Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-2369
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-04T18:01:57.331020Z
status: published
lens: p_space
sector: health
headline: Influence of dominant sequence artifacts on generative AI for virtual contrast-enhancement in breast MRI
dek: Virtual contrast-enhanced (vCE) breast MRI is a deep learning approach for generating contrast-like images from non-contrast acquisitions. As diffusion-weighted imaging (DWI) has emerged as a major contributor to enhancement synthesis, understanding the influence of DWI-related image artifacts is important for the robustness and potential clinical implementation of AI-based contrast-free imaging approaches. We evaluated the robustness of a pix2pix-based vCE synthesis model to artifact-contaminated DWI during tra…
gain_title: (none)
problem_title: In 61 test lesions, the lesion-level error of AI-generated virtual contrast breast MRI rose with the share of slices showing severe signal-loss artifacts (ρ = 0.478 in 12 lesions), so poor-quality input scans can distort lesion depiction.
trace_subject: (none)
gain_reading: (none)
gain_evidence: (none)
problem_reading: In 61 test lesions, the lesion-level error of AI-generated virtual contrast breast MRI rose with the share of slices showing severe signal-loss artifacts (ρ = 0.478 in 12 lesions), so poor-quality input scans can distort lesion depiction.
problem_evidence: severe signal-loss artifacts during inference impaired local performance
quick_read: Researchers at University Hospital Erlangen tested how image artifacts in diffusion-weighted MRI affect a pix2pix AI model that synthesizes contrast-enhanced breast images without injecting gadolinium. They used 1,974 breast MRI exams from 2017 to 2022, split into 1,521 training, 237 validation and 216 test exams.

Leaving artifact-affected slices out of training barely mattered. Whole-image structural similarity rose from 0.843 to 0.849 (p < 0.001, effect size 0.20), while lesion-based metrics and reader ratings of lesion enhancement and image quality were comparable between the two models.

Artifacts at inference time mattered more locally. In the 12 lesions with severe signal-loss slices, lesion error rose with the share of affected slices (Spearman ρ = 0.478, p = 0.025). The authors say this underlines the need for image quality control before clinical deployment, though the analyses were exploratory.
limitation: Correlation analyses were exploratory with no multiplicity correction, the key signal-loss finding rested on only 12 lesions, and reader agreement was low.
tag: Evidence-backed problem
key_points: The cohort had 1,974 breast MRI exams from two 3T scanners at University Hospital Erlangen, with a mean patient age of 49.4 years. | An artifact-detection network found 60.5% of high b-value DWI slices free of high-intensity artifacts and 59.6% free of low-intensity artifacts. Severe artifacts made up 11% of slices on the combined score. | Excluding severe-artifact slices (11% of training slices) raised whole-image SSIM by 0.0065 (95% CI 0.0057–0.0073), but whole-image MSE did not differ significantly (p = 0.087). | Lesion-based metrics for 61 segmented lesions showed no significant differences between the two models, for example KL divergence of 0.293 versus 0.277 (p = 0.117). | Three readers with 0.5, 5 and over 10 years of experience showed no consistent preference for either model. Agreement was low to moderate, with Cohen's κ from −0.035 to 0.380. | Scanner type mattered. The SSIM gain from artifact-aware training was 0.0115 on the Skyra Fit versus 0.0028 on the Vida, and the Vida showed more moderate-to-severe artifacts. | Only low-intensity artifacts, not high-intensity ones, were linked to higher lesion error in the restricted analysis.
rundown: The study responds to interest in supplemental breast MRI for women with dense breasts, where gadolinium contrast adds cost, supervision needs and rare side effects such as gadolinium deposition. Virtual contrast enhancement aims to produce contrast-like images from non-contrast sequences, and diffusion-weighted imaging has been identified as a dominant input.

Two versions of the same pix2pix GAN were trained for 65 epochs on T2-weighted, T1-weighted and three DWI inputs. One used all slices (vCE-all) and one dropped slices with severe artifacts (vCE-clean). Both were evaluated on the same 216 test exams, including 61 with mass lesions (BI-RADS 4–6).

In the reader study, 100 slices with severe high-intensity artifacts and 100 with severe low-intensity artifacts from 62 patients were rated. Readers split between the models, and vCE-clean kept comparable image quality despite never seeing artifact-corrupted inputs during training.

Correlations between artifact score and whole-image SSIM were near zero, and whole-image MSE correlations were small. The lesion-level link appeared only for signal-loss artifacts in a very small subset, and the authors conclude that no definitive clinical impact can be drawn.
sources:
- peer_reviewed | BMC Medical Imaging | https://doi.org/10.1186/s12880-026-02996-8 | 2026-11-03
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

8eb4322b3d91e1f166a6dd6ff9042720567661dc6a33b1bb63d408d9cb4e840c

previous

0000000000000000000000000000000000000000000000000000000000000000