Influence of dominant sequence artifacts on generative AI for virtual contrast-enhancement in breast MRI.
Authors
Affiliations (8)
Affiliations (8)
- Institute of Radiology, Universitätsklinikum Erlangen, Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), Maximiliansplatz 3, 91054, Erlangen, Germany. [email protected].
- Department Artificial Intelligence in Biomedical Engineering, Friedrich-Alexander- Universität Erlangen-Nürnberg, Erlangen, Germany. [email protected].
- Institute of Radiology, Universitätsklinikum Erlangen, Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), Maximiliansplatz 3, 91054, Erlangen, Germany.
- Institute of Computer Science, Polish Academy of Science, Warsaw, Poland.
- Department for Computer Science 5 (Pattern Recognition), Friedrich-Alexander-University Erlangen-Nuremberg (FAU), Martensstrasse 3, 91058, Erlangen, Germany.
- Medical Informatics, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany.
- Diagnostische Radiologie, Radiologie München, München, Germany.
- Department Artificial Intelligence in Biomedical Engineering, Friedrich-Alexander- Universität Erlangen-Nürnberg, Erlangen, Germany.
Abstract
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 training and inference. This IRB-approved, retrospective study included 1,974 multiparametric breast MRI examinations (mean age: 49 years) acquired with two clinical routine 3T MRI scanners (2017-2022). High b-value (b = 1500 s/mm<sup>2</sup>) DWI images were automatically classified for high-intensity (HI) and low-intensity (LI) artifacts using a pretrained neural network. A pix2pix 2D-GAN was trained either after excluding artifact-affected slices (vCE-clean) or using all available slices (vCE-all). Performance was evaluated using image similarity metrics, lesion-based analysis, reader assessment, and correlation of artifact scores with reconstruction errors. Excluding artifact-affected slices during training resulted in a small but statistically significant increase in whole-image structural similarity index (SSIM) with small effect size (vCE-all 0.843 ± 0.034; vCE-clean 0.849 ± 0.031, p < 0.001, d<sub>av</sub> = 0.20). However, lesion-based metrics and qualitative ratings of lesion enhancement and image quality were comparable between models. During Inference, strong LI DWI artifacts were associated with higher lesion-based errors. Controlling for artifacts on diffusion weighted imaging (DWI) during vCE training only minorly influenced technical metrics but not perceived image quality/lesion enhancement. In contrast, severe signal-loss artifacts during inference impaired local performance, highlighting the importance of image quality control during potential deployment.