Pooja Subramaniam: Synthesis of Time-of-Flight Magnetic Resonance Angiography images using 3D Generative Adversarial Network

BCCN Berlin / Charité-Universitätsmedizin Berlin




Data sharing in medical imaging is not always feasible owing to patient privacy. This becomes particularly relevant when using deep learning methods as they require large amounts of data. For instance, automating blood vessel segmentation of 3D volumes of Time-of-Flight Magnetic Resonance Imaging (TOF-MRA) requires quality labelled data. Anonymization would allow data sharing. However, standard anonymization methods have been shown to be reversible especially in the case of neuroimaging. Synthesizing artificial data using Generative Adversarial Network (GAN) can play a crucial role here.


In this work, 3D TOF-MRA patches along with their brain vessel segmentation labels were generated using three different GAN architectures, deep convolutional GAN (DCGAN), Wasserstein GAN with gradient penalty (WGAN-GP) and WGAN-GP with spectral normalization (WGAN-GP-SN). The qualitative assessment of the patch-label pairs was done through visual inspection. For a quantitative assessment, two methods were employed. In the first method, precision recall distribution (PRD) curves of different generated patches with reference to real data distribution were plotted. Next, a 3D Unet was trained with the generated patch-label pairs and tested on real data. To assess the segmentation performance, Dice similarity coefficient (DSC) and 95th percentile Hausdorff distance (95HD) were used.


The 3D WGAN-GP-SN performed the best compared to the other two architectures in all the three assessments. It produced visually good quality images. The PRD curve of generated images from that architecture was closest to the PRD curve of real data. It achieved DSC of 0.766 and 95HD of 60.26 voxels when compared to the 3D Unet performance of real data with 0.902 DSC and 39.97 voxels 95HD.


In this project, three different 3D GANs for generating 3D TOF-MRA patch-label pairs were trained. The generated patch-label pairs captured the vessel structure well, and the artificial data can be shared widely. To the best of our knowledge, no study has attempted to synthesize 3D medical data and its corresponding label for anonymization. The successful application of a 3D GAN architecture to synthesize 3D image-label pairs provides a reusable framework for similar 3D medical imaging data.


Additional Information

Master Thesis Defense

If you would like to join the online talk, please contact graduateprograms(at)bccn-berlin.de


Organized by

Prof. Dr. Kerstin Ritter / Lisa Velenosi

Location: Zoom (Online)

Go back