Hannes Schreiter^{1,2}, Vishal Sukumar^{1,2}, Lorenz Kapsner^{1}, Lukas Folle^{2}, Sabine Ohlmeyer^{1}, Frederik Bernd Laun^{1}, Evelyn Wenkel^{1}, Michael Uder^{1}, Andreas Maier^{2}, Sebastian Bickelhaupt^{1}, and Andrzej Liebert^{1}

^{1}Institute of Radiology, University Hospital Erlangen, Friedrich-Alexander University Erlangen-Nürnberg (FAU), Erlangen, Germany, ^{2}Patter Recognition Lab, Department of Computer Science, Friedrich-Alexander University Erlangen-Nürnberg (FAU), Erlangen, Germany

Magnetic resonance imaging (MRI) examinations of the breast require intravenous administration of gadolinium based contrast agents (GBCA) for comprehensive characterization of the tissue. Novel approaches reducing the need for GBCA might therefore be of value. Here a virtual dynamic contrast enhancement (vDCE) using a U-net architecture is investigated in a cohort of n=540 patients. The vDCE generates T1 subtraction images for five consecutive time points predicting the perfusion maps based on native T1-weighted, T2-weighted, and multi-b-value diffusion weighted acquisitions. A mean structural similarity index (SSIM) value over a test group of 82 patients of 0.848±0.025 was achieved.

In recent works of Kleesiek

A U-net architecture was implemented with 3 encoder-decoder-stages. The network used as input MRI dataset comprised of T1-weighted, T2-weighted, and DWI acquisitions with three different b-values (50-1500 s/mm

Loss function 1: $$L(\theta)=\frac{1}{N}\sum_{i=1}^{N}\exp(-s_{i})*|y_{i}-y_{i}'|+s_i$$

In which

Loss function 2: $$L(\theta)=\biggl(1-\frac{(2\mu_y\mu_{y'}+C_1)(2\sigma_{yy'}+C_2}{(\mu_y^2+\mu_{y'}^2+C_1)(\sigma_y^2\sigma_{y'}^2+C_2)}\biggr)+\sum_{i=1}^N |y_i - y_{i}'|$$

Is further called Combined Loss as it combines the SSIM and L1 norm and which is a slightly changed loss function from the work of Chen

For a quantitative evaluation, a mean structural similarity index measure (SSIM) and peak signal-to-noise ratio (PSNR) over the test group was calculated for each single time point. Additionally Pearson correlation coefficients between the original and generated perfusion curves were calculated for 18 patients in both normal appearing breast tissue and in lesions.

In Figure 2 we show a comparison of the perfusion curves of both DCE and vDCE using the two different loss function in a representative case. An increase of the mean square error between the original and generated perfusion curve can be noted for the Bayesian Loss method. The mean Pearson correlation coefficient in the test group was 0.948±0.055 in the normal appearing breast tissue and 0.967±0.041 in lesions for the combined loss function. For the Bayesian loss values of 0.949±0.055 and 0.959±0.048 could be achieved respectively.

In comparison to previous works by Kleesiek

Additionally, the variability of shape is much higher in breast MRI than in brain MRI. In comparison to Chen

EPO Patent application No. 21197259.1 - Apparatus and method for generating a perfusion image, and method for training an artificial neural network therefor, Friedrich-Alexander-Universität Erlangen-Nürnberg, Bickelhaupt S., Liebert A**.**, Schreiter H., Sukumar V.,

The financial support of BMBF GoBioInitial project "SMART SELECT MR" to A.L. and H.S. is gratefully acknowledged.

1. Bickelhaupt S, Laun FB, Tesdorff J, et al. Fast and Noninvasive Characterization of Suspicious Lesions Detected at Breast Cancer X-Ray Screening: Capability of Diffusion-weighted MR Imaging with MIPs. *Radiology. *2016;278(3):689-697.

2. Kleesiek J, Morshuis JN, Isensee F, et al. Can virtual contrast enhancement in brain MRI replace gadolinium?: a feasibility study. *Investigative radiology. *2019;54(10):653-660.

3. Chen C, Raymond C, Speier B, et al. Synthesizing MR Image Contrast Enhancement Using 3D High-resolution ConvNets. arXiv preprint arXiv:210401592. 2021.

** **

Figure 1: Original DCE T1 subtraction series and vDCE results
for the two different loss functions (columns 2 and 3). The Combined loss
function shows better results in terms of both SSIM and PSNR. A small blurring
of the image can be noted for the Baysian Loss version of the vDCE

Table 1: Mean SSIM and mean PSNR over the whole test group for the each different time point generated using the vDCE with both loss functions. An overall decrease of the PSNR and an increase of the SSIM value can be noted in the later time points. The Combined loss function shows slightly higher SSIM and PSNR values.

Figure 2: Marked Regions of Interest (1^{st} column) together with perfusion curves for DCE and vDCE with different loss functions in normal appearing breast tissue (1st Row, A, B) and in a lesion (2nd row, C, D). The perfusion curves were normalized with the number of voxels of the region of interest. An increased mean square root error can be observed for the use of Bayesian Loss function.

DOI: https://doi.org/10.58530/2022/1522