Dataset for: Unsupervised segmentation of 3D microvascular photoacoustic images using deep generative learning
- Abstract:
- VAN-GAN (Vessel Segmentation Generative Adversarial Network) Sweeney, P.W. et al. (2024) ''' Synthetic Paired Dataset ''' The raw data for physic-driven photoacoustic (PA) simulations and the synthetic vascular segmentation masks are to be placed in 'raw_data/PA_Simulations' and 'raw_data/Synthetic_Segmentations', respectively. PA Simulations are split across 5 zip files 'PA_Simulations_*'. Synthetic segmentations are in a single zip. Each dataset consist of 459 tiff stacks. Raw image size: 512 x 512 x 140 (X x Y x Z) with isotropic voxels with 20 um width/height/depth (for all images in datasets). ''' Preprocessing ''' Preprocessing steps are available in the VAN-GAN code. If the option is selected to saved filtered data, tiff stacks will be saved into their train/test/val subfolder in 'filtered'. Data for VAN-GAN is stored as numpy files in train/test/val folders in the main directory. A or B indicates whether the files belong to the PA simulation or segmentation mask dataset. Example outputs are provide in 'Example_Outputs'. Here, data was saved following training of the 200th epoch. The provided tiff stacks are either a predicted segmentation mask or a fake PA simulation.
- Authors:
- P Sweeney, L Hacker, T Lefebvre, E Brown, J Gröhl, S Bohndiek
- Publication date:
- 1st Apr 2025
- Full text
- DOI