noticeable artifacts to those of conventional QSM maps. Moreover, the susceptibility values of QSMnet + in hemorrhagic lesions are better matched to those of the conventional QSM method than those of QSMnet when analyzed using linear regression (QSMnet + : slope = 1.05, intercept = -0.03, R 2 = 0.93; QSMnet: slope = 0.68, intercept = 0.06, R 2 = 0.86), consolidating improved linearity in QSMnet + . This study demonstrates the importance of the trained data range in deep neural network-powered parametric mapping and suggests the data augmentation approach for generalization of network.The new network can be applicable for a wide range of susceptibility quantification.

xQSM: quantitative susceptibility mapping with octave convolutional and noise‐regularized neural networks

“…Previous works have proposed to solve QSM dipole inversion with deep neural networks, including QSMnet, 26 QSMnet + , 28 QSMGAN, 29 VaNDI, 31 autoQSM 32 and DeepQSM 34 . The first five frameworks require reconstructing QSM first using conventional methods as the training labels, which may deviate from the ground truth.…”

Overview of quantitative susceptibility mapping using deep learning: Current status, challenges and opportunities

“…One example is a recent study by Jung et al, reporting the effects of susceptibility range. 56 In the study, the network was trained using healthy volunteer data, limiting the training range to healthy tissue values (Figure 3a). When the network generated the susceptibility map of a hemorrhagic patient, it produced underestimated susceptibility values in hemorrhagic lesions, which have far higher susceptibility values than those of normal tissue ( Figure 3).…”

“…Similarly, Jung et al showed that data augmentation using susceptibility value scaled maps can cover a wider range of susceptibility values, successfully reconstructing QSM maps for pathology. 56 Different from other fields, which require additional data acquisition for data augmentation, QSM can be benefited from the dipole model in generating augmentation data.…”

SEPIA - SuscEptibility mapping PIpeline tool for phAse images

“…Another focus will be to explore the possibility to incorporate processing methods that are not developed in MATLAB. Emerging techniques using deep learning have already shown very promising results with further improvement in QSM artefact reduction (Bollmann et al, 2019;Chen et al, 2019;Jung et al, 2020;Polak et al, 2020;Wei et al, 2019;Yoon et al, 2018;Zhang et al, 2020). As the implementations of these methods are primarily in Python, while there are rich resources already available in MATLAB and supported in SEPIA to perform tasks such as phase unwrapping and background field removal, it would be valuable for SEPIA to connect the two.…”

“…As the implementations of these methods are primarily in Python, while there are rich resources already available in MATLAB and supported in SEPIA to perform tasks such as phase unwrapping and background field removal, it would be valuable for SEPIA to connect the two. Up to now, most deep learning methods do not yet have the flexibility to cope with arbitrary slice orientation and resolution (Jung et al, 2020;Yoon et al, 2018), this would therefore require data re-sampling to a pre-defined space which can be done with a plethora of software available for neuroimaging data, but we believe this is to be outside the scope of this toolbox. We also consider the compatibility between the SEPIA input and outputs data with BIDS (Gorgolewski et al, 2016), moving toward the standardisation of input and output phase data format and structure.…”