Synthesising Brain Iron Maps from Quantitative Magnetic Resonance Images Using Interpretable Generative Adversarial Networks

Abstract

Accurate spatial estimation of brain iron concentration in-vivo is vital to elucidate the role of iron in neurodegenerative diseases, among other applications. However, ground truth quantitative iron maps of the brain can only be acquired post-mortem from ex-vivo samples. Quantitative magnetic resonance imaging (QMRI) methods are iron-sensitive and hold potential to quantitatively measure brain iron. We hypothesise interpretability methods can identify the most salient QMRI parameter(s) for iron prediction. In this study, a generative adversarial network with spatially adaptive normalisation layers (SPADE) was trained to synthesise maps of brain iron content from QMRI parameters, including those from relaxometry, diffusion and magnetisation transfer MRI. Ground truth maps of iron content were obtained by synchrotron radiation X-ray fluorescence (SRXRF). QMRI and SRXRF datasets were registered, and a distribution-based loss was proposed to address misalignment from multi-modal QMRI-to-SRXRF registration. To enable interpretation, channel attention was incorporated to learn feature importance for QMRI parameters. Attention weights were compared against occlusion and local interpretable model-agnostic explanations. Our model achieved dice scores of 0.97 and 0.95 for grey and white matter, respectively, when comparing tissue boundaries of synthesised vs. MRI images. Examining the contrast in predicted vs. ground truth iron maps, our model achieved 15.2% and 17.8% normalised absolute error for grey and white matter, respectively. All three interpretable methods ranked fractional anisotropy as the most salient, followed by myelin water fraction and magnetisation transfer ratio. The co-location of iron and myelin may explain the finding that myelin-related QMRI parameters are strong predictors of iron.