MAMA-FLUX.2: Image-to-Image Synthesis of Post-Contrast Breast DCE-MRI for the MAMA-SYNTH Challenge

• Virtual contrast enhancement synthesizes post-contrast breast MRI from pre-contrast acquisitions, potentially reducing gadolinium-related examination complexity and patient burden.Source: §1 · p. 1
• MAMA-SYNTH evaluates peak-enhanced post-contrast image synthesis while requiring plausible enhancement and preservation of patient-specific anatomy.Tumor enhancement is often localized, making global image-level objectives potentially insufficient.Source: §1 · p. 2
• Latent generative modeling, flow matching, and rectified-flow formulations provide an efficient framework for conditional medical image synthesis requiring realism and structural consistency.Source: §1 · p. 2
• The proposed model adapts FLUX.2-Klein-4B by conditioning on the pre-contrast image and predicting the flow field for a noised post-contrast target latent.Source: §1 · p. 2
• A three-component objective combines global latent flow matching, tumor-region supervision, and stable-foreground preservation to discourage unnecessary changes outside tumors.Source: §1 · p. 2
• The study evaluates LoRA rank, LoRA scaling, and regional loss weighting using image-quality and tumor-focused metrics.Source: §1 · p. 2

2.1 Conditional Latent Flow Matching

• Pre-contrast and post-contrast images are encoded into VAE latents for conditional latent rectified-flow synthesis.The model uses the pre-contrast latent as conditioning and reconstructs the synthetic post-contrast image through VAE decoding.Source: §2 · p. 2, §2.1 · p. 3
• At sampled noise level t, the method constructs a noised target latent and trains a transformer to predict its corresponding flow field.The conditioning includes the pre-contrast latent, text embedding, and timestep embedding.Source: §2.1 · p. 3

2.2 Tumor-Aware and Anatomy-Preserving Objective

• The training objective combines standard global flow matching with tumor-region supervision and stable-foreground preservation.The global loss averages squared flow-matching errors over all latent elements, while regional terms target clinically relevant structures.Source: §2.2 · pp. 3–4
• Tumor supervision resizes and broadcasts the tumor mask to latent-error resolution, emphasizing synthesis accuracy within the lesion region.The tumor-region loss is computed using the masked latent errors.Source: §2.2 · p. 4
• Stable-foreground preservation targets anatomy outside the tumor using an image-space mask that suppresses background and excludes strongly enhancing tissue.The thresholds are set empirically to τfg = −0.45 and τenh = 0.5 before resizing and broadcasting the mask.Source: §2.2 · p. 4
• The final objective weights the tumor-region and stable-foreground losses relative to the global loss using λtumor and λstable.These coefficients control the relative contributions of the two regional objectives.Source: §2.2 · p. 4

Data and preprocessing

• Paired pre-contrast and peak-enhanced post-contrast 2D MHA slices from MAMA-SYNTH support regional supervision and evaluation with tumor masks.The preprocessing follows the official protocol and preserves spatial metadata for evaluation.Source: §3.1 · p. 4, §3.2 · p. 4

Anatomical plane routing

• A binary classifier routes each slice to an axial or sagittal LoRA adapter before conditional synthesis with the frozen FLUX.2-Klein-4B backbone.Routing uses image dimensions, intensity statistics, foreground descriptors, and occupancy profiles.Source: Anatomical Plane Routing · p. 5, Figure 1 · p. 5

Model and fine-tuning

• The pipeline encodes the pre-contrast slice into VAE latent space as spatial conditioning while optimizing only attention-projection LoRA adapters.The VAE, text encoder, and pretrained transformer weights remain frozen.Source: FLUX.2 Fine-Tuning · p. 5, Figure 2 · p. 6

Ablation design

• The main axial-slice search varies LoRA rank and scaling, intensity-window upper bounds, and regional loss weights before transferring the selected configuration to sagittal slices.The reported final window uses a z-score range of [−0.51, 25].Source: §3.2 · p. 5, FLUX.2 Fine-Tuning · p. 6

Evaluation protocol

• Inference uses 50 rectified-flow sampling steps with guidance scale 4.0, while Table 1 reports complete ablation configurations as mean ± standard deviation.Runtime profiling covers spatial sizes from 256×256 to 512×512.Source: FLUX.2 Fine-Tuning · p. 6, Table 1 · p. 7

Quantitative ablation

• λtumor = 0.25 and λstable = 0.1 achieve the highest DSC and lowest HD95 in the 10-case development ablation.The analysis evaluates DSC, HD95, tumor SSIM, MSE, and LPIPS.Source: §4 · p. 7
• Increasing stable-foreground weighting improves MSE and LPIPS but degrades segmentation performance when λtumor = 0.25.Source: §4 · p. 7
• Stronger tumor-region weighting improves tumor SSIM but does not provide the best overall balance across metrics.Source: §4 · p. 7

Qualitative results

• The selected model preserves main anatomical structure and produces spatially aligned enhancement patterns in the qualitative examples.The selected configuration uses LoRA rank/α = 64/64, MHAmax = 25, λtumor = 0.25, and λstable = 0.1.Source: Figure 3 · p. 8
• A challenging case shows higher reconstruction error, with discrepancies concentrated around enhancing tissue and tumor-adjacent regions rather than the background.Source: §4 · p. 8

Scope and generalization

• The ablation uses only 10 development cases, so selected hyperparameters may not generalize to the hidden test set or external cohorts.Source: §5 · p. 8

Method constraints

• Slice-wise training does not enforce volumetric consistency, and intensity windows, regional loss weights, and plane routing were selected empirically.Source: §5 · p. 8

Scope and generalization

• The design choices should be validated on larger independent datasets and across different acquisition protocols.Source: §5 · p. 8