VIRTUAL GADOLINIUM: DEEP LEARNING-BASED VIRTUAL CONTRAST MRI SYNTHESIS AND BRAIN TUMOR SEGMENTATION

Abstract

The Gadolinium-Based Contrast Agents have been employed as a significant component of Magnetic Resonance Imaging (MRI) so as to augment the visibility of pathological features. However, the problem of accumulation of gadolinium into the central nervous system and hypothetically undesirable effects has prompted the desire to explore other methodologies. The study produces a Deep Learning-based model that is able to either produce a Gadolinium-Enhanced T1 -weighted MRI image from unenhanced T1 -weighted images (without the use of exogenous contrast administration).  In order to represent the reality as closely as possible, the complicated nature of the transformation between unenhanced and contrast-enhanced representation, we assume the most current U-Net architecture with residual modeling. In the training database, more than a hundred cases of patients, including the presence of the paired axial T1pre (unenhanced) and T1post (contrast -enhanced) slices, were included. Using preprocessing, the 3-dimensional NIfTI volumes are transformed into 2-dimensional slices, normalised intensity is performed, and other data augmentation transformations are applied.  A composite loss, which was estimated to have taken into account both the Structural Similarity Index and mean absolute error (L1 loss), was used to optimise the network.  The encouraging scores on another sample of the tests were with a mean Peak Signal to Noise Ratio at 32.70 dBA and SSIM of 0.9540.  The proposed synthetic contrast protocol, in turn, introduces Gadolinium-Free Imaging and a less expensive alternative to Gadolinium-based enhancement without biasing diagnostic fidelity, and, hopefully, it might also prove helpful in lowering patient exposure to contrast agents and associated risks.