Source-linked AI summary
Anatomy-Aware Synthesis of Post-Contrast Breast MRI from Pre-Contrast Images
Zhengbo Zhou, Dooman Arefan, Lin Gu, Ufara Zuwasti Curran, Shandong Wu
TL;DR
Administered contrast in breast MRI adds cost, workflow complexity, and potential safety concerns, while synthesizing post-contrast images must preserve localized enhancement and shared anatomy. This paper adds anatomy-aware supervision to pre-contrast-to-post-contrast synthesis and reports improved whole-image and regional quantitative performance over baseline methods, with no significant Ki-67 classification differences across real/synthetic settings.
Problem
Contrast administration adds cost, workflow complexity, and potential safety concerns, while synthesis must preserve diagnostically important tumor enhancement, BPE, and shared anatomy.
Source: §1 · p. 2Method
The framework extends Pix2PixHD with breast-region consistency and lesion- and BPE-region supervision to synthesize post-contrast images from pre-contrast inputs.
Source: §1 · p. 2, §2.3 · p. 4Results
The method improved quantitative synthesis performance over conventional GAN, Pix2PixHD, and diffusion-based baselines on whole-image and targeted regional evaluations; Ki-67 classification showed no significant differences across real/synthetic configurations.
Source: §4 · pp. 12–13Takeaways & Limitations
The findings support continued investigation of synthetic contrast enhancement as a step toward reducing reliance on contrast administration in selected breast MRI applications.
Source: §4 · p. 13The proof-of-concept study used data from a single scanner and institution, limiting the scope of its generalizability evidence.
Source: §4 · p. 13
Abstract
from arXiv · showhide
We developed an anatomy-aware deep learning framework to synthesize post-contrast breast MRI from pre-contrast images, emphasizing tumor and background parenchymal enhancement (BPE) regions. This retrospective study included 649 patients with 6,251 paired pre-contrast and post-contrast images. The framework integrates breast mask consistency, lesion-region supervision, and BPE-region supervision into an image-to-image translation model. Evaluation included quantitative image quality metrics, a reader study with two breast radiologists, and downstream Ki-67 classification. The proposed method outperformed Pix2Pix, Pix2PixHD, diffusion-based synthesis, and mask-supervised baselines in whole-image and regional evaluations. Ki-67 classification showed no statistically significant performance differences across real- and synthetic-image training and testing settings, although this does not establish equivalence. These findings suggest that anatomy-aware supervision improves synthesis fidelity and support further investigation of synthetic post-contrast MRI for contrast-free imaging workflows.
1 Introduction
Contrast-enhanced breast MRI supports lesion characterization and treatment assessment, but administered gadolinium adds cost, workflow complexity, and potential safety concerns. The paper proposes anatomy-aware synthesis to address gaps in preserving enhancement-related regions and evaluating both human-perceived quality and downstream utility.
1 Introduction
- Gadolinium-based contrast administration adds cost, workflow complexity, and potential safety concerns, motivating synthesis from pre-contrast scans while preserving diagnostically relevant enhancement patterns.The proposed computational approach could reduce reliance on administered contrast.Source: §1 · p. 2
- Prior work examined lesion-focused GANs, downstream tumor segmentation, temporal diffusion, and limits of similarity metrics, but rarely combined these perspectives.The introduction presents generation strategy and evaluation protocol as important to reliable breast MRI synthesis.Source: §1 · p. 2
- A unified framework that preserves shared anatomy and both tumor and BPE enhancement while combining reader and downstream evaluation remains underexplored.Prior work suggests localized, anatomically meaningful regions carry diagnostically important information, and reader-based assessment and downstream modeling are rarely examined together.Source: §1 · p. 2
- The proposed Pix2PixHD-based framework adds breast-, lesion-, and BPE-region supervision and evaluates synthesized images through a blinded reader study and downstream Ki-67 prediction.The study involved two experienced breast radiologists and assesses functional utility as well as visual realism.Source: §1 · p. 2
2 Materials and Methods
The study uses patient-level cross-validation on a single-scanner breast MRI cohort and extends Pix2PixHD with anatomy- and region-specific supervision. Evaluation combines image-quality metrics, a reader study, and downstream Ki-67 prediction.
2.1 Study Cohorts and Datasets
- 649 patients contributed 6,251 paired pre-contrast and post-contrast images, with patient-level five-fold cross-validation preventing slices from one patient entering both training and test sets.Table 1 stratifies clinicopathologic characteristics by Ki-67 status, defining high as >20 and low as ≤20.Source: §2.1 · p. 3, Table 1 · p. 3
2.2 Breast DCE-MRI
- All women underwent prone imaging on a 1.5-T GE SIGNA scanner with a dedicated seven-channel breast coil; parameters included 512 × 512 in-plane resolution and 2–3 mm slice thickness.The field of view was 28–34 cm.Source: §2.2 · p. 3
2.3 Model Architecture
- The Pix2PixHD-based Pix2PixH-BTC model uses pre-contrast images as its only test-time input, with training supervision from breast masks, lesion regions, and BPE regions.Its full- and half-resolution discriminators target fine detail and global anatomical layout; auxiliary losses include perceptual, lesion, BPE, and shared-region consistency terms.Source: §2.3 · p. 4, Figure 1 · p. 5
2.4 Automated BPE Quantification
- Automated BPE quantification segmented breast and fibroglandular tissue, excluded tumor regions, then thresholded pre-/post-contrast differences to isolate enhancement areas.The method used a pretrained U-Net for breast segmentation and fuzzy c-means clustering to select the brightest fibroglandular-tissue cluster.Source: §2.4 · pp. 4–5
2.5 Quantitative Evaluation
- Image quality was assessed with SSIM, PSNR, and NRMSE for whole images, tumor areas, and tumor-plus-BPE regions.Representative comparisons included ground-truth post-contrast images, GAN outputs, Pix2PixHD outputs, and the proposed method.Source: §2.5 · p. 5
2.6 Reader Study
- Two breast radiology fellows reviewed 130 real and 130 synthesized middle-tumor-slice images, scoring lesion and enhancement features and perceived synthetic appearance.Feature similarity was evaluated with chi-square or Kolmogorov–Smirnov tests and summarized using Cramer's V or Cohen's d; the realism scale ranged from very real to very synthetic.Source: §2.6 · pp. 5–6, Figure 2 · p. 6
2.7 Downstream Biomarker Prediction
- Ki-67 prediction used radiomics and deep-learning workflows, with high expression defined by a threshold of 20 and five-fold cross-validation on 648 cases.Deep-learning inputs included whole images, tumor-masked regions, and tumor crops with an 8-pixel margin; experiments included real-to-real, fake-to-fake, fake-to-real, and real-to-fake settings.Source: §2.7 · pp. 6–7, Figure 3 · p. 7
3 Results
The proposed method led quantitative comparisons and appeared closer to ground truth in qualitative cases; reader findings showed both feature similarity and remaining detectable differences.
3.1 Quantitative Synthesis Performance
- Tumor-area evaluation favored the proposed method (SSIM 64.4, PSNR 13.2, NRMSE 38.8), as did tumor-plus-BPE evaluation (64.7, 13.4, 38.9).The authors state that these results suggest improved fidelity in clinically meaningful enhancement regions.Source: §3.1 · p. 8
- The proposed method led all models in whole-image SSIM (70.8), PSNR (25.1), and NRMSE (41.6), versus Pix2PixHD-BTM (70.1, 24.8, 43.5) and Pix2PixHD (69.3, 24.5, 45.3).Table 2 also covers tumor-area and tumor-plus-BPE evaluations.Source: §3.1 · pp. 7–8, Table 2 · p. 8
3.2 Qualitative Findings
- Figure 4 reports more realistic enhancement and better preservation of anatomy and lesion appearance with the proposed method; Figure 5 shows a second case with magnified lesion insets.The qualitative comparisons include GAN, Pix2PixHD, and diffusion outputs alongside pre-contrast inputs and ground-truth post-contrast images.Source: Figure 4 · p. 9, Figure 5 · p. 10
3.3 Radiologist Assessment
- Readers found high-to-very-high similarity for lesion features; NME differed significantly but had a modest effect size (P = 0.0010; Cohen’s d = -0.3967).Lesion shape, margin, and internal enhancement showed no significant differences between real and synthesized images.Source: §3.3 · p. 11
- Readers distinguished synthetic from real images with moderate performance (mean AUC 0.779; accuracy 0.735).The reader study evaluated 260 images: 130 real and 130 synthesized.Source: §3.3 · p. 11
3.4 Downstream task for Ki-67 Prediction
- Ki-67 AUCs did not differ significantly across synthetic-image configurations versus acquired-image reference for either model (all P ≥.46).EfficientNet-B2 mean AUCs were 0.602 (GT→GT), 0.600 (synthetic→synthetic), 0.598 (synthetic→acquired), and 0.595 (acquired→synthetic).Source: §3.4 · pp. 11–12, Table 4 · p. 12
4 Discussion
An anatomy-aware Pix2PixHD framework improved whole-image and lesion/enhancement-region synthesis over conventional GAN, Pix2PixHD, and diffusion baselines. The discussion reports reader-feature similarity and no significant differences across Ki-67 settings, while noting residual image differences and study-scope limitations.
4 Discussion
- The anatomy-aware framework improved quantitative synthesis over conventional GAN, Pix2PixHD, and diffusion-based baselines, including in clinically relevant regions.The study concludes that this synthesis is feasible and improves fidelity in clinically relevant regions.Source: §4 · pp. 12–13
- Lesion and BPE supervision with shared-region consistency encourages coherent enhancement within anatomy and better preserves focal lesion appearance than baseline methods.The approach addresses the challenge of spatially sparse, anatomically constrained contrast changes that standard image-level losses may underemphasize.Source: §4 · p. 12
- Radiologist-extracted lesion and enhancement features did not differ significantly between synthetic and real images, yet perceptible image differences remained.Artifacts outside the breast may reveal synthesis without directly affecting breast or lesion analysis; future work should investigate these differences.Source: §4 · pp. 12–13
- No significant differences across real/synthetic train-test configurations indicated that synthetic images preserved quantitative information relevant to Ki-67 classification.The downstream evaluation adds a clinical-utility assessment beyond technical image-quality evaluation.Source: §4 · p. 13
- The proof-of-concept study used single-institution, single-scanner data, only pre-contrast MRI inputs, and downstream evaluation limited to Ki-67 with models not optimized for that task.The authors identify multi-scanner data, additional MR sequences, further disagreement analysis, and additional downstream tasks as future work.Source: §4 · p. 13