Deep Learning to Simulate Contrast-enhanced Breast MRI of Invasive Breast Cancer

Abstract

Background There is increasing interest in noncontrast breast MRI alternatives for tumor visualization to increase the accessibility of breast MRI. Purpose To evaluate the feasibility and accuracy of generating simulated contrast-enhanced T1-weighted breast MRI scans from precontrast MRI sequences in biopsy-proven invasive breast cancer with use of deep learning. Materials and Methods Women with invasive breast cancer and a contrast-enhanced breast MRI examination that was performed for initial evaluation of the extent of disease between January 2015 and December 2019 at a single academic institution were retrospectively identified. A three-dimensional, fully convolutional deep neural network simulated contrast-enhanced T1-weighted breast MRI scans from five precontrast sequences (T1-weighted non-fat-suppressed [FS], T1-weighted FS, T2-weighted FS, apparent diffusion coefficient, and diffusion-weighted imaging). For qualitative assessment, four breast radiologists (with 3-15 years of experience) blinded to whether the method of contrast was real or simulated assessed image quality (excellent, acceptable, good, poor, or unacceptable), presence of tumor enhancement, and maximum index mass size by using 22 pairs of real and simulated contrast-enhanced MRI scans. Quantitative comparison was performed using whole-breast similarity and error metrics and Dice coefficient analysis of enhancing tumor overlap. Results Ninety-six MRI examinations in 96 women (mean age, 52 years ± 12 [SD]) were evaluated. The readers assessed all simulated MRI scans as having the appearance of a real MRI scan with tumor enhancement. Index mass sizes on real and simulated MRI scans demonstrated good to excellent agreement (intraclass correlation coefficient, 0.73-0.86; P < .001) without significant differences (mean differences, -0.8 to 0.8 mm; P = .36-.80). Almost all simulated MRI scans (84 of 88 [95%]) were considered of diagnostic quality (ratings of excellent, acceptable, or good). Quantitative analysis demonstrated strong similarity (structural similarity index, 0.88 ± 0.05), low voxel-wise error (symmetric mean absolute percent error, 3.26%), and Dice coefficient of enhancing tumor overlap of 0.75 ± 0.25. Conclusion It is feasible to generate simulated contrast-enhanced breast MRI scans with use of deep learning. Simulated and real contrast-enhanced MRI scans demonstrated comparable tumor sizes, areas of tumor enhancement, and image quality without significant qualitative or quantitative differences. © RSNA, 2022 Online supplemental material is available for this article. See also the editorial by Slanetz in this issue. An earlier incorrect version appeared online. This article was corrected on January 17, 2023.

Figure 1:

Schematic of the Deep Learning, Fully Convolutional Neural Network Architecture The deep learning, convolutional neural network was trained on 80×80×80 voxels training patches from preprocessed images. Our network architecture consisted of a deep learning, convolutional neural network with three-dimensional convolution bottleneck residual blocks (blue blocks), strided convolution downsampling (yellow trapezoids), transpose convolution upsampling (orange trapezoids), and long-range skip connections with feature concatenation (dashed lines). A 1×1×1 convolutional layer (orange block) mapped features to final output image patches. A schematic of the 3×3×3 bottleneck residual block is included (bottom left).

Figure 2:

Study Flow Chart Ductal carcinoma in situ (DCIS), lobular carcinoma in situ (LCIS).

Figure 3:

Quantitative Similarity and Error Metrics for Real versus Simulated Contrast-enhanced MRI across the Whole Breast Overall, there was strong similarity and low voxel-wise error across the whole breast. Similarity metrics: structural similarity index (SSIM), histogram mutual information (MI), and normalized neighborhood cross correlation (CC). Four error metrics: normalized root mean square error (NRMSE), symmetric mean absolute percent error (SMAPE), median symmetric accuracy (MdSA), and log accuracy ratio (LOGAC).

Figure 4.

Real versus Simulated Contrast-enhanced T1-weighted Axial Breast MRIs of Patients with Invasive Breast Cancer Pairs of real (top) and simulated (bottom) contrast-enhanced breast MRI from 15 patients with invasive breast cancer (arrows). Intrathoracic and extramammary structures were masked in all images.

Figure 5:

Failed Enhancement of the Index lesion on the Simulated Contrast-enhanced MRI Simulated contrast-enhanced MRI (bottom) demonstrated failed index lesion enhancement (arrows) compared with the real contrast-enhanced MRI (top).

Figure 6.

Agreement of Tumor Sizes on Real and Simulated Contrast-enhanced Breast MRI Modified Bland-Altman plot of agreement of tumor sizes on real versus simulated contrast-enhanced breast MRI. The dotted red lines and black dashed lines represent bias and limits of agreement lines (two standard deviations above and below the mean), respectively.