MIRAGE: Multi-scale Lesion-Informed Representation with Auxiliary Guidance for MRI Contrast Enhancement

2026-07-21 · Source: Artificial Intelligence · Field: Science & Research — Health & Medical Research, Artificial Intelligence & Machine Learning, Image and Video Processing · Depth: Expert, quick

Summary

MIRAGE, a residual 2D U-Net, addresses the underdetermined problem of inferring contrast enhancement from pre-contrast breast MRI slices. Unlike methods optimizing only pixel fidelity or using adversarial objectives, MIRAGE combines global reconstruction and perceptual losses with three lesion-aware supervisions available during training: an asymmetric penalty for missed tumor enhancement, multi-scale auxiliary tumor segmentation, and guidance via a frozen post-contrast tumor segmentation nnU-Net. Evaluated on 301 cases from the multi-centre MAMA-SYNTH data using eight metrics, MIRAGE ranks first on six and significantly improves downstream lesion localization over pix2pix, conditional diffusion, and latent bridge-matching baselines. While generative alternatives show advantages in LPIPS or contrast classification, indicating a fidelity-utility trade-off, ablations confirm distinct effects of losses on appearance, radiomics, and boundary accuracy.

Key takeaway

For AI Scientists or Machine Learning Engineers developing MRI contrast enhancement models, MIRAGE offers a superior approach to tumor localization. Its lesion-aware supervision during training significantly outperforms general generative models. You should consider the inherent fidelity-utility trade-off and carefully select evaluation metrics that align with the specific downstream clinical utility of your synthesized images, rather than relying solely on perceptual or classification scores.

Key insights

MIRAGE enhances MRI contrast by integrating lesion-aware supervision, improving tumor detection while balancing fidelity and utility.

Principles

• Lesion-aware supervision improves MRI contrast enhancement.
• Fidelity-utility trade-offs exist in generative medical imaging.
• Task-aware synthesis optimality is metric-dependent.

Method

MIRAGE uses a residual 2D U-Net with global reconstruction and perceptual losses, augmented by asymmetric tumor enhancement penalties, multi-scale auxiliary segmentation, and nnU-Net guidance during training.

In practice

• Integrate asymmetric penalties for missed tumor enhancement.
• Employ multi-scale auxiliary segmentation for guidance.
• Utilize frozen segmentation models for training supervision.

Topics

• Breast MRI
• Contrast Enhancement
• U-Net
• Lesion Segmentation
• Medical Imaging
• Generative Models

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• MIRAGE: Multi-scale Lesion-Informed Representation with Auxiliary Guidance for MRI Contrast Enhancement — MIRAGE enhances MRI contrast inference by integrating lesion-aware supervision with global reconstruction, improving tumor localization.
• Multi-Contrast MRI Motion Correction via Parameter-Informed Disentanglement and Adaptive Experts — A unified MRI motion correction framework disentangles contrast and adapts to severity for robust, generalizable artifact removal.
• Enhancing Brain MRI Anomaly Detection and Reasoning with ROI Rethink and Synthetic Data — Explicit hypothesis-verification grounding with ROI marking enhances trustworthy brain MRI anomaly detection and reduces hallucination.
• Region-Grounded Vision-Language Learning for Detection-Guided Mammographic Lesion Classification — Region-grounded vision-language learning improves mammographic lesion classification by mirroring radiologists' localized diagnostic approach.
• Seeing What Matters: Lesion-Aware High-Resolution Patch Discovery and Fusion for Chest X-ray Report Generation — LePaX enables efficient high-resolution CXR perception for RRG by adaptively allocating resolution and fusing regional evidence without token inflation.
• Prediction of microvascular invasion in hepatocellular carcinoma using contrast-enhanced ultrasound and deep learning — MAPUSE non-invasively predicts hepatocellular carcinoma microvascular invasion using deep learning on CEUS, linking to immune infiltration.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.