Using Machine Learning to Reduce the Need for Contrast Agents in Breast MRI through Synthetic Images

Abstract

Accurate and automatic segmentation of fibroglandular tissue in breast MRI screening is essential for the quantification of breast density and background parenchymal enhancement. In this retrospective study, we developed and evaluated a transformerbased neural network for breast segmentation (TraBS) in multi-institutional MRI data, and compared its performance to the well established convolutional neural network nnUNet. TraBS and nnUNet were trained and tested on 200 internal and 40 external breast MRI examinations using manual segmentations generated by experienced human readers. Segmentation performance was assessed in terms of the Dice score and the average symmetric surface distance. The Dice score for nnUNet was lower than for TraBS on the internal testset (0.909±0.069 versus 0.916±0.067, P<0.001) and on the external testset (0.824±0.144 versus 0.864±0.081, P=0.004). Moreover, the average symmetric surface distance was higher (=worse) for nnUNet than for TraBS on the internal (0.657±2.856 versus 0.548±2.195, P=0.001) and on the external testset (0.727±0.620 versus 0.584±0.413, P=0.03). Our study demonstrates that transformer-based networks improve the quality of fibroglandular tissue segmentation in breast MRI compared to convolutional-based models like nnUNet. These findings might help to enhance the accuracy of breast density and parenchymal enhancement quantification in breast MRI screening.

AI as a New Frontier in Contrast Media Research

“…Because the contrast-to-noise ratio (CNR) always depends on both the image noise and the contrast, an artificially low CNR can be created this way. 5 However, this approach fundamentally changes the image impression and is very far from real low-dose data. In future studies, these different approaches must be tested and compared with real low-dose and normal-dose datasets.…”

“…5). Because the contrast-to-noise ratio (CNR) always depends on both the image noise and the contrast, an artificially low CNR can be created this way 5 . However, this approach fundamentally changes the image impression and is very far from real low-dose data.…”

“…The CNR is reduced by adding a noise map to generate virtually contrast-reduced image data. It corresponds to a strategy recently published by Müller-Franzes et al 5 that we transferred to the same image data of Figure 4 for comparative purposes.…”

“…A rtificial intelligence (AI) techniques, particularly generative adversarial networks (GANs), have shown significant potential in processing medical image data. In recent studies, AI-based approaches have demonstrated the possibility of reducing the amount of gadolinium-based contrast agents (GBCAs) in magnetic resonance imaging (MRI) [1][2][3][4][5][6][7][8] and iodine-based contrast media (ICM) in computed tomography (CT). [9][10][11][12][13][14] For both modalities, the amount of contrast agent respectively can be technically optimized in completely different ways, for example, by means of applications for the optimization of the radiation spectrum such as CareKV 15 for CT or by means of sequence, 16 molecule relaxivity, 17,18 and acquisition time optimization for MRI.…”

Amplifying the Effects of Contrast Agents on Magnetic Resonance Images Using a Deep Learning Method Trained on Synthetic Data

“…Synthetic data can be used to enrich and extend existing data sets, thus improving generalizability and reproducibility of AI approaches, or even to replace real-world data when these are not available. 31,32 Here, we describe a DL approach based on synthetic training data to boost image contrast in CA-enhanced MR images. Specifically, we leverage a physical model to simulate different levels of MR image contrast from a GBCA, and we apply the resulting data sets to train a CNN (as originally proposed by Lee et al 33 ) to predict contrast at GBCA doses several times larger than those used in the clinical practice.…”