MAMA SYNTH Pix2PixHD Baseline Debug


Algorithm Logo

About

Editor:
Contact email:
Image Version:
9af7e9c3-bdca-48c5-b3f2-6c2d6c4b0bd4 — May 30, 2026
Model Version:
95f27ff8-5cc2-4516-8818-f597893f5bf5 — May 28, 2026

Summary

Pix2PixHD GAN Baseline for virtual contrast enhancement in breast DCE-MRI. The algorithm synthesizes post-contrast breast MRI slices from pre-contrast inputs using a conditional generative adversarial network (Pix2PixHD, medigan model 00023).

Mechanism

  • Target population: Women undergoing breast DCE-MRI for cancer screening, diagnosis, or treatment monitoring. The algorithm aims to reduce or eliminate the need for gadolinium-based contrast agents by synthesizing virtual post-contrast images from pre-contrast acquisitions.

  • Algorithm description: The algorithm employs a Pix2PixHD conditional generative adversarial network (cGAN) to perform image-to-image translation from pre-contrast to post-contrast breast DCE-MRI. The generator uses a global architecture with 9 residual blocks and instance normalization, trained with a multi-scale discriminator, feature matching loss, and VGG perceptual loss on paired 2D slices from the MAMA-MIA dataset.

    Since the model was trained on 8-bit grayscale images, the inference pipeline bridges the z-score normalized challenge inputs to the model's expected distribution by inverting the normalization (using dataset-level pre-contrast statistics: mean=107.41, std=219.96), applying per-image min-max scaling to [0, 255], running the generator, and then converting the output back to z-score space. All spatial resizing (to/from the model's native 512×512 resolution) is handled via bicubic interpolation.

  • Inputs and Outputs: The input is a single 2D pre-contrast breast DCE-MRI slice (.mha, z-score normalized float32). The output is a synthetic post-contrast slice in the same format and spatial dimensions. The algorithm additionally handles raw uint16 inputs and 3D volumes (by selecting the middle slice) for robustness against format variations.


Interfaces

This algorithm implements all of the following input-output combinations:

Inputs Outputs
1
    Pre-Contrast DCE-MRI Slice Breast
    Synthetic-Contrast DCE-MRI Slice Breast

Validation and Performance


Challenge Performance

Date Challenge Phase Rank
May 31, 2026 mamasynth Debug Phase 10

Uses and Directions

This algorithm was developed for research purposes only.

Warnings

Common Error Messages

Information on this algorithm has been provided by the Algorithm Editors, following the Model Facts labels guidelines from Sendak, M.P., Gao, M., Brajer, N. et al. Presenting machine learning model information to clinical end users with model facts labels. npj Digit. Med. 3, 41 (2020). 10.1038/s41746-020-0253-3