Abstract
Dynamic contrast-enhanced breast magnetic resonance imaging (DCE-MRI) provides clinically valuable lesion information but requires intravenous gadolinium-based contrast administration. The 2026 MAMA-SYNTH Challenge (Synthesizing Virtual Contrast-Enhancement in Breast MRI) asks participants to synthesize a two dimensional peak-enhancement post-contrast slice from the corresponding pre-contrast slice, requiring preservation of global anatomy while estimating tumor and peritumoral enhancement from limited input information.
We formulate the task as enhancement-residual prediction and introduce the Tumor-Aware Residual nnU-Net, implemented with a 2D nnU-Netv2 ResidualEncoderUNet backbone. Training combines image-level reconstruction and structural similarity with a training-only tumor segmentation objective, differentiable tumor-ROI summary-statistic alignment, and per-case tumor-ROI L1. Inference uses only the pre-contrast image. The model was trained on 1,080 cases and evaluated on a disjoint patient-level partition of 426 cases. Compared with identity, mean MSE decreased from 1.3000 to 0.5206, tumor SSIM increased from 0.1294 to 0.4892, Dice increased from 0.1378 to 0.5425, and HD95 decreased from 325.16 to 109.23. Pre/post AUROC increased from 0.5000 to 0.8136, whereas tumor/non-tumor AUROC had a point estimate of 0.4769, below the nominal chance level of 0.5. Identity retained better LPIPS and FRD, and LPIPS was significantly worse in the paired analysis. On this evaluation partition, the proposed Tumor-Aware Residual nnU-Net achieved better pixel-level accuracy, tumor-region structure, and segmentation oriented utility relative to identity, while perceptual similarity, radiomics distribution alignment, and tumor-specific discrimination remained areas for improvement.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MAMA_SYNTH_012.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=S09lsDDfjD
BibTex
@InProceedings{HidKei_TumorAware_MICCAISAT2026,
author = { Hidaka, Keisuke},
title = { { Tumor-Aware Residual nnU-Net for Virtual Peak-Enhancement Breast DCE-MRI Synthesis from a Single Pre-Contrast Slice } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17256},
month = {pending},
page = {pending}
}