DCE-Diff: Diffusion Model for Synthesis of Early and Late Dynamic Contrast-Enhanced MR Images from Non-Contrast Multimodal Inputs

Abstract

Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) is pivotal in delineating abnormal lesions and cancerous regions in the anatomy of interest. However, DCE-MRI requires the injection of gadolinium (Gad)based contrast agents during acquisition which is known to have potential toxic effects, posing radiological concerns. Previous deep learning models employed for synthesizing DCE-MRI images consider unimodal structural MRI inputs lacking information about perfusion or perform early to late response predictions requiring Gad-based MRI sequences as input to drive the synthesis. In this work, we consider the heterogeneity in (i) the multimodal MRI structural inputs offering diverse and complementary anatomical features, (ii) the scanner settings and acquisition parameters, and (iii) the importance of incorporating the perfusion information in Apparent Diffusion Coefficient (ADC) data, which is essential to learn the hyperintense features for DCE-MRI synthesis. We propose DCE-diff, a deep generative diffusion model for multimodal image-to-image mapping from non-contrast structural MRI sequences and ADC maps to synthesize early and late response DCE-MRI images to circumvent Gad contrast injection to patients. Comparative studies using ProstateX and Prostate-MRI datasets against previous methods show that our model demonstrates (i) better synthesis quality with improvement margins of +0.85 dB in PSNR, +0.04 in SSIM, -22.8 in FID, and -0.02 in MAE (ii) better adaptability to different scanner data with deviated settings, showcasing a +8.7 dB improvement in PSNR, +0.22 in SSIM, -40.4 in FID, and -0.1 in MAE, and (iii) the importance of ADC maps in the DCE-MRI synthesis.