Application of Generative Adversarial Networks for virtual dynamic contrast enhanced MRI of the breast using a non-enhanced acquisition protocol

点击翻译按钮获取中文摘要

Breast cancer is the most common cancer in women. One in eight women will be affected by breast cancer during their life time. Breast cancer screening significantly contributed to earlier detection and decreased mortality. Population based breast cancer screening is commonly performed using X-ray mammography, which is a long-established diagnostic procedure for detecting suspicious changes in the breast. However, X-ray mammography is increasingly limited in women with increasing breast density. Studies describe about 40% of all women to have heterogeneously or extremely dense breast. Therefore, complemenatry diagnostic modalities, such as magnetic resonance imaging (MRI), are being increasingly investigated as a supplement (or even potential alternative) to X-ray mammography. Breast MRI show the highest sensitivity among all modalities for the detection of small lesions and at the same time provides imaging without the use of ionizing radiation or breast compression. However, breast MRI requires intravenous administration of gadolinium-based contrast agents (GBCA) for the visualization of tissue perfusion and its pathologic alterations, which is typically associated to suspicious lesions. Whilst of undoubted diagnostic value in MRI, GBCA administration is not without rare, but potentially relevant side effects – which needs to be considered as well for a screening environment. Additionally, environmental aspects of gadolinium such as the anthropogenic contamination of water and the mining and manufacturing process are increasingly investigated. Finally, the contrast agent administration causes a considerable financial burden and periprocedural expenditure of time related to the application process. This project, thus aims to develop a generative-adversarial network (GAN) dedicated to breast MRI which will derive perfusion tissue properties out of a comprehensive set of non-contrast enhanced acquisitions. The derived data shall be visualizable as routine dynamic subtraction series and will allow for curve analytics of suspicious lesions. A GAN network is a system of two neural networks, a generator network which creates synthetic images and a discriminator which tries to learn to distinguish between synthetic and real images. The GAN network in this project will be based on previous works on virtual dynamic contrast enhancement algorithms for breast MRI which are able to derive tissue perfusion properties using a classic U-net architecture. This previously developed U-Net architecture will be used as the generator network of the GAN system. During the project multiple different GAN configurations will be investigated with different setups focusing on advancing the GAN technology of the discriminator networks. This shall enable to improve the diagnostic value and validity of the generated perfusion data.