MIRAGE: M ULTI - SCALE L ESION -I NFORMED R EPRESENTATION WITH AUXILIARY G UIDANCE FOR MRI C ONTRAST E NHANCEMENT
Andrea Borghesi Netherlands Cancer Institute
Xin Wang Netherlands Cancer Institute
Jonas Teuwen Netherlands Cancer Institute [email protected]
arXiv:2607.19137v1 [eess.IV] 21 Jul 2026
George Yiasemis Netherlands Cancer Institute
A BSTRACT Inferring contrast enhancement from one pre-contrast breast MRI slice is underdetermined: postcontrast appearance contains physiological information that is not uniquely encoded in baseline anatomy. Optimizing only paired pixel fidelity can suppress uncertain lesion enhancement, whereas adversarial or stochastic generative objectives can favor realistic post-contrast appearance without guaranteeing patient-specific lesion fidelity. We introduce MIRAGE, a residual 2D U-Net that combines global reconstruction and perceptual losses with three forms of lesion-aware supervision available only during training: an asymmetric penalty for missed tumor enhancement, multi-scale auxiliary tumor segmentation, and guidance through a frozen post-contrast tumor segmentation nnUNet. We evaluate the method on 301 cases from the multi-centre MAMA-SYNTH data using eight complementary image-, region-, radiomics-, and segmentation-based metrics. MIRAGE ranks first on six metrics and markedly improves downstream lesion localization over tuned pix2pix, conditional diffusion, and latent bridge-matching baselines. The generative alternatives retain advantages in LPIPS or contrast classification, revealing a clear fidelity-utility trade-off. Leave-one-in and leaveone-out ablations show that the losses are partly redundant for lesion localization but exert distinct effects on appearance, radiomics, and boundary accuracy. These results support task-aware synthesis while also showing that its apparent optimality is conditional on the downstream models and metrics used to define utility. Keywords Breast MRI · DCE MRI · Virtual contrast enhancement · Image synthesis · Task-aware learning · Deep supervision
1
Introduction
Dynamic contrast-enhanced (DCE) MRI is the most sensitive breast-imaging modality and supports lesion detection, disease-extent assessment, treatment planning, and response monitoring [1]. Its value derives from signal changes after administration of a gadolinium-based contrast agent (GBCA). Although contemporary GBCAs have a strong safety record, contrast administration still requires intravenous access, adds time and complexity, is restricted in selected patients, and contributes to environmental contamination [2, 3]. These considerations motivate contrast-reduced or contrast-free protocols, provided that clinically relevant enhancement information can be recovered reliably. Virtual contrast enhancement predicts a post-contrast image y from non-contrast input x. Because enhancement depends on physiological factors not fully determined by pre-contrast anatomy, multiple targets may be compatible with the same input, particularly when only a single T1-weighted slice is available. Methods must therefore balance patient-specific fidelity with realistic post-contrast appearance, reflecting the perception–distortion trade-off [4]. In medical imaging, however, plausible synthesis may still add, shift, or suppress clinically relevant features [5].
A NDREA B ORGHESI ET AL .
Figure 1: MIRAGE architecture. A four-scale U-Net with bottleneck B and skip connections El → Dl maps pre-contrast x to synthetic post ŷ = x + fθ (x), where the image head Himg predicts the subtraction enhancement fθ (x) and a residual connection adds it back to x. At inference, only ŷ is returned. During training, segmentation heads on each decoder scale produce tumour-mask logits sℓ , supervised against the ground-truth ROI mask m, and a frozen challenge nnU-Net U segments ŷ as z (after z-normalization) using the same mask for supervision. These segmentation paths are outlined with dotted lines and are discarded at inference; m never enters the synthesiser input.
Breast virtual-contrast studies have used multi-sequence inputs, simulated low-dose acquisitions, diffusion-weighted MRI, lesion-weighted adversarial objectives, and downstream segmentation guidance [6–10]. Conditional diffusion has also shown that subtraction targets and tumor-aware losses improve lesion fidelity, while mask conditioning further helps but requires lesion localization at inference [11]. These results suggest that global pixel or perceptual criteria alone are insufficient. Existing synthesis families make different compromises: pix2pix uses adversarial learning to encourage sharp target-domain appearance [12]; denoising diffusion models represent flexible conditional distributions through iterative sampling [13]; and latent bridge matching (LBM) enables one-step translation in latent space [14]. Direct paired regression is more tightly anchored to the observed anatomy but may average uncertain detail. We investigate this tension under the MAMA-SYNTH protocol [15], which maps a single fat-suppressed pre-contrast slice to its peak-enhancement counterpart and evaluates global fidelity, tumor-region similarity, radiomic classification, and segmentation by a fixed post-contrast nnU-Net. We propose MIRAGE (Multi-scale Lesion-Informed Representation with Auxiliary Guidance for Contrast Enhancement), a residual U-Net trained with image-, lesion-, and task-aware supervision. tumor masks guide training but are not supplied at inference. Our contributions are as follows: (i) A lesion-informed objective combining asymmetric enhancement preservation, multi-scale segmentation, and frozen downstream-model guidance; (ii) Complementary leave-one-in and leave-one-out analyses that distinguish isolated benefit from redundancy in the full objective; and (iii) A comparison with adversarial, diffusion, and latent-bridge alternatives, exposing the trade-off between appearance realism and lesion utility.
2
Materials and Methods
2.1
Data and task
We use the MAMA-SYNTH release derived from MAMA-MIA [15, 16]. It contains 1,506 pretreatment breast DCEMRI examinations from four public cohorts (Duke [17], ISPY1 [18], ISPY2 [19], NACT [20]), acquired at 1.5 or 3.0 T across multiple vendors. For each examination, preprocessing selects the post-contrast phase with the highest mean signal inside the annotated tumor and the 2D slice with the largest tumor area. Each case comprises a fat-suppressed pre-contrast image x, its peak-enhancement target y, and a binary tumor mask m. A global mean and standard deviation estimated from training pre-contrast images are applied to both x and y, preserving their relative intensity change. We use 1,205 cases for training and a fixed 301-case internal validation cohort. Masks are never provided to the synthesizer at inference. 2
MIRAGE: M ULTI - SCALE L ESION -I NFORMED R EPRESENTATION WITH AUXILIARY G UIDANCE FOR MRI C ONTRAST E NHANCEMENT
2.2
Architecture and objective
MIRAGE is a four-scale residual 2D U-Net [21] with encoder/decoder widths [64, 128, 256, 512], a 1024-channel bottleneck B, InstanceNorm, and LeakyReLU. Rather than predicting y directly, an 1×1 image head Himg on the finest decoder stage D0 reads out the predicted subtraction fθ (x) and ŷ = x + fθ (x),
(1)
with skip connections from each encoder El to Dl , l=0, . . . , 3 (l=0 full res.). For training-only supervision, four 1×1 segmentation heads Hlseg attach to decoder stage Dl , l=0, . . . , 3, producing tumor-mask logits at each scale sl . A frozen downstream segmenter (see below) is applied likewise only during training. At inference the network receives only x and returns ŷ; Hlseg and the segmenter path are discarded. The full architecture is shown in Figure 1. The full objective is L = L1 + 0.3 LLPIPS + 0.05 Lunder + 0.1 Lseg + 0.2 Ldseg , (2) grouped as fidelity, ROI realism, and segmentation. The coefficients were chosen through coarse development sweeps and define one operating point, not a universal optimum. With valid-pixel mask v, L1 =
∥v ⊙ (ŷ − y)∥1 . ∥v∥1
(3)
LLPIPS is the VGG-based perceptual loss [22] on the reconstructed post-contrast image. To penalise insufficient enhancement in the lesion, define e = relu(y − x) and ê = relu(ŷ − x). The asymmetric term Lunder =
∥m ⊙ v ⊙ relu(e − ê)∥1 ∥m ⊙ v∥1
(4)
penalises only ROI pixels where ê < e; over-enhancement inside the tumor is not penalised by this term and is handled by the global reconstruction losses. For internal deep supervision, decoder head l predicts logits sl against level-l targets ml and vl (l = 0 full resolution; l > 0 obtained by adaptive average pooling of m and v followed by binarization). Lseg is the uniform sum over four scales of masked binary cross-entropy plus soft Dice, Lseg =
3 h X
i BCE(sl , ml ; vl ) + Dice(sl , ml ; vl ) .
(5)
l=0
For external task guidance, let U be the released post-contrast nnU-Net [23] from MAMA-SYNTH [15], and let N (·; v) denote per-image z-scoring over valid pixels v. With tumor logits z = U (N (ŷ, v)), Ldseg = 0.4 LBCE (z, m; v) + 0.6 LDice (z, m; v).
(6)
U remains frozen, but gradients flow through ŷ. Training applies a single full-resolution forward pass on the padded canvas, without the evaluator’s sliding-window inference, test-time mirroring, or full nnU-Net preprocessing pipeline; Ldseg is therefore a differentiable surrogate of, not an exact reproduction of, the deployed segmentation score. 2.3
Optimization, baselines, and evaluation
MIRAGE is trained with Adam, initial learning rate 5×10−4 , warm-up cosine decay, batch size 4, and 59,000 iterations on a single NVIDIA H100 GPU. We compare with tuned paired-translation implementations of pix2pix [12], a conditional DDPM [13] predicting the subtraction image, and one-step LBM [14]. All models are evaluated on the validation cohort using the eight metrics defined by the MAMA-SYNTH protocol [15]. The metrics cover four complementary aspects of synthesis quality. Global image fidelity is measured with MSE and LPIPS [22], while tumor-region realism is assessed with tumor-masked SSIM [24] and Fréchet radiomic distance (FRD), which compares radiomic-feature distributions extracted from synthetic and reference tumor regions [25]. Two fixed radiomics classifiers provide task-oriented measures. AUROCC evaluates whether the synthetic images exhibit a post-contrast enhancement phenotype by measuring their separability from pre-contrast images. AUROCR 3
A NDREA B ORGHESI ET AL .
distinguishes the annotated tumor ROI from its mirrored contralateral counterpart and hence probes whether tumorassociated enhancement is present at the correct region. Both classifiers use the reference tumor mask and consequently assess enhancement characteristics rather than automatic tumor localization. Downstream utility is evaluated with a released by MAMA-SYNTH single-fold 2D nnU-Net. Predictions on the synthetic images are compared with the reference tumor masks using Dice and the 95th-percentile Hausdorff distance (HD95). This evaluation uses the complete released nnU-Net preprocessing and inference pipeline, in contrast to the simplified differentiable forward pass used to compute Ldseg during training. For experiments repeated across random seeds, results are reported as mean ± std. Differences between configurations are assessed using the two-sample standard error SE∆ =
q
σ12 /n1 + σ22 /n2 ,
(7)
rather than the variability of either configuration alone. A difference is marked when its magnitude is at least 2, SE∆ .
3
Results
Figure 2 shows representative results from different cohorts and acquisition orientations. MIRAGE preserves the overall breast anatomy and concentrates the predicted enhancement within the annotated tumor region, although fine differences in intratumor texture and surrounding parenchymal enhancement remain visible relative to the reference. 3.1
Loss ablations
The contribution of each auxiliary objective is examined from two complementary directions. The leave-one-in experiments in Table 1 add each term individually to the masked-L1 baseline, whereas the leave-one-out experiments in Table 2 remove one term at a time from the complete objective. Table 1: Leave-one-in ablation over three seeds. Each loss is added to the masked-L1 baseline. ∗ : mean difference > 2 combined standard errors relative to L1 . Image fidelity
ROI realism
Classification
Segmentation
Recipe
MSE ↓
LPIPS ↓
FRD ↓
SSIMtum ↑
AUROCC ↑
AUROCR ↑
Dice ↑
HD95 ↓
L1 only + LLPIPS + Lseg + Ldseg + Lunder
0.597±0.009 0.603±0.004 0.568∗±0.009 0.596±0.011 0.599±0.004
0.173±0.003 0.091∗±0.002 0.172±0.005 0.164∗±0.005 0.170±0.002
10.47±0.36 7.23∗±0.12 10.50±1.47 7.77∗±0.64 9.81±0.83
0.482±0.017 0.498±0.013 0.543∗±0.018 0.511∗±0.008 0.554∗±0.004
0.721±0.009 0.726±0.006 0.660∗±0.006 0.795∗±0.002 0.698∗±0.009
0.683±0.005 0.694∗±0.007 0.712∗±0.003 0.681±0.007 0.713∗±0.016
0.495±0.038 0.562∗±0.007 0.635∗±0.010 0.634∗±0.016 0.632∗±0.025
121.7±17.6 98.7∗±3.8 56.5∗±7.6 58.1∗±10.8 54.9∗±11.3
The masked-L1 model performs poorly on downstream segmentation. Adding any of the four auxiliary terms significantly improves both Dice and HD95. The three mask-aware terms reach Dice scores within 0.003 of one another, showing that each can independently recover a large part of the localization performance. The terms differ more clearly outside the segmentation metrics. Adding either LLPIPS or Lunder improves image fidelity, tumor-masked SSIM, and radiomic agreement. Adding Lseg or Ldseg produces the largest gains in tumor localization and also improves tumor-region realism. The complete objective achieves better segmentation performance than any leave-one-in configuration. In the leave-oneout analysis, removing Lseg is the only ablation that significantly reduces Dice. Removing Ldseg significantly worsens both FRD and HD95, while Dice remains essentially unchanged. Removing LLPIPS worsens LPIPS and also increases FRD. The effect of Lunder differs between the two analyses. It provides clear improvements when added alone to the masked-L1 model, but removing it does not materially alter FRD, Dice, or HD95. Its remaining effect is a small improvement in SSIMtum accompanied by a small increase in MSE. All four leave-one-out configurations produce a higher mean AUROCC than the complete objective. None of these differences reaches the 2σ criterion, however, and the complete model has the largest variability for this metric, with a standard deviation of ±0.031. 4
MIRAGE: M ULTI - SCALE L ESION -I NFORMED R EPRESENTATION WITH AUXILIARY G UIDANCE FOR MRI C ONTRAST E NHANCEMENT
Figure 2: Qualitative synthesis on three held-out cases from different cohorts, each of single-breast view. The contour indicates the expert annotated tumor.
3.2
Comparison across synthesis paradigms
Table 3 compares MIRAGE with pix2pix, conditional DDPM, and LBM on the same validation cohort and evaluation pipeline. The methods occupy different operating points across the metric groups. LBM achieves the best LPIPS, while DDPM obtains the highest AUROCC . MIRAGE achieves the best downstream segmentation performance and the lowest FRD. All three alternative synthesis models perform worse than MIRAGE on Dice and HD95. They also produce higher FRD values, despite obtaining stronger results on selected appearance-oriented metrics. Thus, the models producing the most perceptually favorable or strongly contrast-like outputs do not produce the most accurate tumor localization or the closest tumor-region radiomic distribution. 5
A NDREA B ORGHESI ET AL .
Table 2: Leave-one-out ablation over three seeds. Each loss is removed from the full objective. ∗ : mean difference > 2 combined standard errors; † : complete seed separation without meeting that criterion. Image fidelity
ROI realism
Classification
Segmentation
Recipe
MSE ↓
LPIPS ↓
FRD ↓
SSIMtum ↑
AUROCC ↑
AUROCR ↑
Dice ↑
HD95 ↓
Full − Ldseg − Lseg − LLPIPS − Lunder
0.607±0.008 0.590∗±0.010 0.601±0.013 0.608±0.009 0.581∗±0.004
0.095±0.001 0.093∗±0.001 0.095±0.001 0.167∗±0.001 0.094±0.002
6.56±0.15 6.98∗±0.22 6.78±0.17 8.74†±2.13 6.79±0.51
0.540±0.007 0.553∗±0.004 0.543±0.005 0.521∗±0.010 0.530∗±0.005
0.711±0.031 0.712±0.011 0.758±0.035 0.732±0.008 0.718±0.009
0.718±0.013 0.712±0.013 0.708±0.003 0.735†±0.009 0.717±0.011
0.679±0.009 0.667±0.010 0.663∗±0.004 0.671†±0.002 0.674±0.001
34.8±2.3 45.5∗±4.1 45.2∗±3.1 36.1±3.0 34.4±2.5
Table 3: Cross-paradigm comparison on the validation cohort. Average rank is the mean of the eight per-metric ranks and is descriptive rather than the benchmark’s group-wise ranking. Image fidelity
ROI realism
Classification
Segmentation
Model
MSE ↓
LPIPS ↓
FRD ↓
SSIMtum ↑
AUROCC ↑
AUROCR ↑
Dice ↑
HD95 ↓
Avg rank ↓
MIRAGE (ours) DDPM pix2pix LBM
0.607 0.795 0.628 0.612
0.095 0.122 0.169 0.079
6.56 9.07 11.75 8.12
0.540 0.389 0.497 0.529
0.711 0.816 0.755 0.792
0.718 0.632 0.677 0.657
0.679 0.495 0.509 0.493
34.8 122.4 106.2 126.6
1.50 3.13 2.88 2.50
4
Discussion and Conclusion
The ablations show that the auxiliary objectives provide overlapping but distinct constraints. All improve the masked-L1 baseline individually, while the complete objective achieves better downstream segmentation than any single-term addition. The final performance therefore cannot be attributed to one dominant loss. The segmentation objectives are particularly complementary. Removing Lseg is the only ablation that significantly reduces Dice, indicating that direct multiscale supervision remains important for overlap-based localization. Removing Ldseg , by contrast, primarily worsens HD95 and FRD while leaving Dice largely unchanged, suggesting that guidance from the frozen segmenter additionally constrains lesion boundaries and tumor-region appearance. Removing LLPIPS likewise worsens both LPIPS and FRD. This association does not imply that LPIPS directly optimizes radiomic fidelity, but shows that perceptual and radiomic agreement are coupled in the present setting. The clearest redundancy concerns Lunder . Although effective when added alone, it has little effect on FRD, Dice, or HD95 once the other losses are present. Its remaining improvement in SSIMtum is accompanied by a small increase in MSE. The other lesion-aware objectives may therefore already provide much of the pressure needed to preserve tumor enhancement, making Lunder the first candidate for removal when favoring a simpler objective. The classification and cross-paradigm results reveal a broader trade-off. Removing any auxiliary term slightly increases AUROCC , although none of these differences reaches the predefined 2σ threshold. Similarly, LBM achieves the best LPIPS and DDPM the highest AUROCC , yet pix2pix, DDPM, and LBM all perform worse than MIRAGE on segmentation and FRD. Images that appear sharper or more characteristically post-contrast are therefore not necessarily better aligned with the patient-specific tumor location or radiomic distribution. The fixed classifier may reward a more stereotyped enhancement phenotype, whereas the lesion-aware losses emphasize spatial correspondence, although the observed AUROCC trend remains tentative. These differences are consistent with the respective objectives. Adversarial and generative distribution-learning methods encourage outputs that resemble the post-contrast domain, but realistic appearance alone does not guarantee correctly localised enhancement. The results concern the evaluated implementations and should not be interpreted as showing that adversarial, diffusion, or bridge-matching methods are intrinsically unsuitable. The selected operating point is nevertheless conditional. Loss weights were chosen through coarse development sweeps on one internal validation cohort rather than exhaustive multi-objective optimisation, and different applications may prioritise different compromises between sharpness, localization, and downstream utility. Moreover, changes to the downstream model for Ldseg computation could alter the optimal objective and weighting. 6
MIRAGE: M ULTI - SCALE L ESION -I NFORMED R EPRESENTATION WITH AUXILIARY G UIDANCE FOR MRI C ONTRAST E NHANCEMENT
Finally, none of the reported metrics establishes physiological correctness. Pixelwise metrics compare against one observed post-contrast acquisition, radiomic metrics capture selected image properties, and fixed classifiers and segmenters inherit their own biases. External cohorts, multiple acquisition protocols and downstream models, and reader studies are therefore needed to determine whether synthetic enhancement preserves diagnostically relevant findings without introducing, displacing, or suppressing lesions. In conclusion, MIRAGE synthesizes peak post-contrast breast MRI from a single pre-contrast slice using tumor-informed supervision available only during training. It provides a stronger overall balance of patient-specific fidelity, tumor-region radiomic agreement, and downstream segmentation than the evaluated generative baselines, while conceding selected appearance-oriented metrics. More broadly, the findings support task-aware medical image synthesis in which pretrained downstream models guide outputs towards preserving clinically relevant information, rather than merely plausible appearance.
7
A NDREA B ORGHESI ET AL .
References [1] Ritse M Mann, Nariya Cho, and Linda Moy. Breast mri: state of the art. Radiology, 292(3):520–536, 2019. [2] Jitka Starekova, Ali Pirasteh, and Scott B Reeder. Update on gadolinium-based contrast agent safety, from the ajr special series on contrast media. American Journal of Roentgenology, 223(3):e2330036, 2024. [3] Helena M Dekker, Gerard J Stroomberg, Aart J Van der Molen, and Mathias Prokop. Review of strategies to reduce the contamination of the water environment by gadolinium-based contrast agents. Insights into Imaging, 15(1):62, 2024. [4] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228–6237, 2018. [5] Joseph Paul Cohen, Margaux Luck, and Sina Honari. Distribution matching losses can hallucinate features in medical image translation. In International conference on medical image computing and computer-assisted intervention, pages 529–536. Springer, 2018. [6] Maggie Chung, Evan Calabrese, John Mongan, Kimberly M Ray, Jessica H Hayward, Tatiana Kelil, Ryan Sieberg, Nola Hylton, Bonnie N Joe, and Amie Y Lee. Deep learning to simulate contrast-enhanced breast mri of invasive breast cancer. Radiology, 306(3):e213199, 2022. [7] Gustav Müller-Franzes, Luisa Huck, Soroosh Tayebi Arasteh, Firas Khader, Tianyu Han, Volkmar Schulz, Ebba Dethlefsen, Jakob Nikolas Kather, Sven Nebelung, Teresa Nolte, et al. Using machine learning to reduce the need for contrast agents in breast mri through synthetic images. Radiology, 307(3):e222211, 2023. [8] Tianyu Zhang, Luyi Han, Anna D’Angelo, Xin Wang, Yuan Gao, Chunyao Lu, Jonas Teuwen, Regina Beets-Tan, Tao Tan, and Ritse Mann. Synthesis of contrast-enhanced breast mri using t1-and multi-b-value dwi-based hierarchical fusion network with attention mechanism. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 79–88. Springer, 2023. [9] Eunjin Kim, Hwan-Ho Cho, Junmo Kwon, Young-Tack Oh, Eun Sook Ko, and Hyunjin Park. Tumor-attentive segmentation-guided gan for synthesizing breast contrast-enhanced mri without contrast agents. IEEE journal of translational engineering in health and medicine, 11:32–43, 2022. [10] Richard Osuala, Smriti Joshi, Apostolia Tsirikoglou, Lidia Garrucho, Walter HL Pinaya, Daniel M Lang, Julia A Schnabel, Oliver Diaz, and Karim Lekadir. Simulating dynamic tumor contrast enhancement in breast MRI using conditional generative adversarial networks. Journal of Medical Imaging, 12(S2):S22014, 2025. [11] Sebastian Ibarra, Javier del Riego, Alessandro Catanese, Julian Cuba, Julian Cardona, Nataly Leon, Jonathan Infante, Karim Lekadir, Oliver Diaz, and Richard Osuala. Comparing conditional diffusion models for synthesizing contrast-enhanced breast mri from pre-contrast images. In Deep Breast Workshop on AI and Imaging for Diagnostic and Treatment Challenges in Breast Care, pages 226–236. Springer, 2025. [12] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 1125–1134, 2017. [13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, volume 33, pages 6840–6851, 2020. [14] Clément Chadebec, Onur Tasar, Sanjeev Sreetharan, and Benjamin Aubin. LBM: Latent bridge matching for fast image-to-image translation. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), pages 29086–29098, 2025. [15] Richard Osuala, Smriti Joshi, Jarek van Dijk, Luyi Han, Maria Laura Cosaka, Daniel Mysler, Lidia Garrucho, Karim Lekadir, Simone Balocco, and Oliver Diaz. The mama-synth challenge: Synthesizing virtual contrastenhancement in breast mri, April 2026. [16] Lidia Garrucho, Kaisar Kushibar, Claire-Anne Reidel, Smriti Joshi, Richard Osuala, Apostolia Tsirikoglou, Maciej Bobowicz, Javier Del Riego, Alessandro Catanese, Katarzyna Gwoździewicz, et al. A large-scale multicenter breast cancer dce-mri benchmark dataset with expert segmentations. Scientific data, 12(1):453, 2025. [17] Ashirbani Saha, Michael R. Harowicz, Lars J. Grimm, Jing Weng, Elizabeth H. Cain, Connie E. Kim, Sujata V. Ghate, Ruth Walsh, and Maciej A. Mazurowski. Dynamic contrast-enhanced magnetic resonance images of breast cancer patients with tumor locations, 2021. [Data set]. [18] David Newitt, Nola Hylton, and on behalf of the I-SPY 1 Network and ACRIN 6657 Trial Team. Multi-center breast DCE-MRI data and segmentations from patients in the I-SPY 1/ACRIN 6657 trials, 2016. [19] Wen Li, David C. Newitt, Jessica Gibbs, Lisa J. Wilmes, et al. I-spy 2 breast dynamic contrast enhanced mri (i-spy2 trial), 2022. 8
MIRAGE: M ULTI - SCALE L ESION -I NFORMED R EPRESENTATION WITH AUXILIARY G UIDANCE FOR MRI C ONTRAST E NHANCEMENT
[20] David Newitt and Nola Hylton. Single site breast DCE-MRI data and segmentations from patients undergoing neoadjuvant chemotherapy, 2016. [Data set]. [21] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: convolutional networks for biomedical image segmentation. In Proc. Med. Image Comput. Comput. Assist. Interv. (MICCAI), pages 234–241, 2015. [22] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 586–595, 2018. [23] Fabian Isensee, Paul F. Jaeger, Simon A. A. Kohl, Jens Petersen, and Klaus H. Maier-Hein. nnU-Net: a selfconfiguring method for deep learning-based biomedical image segmentation. Nature Methods, 18(2):203–211, 2021. [24] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600–612, 2004. [25] Nicholas Konz, Richard Osuala, Preeti Verma, Yuwen Chen, Hanxue Gu, Haoyu Dong, Yaqian Chen, Andrew Marshall, Lidia Garrucho, Kaisar Kushibar, et al. Fréchet radiomic distance (frd): A versatile metric for comparing medical imaging datasets. Medical Image Analysis, page 103943, 2026.
9