Multi-Scale Subtraction Consistency Based Conditional Generative Adversarial Networks for Breast DCE-MRI Synthesis

Sammanfattning

Dynamic contrast-enhanced MRI (DCE-MRI) is widely used for breast cancer diagnosis and treatment monitoring, but repeated injections of gadoliniumbased contrast agents (GBCAs) raise safety, and environmental concerns. This thesis investigates whether contrast-free, multi-sequence breast MRI can be used to synthesize multi-phase DCE-like images under realistic data constraints. We develop a conditional generative adversarial network (cGAN) based on Pix2PixHD that maps pre-contrast T1-weighted images, diffusionweighted imaging (DWI) at multiple b-values, and apparent diffusion coefficient (ADC) maps to several post-contrast phases in a single forward pass. To emphasise enhancement dynamics rather than absolute intensities, we introduce a multi-scale subtraction consistency (MSSC) loss that compares pre-/post-contrast subtraction maps across multiple image scales. Experiments on public breast MRI datasets show that the proposed MSSC model with full multi-sequence input improves standard reconstruction metrics (PSNR, SSIM, LPIPS, lesion-wise RMSE) compared with strong Pix2PixHD baselines. Lesion-level enhancement curves and qualitative assessments further indicate better preservation of tumour morphology and parenchymal texture with fewer artifacts. These results suggest that, given limited public data, a contrast-free, multi-sequence cGAN with scale-aware subtraction supervision can synthesize multi-phase DCE-like breast MRI that is competitive with real DCE-MRI while more faithfully capturing lesion-level enhancement behaviour.