Fig. 5: Ablation studies on Model’s performance.
From: MU-Diff: a mutual learning diffusion model for synthetic MRI with Application for brain lesions
a Visualization of feature maps from the contrast-aware adaptive module on FLAIR, T2, and T1 conditions to synthesize T1CE. b Attentive feature maps (M1, M2) from the critic network applied to predictions from the two generative diffusion networks. c Prediction results from each denoising network (\({{\bf{x}}}^{{p}_{1}}\), \({{\bf{x}}}^{{p}_{2}}\)) and the mutual prediction from MU-Diff. d Quantitative comparison of individual prediction of denoising networks (\({{\mathcal{F}}}_{1}\), \({{\mathcal{F}}}_{2}\)) and mutual prediction from MU-Diff. e Effect of different timestep T in the denoising process. f Ablation on critic network using FID scores. g Ablation results on BraTS Dataset for whole brain regions. h Ablation results on BraTS Dataset for lesion regions. abl1 uses a single network without mutual learning; abl2 uses mutual learning w/o any feature adaptation; abl3 includes ϕ w/o ρ; abl4 w/o mask loss (\({{\mathcal{L}}}_{m}\)). i Qualitative comparison for each ablation.
