Disclosure: The provided dataset is fully simulated and intended for research purposes only. While we have made every effort to ensure its accuracy, the dataset may contain artifacts, imperfections, or bugs. The authors do not assume any responsibility for errors or misinterpretations arising from its use. Users are encouraged to validate results independently and refer to the associated publication for methodological details.

The provided dataset and network weights can be used for Magnetic Resonance-based Electrical Properties Reconstruction. 

A extensive description about the dataset and network can be found in: https://onlinelibrary.wiley.com/doi/full/10.1002/mrm.30338

The dataset and network are fully compatible with the network available in: https://github.com/GiannakopoulosIlias/vision-transformer-network-for-mr-electrical-properties-tomography 

DATASET

Each h5 file in train, val, and test includes four 3D tensors: 

• • er: Relative Permittivity Distribution
  • se: Electric Conductivity Distribution
  • mag_b1p: Magnitude of the Magnetic Transmit Field
  • tpa_b1p: Transceive Phase

Each h5 file in canny_train, canny_val, and canny_test includes one 3D tensor

• • edges: 3D Binary Edge Mask

The dataset is splitted to

• 8065 (train) 3D volumes
• 1463 (val) 3D volumes
• 632 (test) 3D volumes 

er, se:

We generated 10160 tissue-mimicking phantoms, discretized with a voxel resolution of 5 mm³. The phantoms had either an ellipsoidal (8160) or a cylindrical geometry (2000). The ellipsoids had random lengths for the principal semi-axes, which were constrained to be at least 7, 7, and 8.5 cm and at most 9.5, 12, and 11.5 cm in the x, y, and z directions, respectively. Inside every ellipsoid, we randomly placed either 0, 1, or 23 smaller ellipsoids, varying in both size and position. To introduce inhomogeneity, each ellipsoid was randomly assigned tissue-mimicking EP, ranging from 11 to 120 for the relative permittivity and 0.07 to 2.5 S/m for the electric conductivity. The cylindrical phantoms had random length (between 17 and 23 cm) and radius (between 7 and 9.5 cm). They were either homogeneous or inhomogeneous with an additional cylindrical compartment, using random EP values in the same range as for the ellipsoids. The inner cylindrical compartment had the same length as the phantom, whereas its position and radius varied. All models were enclosed in the same cuboid domain of dimensions 19 x 23.5 x 23 cm³ that corresponded to 38 × 47 × 46 voxels. 

mag_b1p, tpa_b1p:

We modeled a 3 tesla (T) high-pass birdcage coil of radius of 12.3 cm, 8 legs, a length of 22 cm, and a copper width of 1 cm. It was discretized using 2990 triangular elements and modeled as perfect electric conductor. The capacitor values were set to 11.634 pF. We simulated the B1(+) in quadrature mode and the B1(−) (receive sensitivity) in anti-quadrature mode for all phantoms and using the hybrid-Volume-Surface Integral Equation and utilized first-order, using RWG to approximate the coil current and piecewise polynomials to approximate the polarization current. The complex B1(+) and B1(−) were corrupted with independent and identically distributed Gaussian noise. The peak SNR was chosen randomly and ranged between 50 and 200. We computed the transceive phase as the sum of the phases of B1(+) and B1(−), divided by 2.

edges:

For each phantom, we computed edge masks outlining the boundaries between different compartments by applying a Canny edge detection filter on every axial plane of the ground-truth 3D conductivity map. 

NETWORK WEIGHTS AND LOGS

TransUNet_FiLM_3

We provide the weights of the network available in https://github.com/GiannakopoulosIlias/vision-transformer-network-for-mr-electrical-properties-tomography  when trained with the provided dataset. In particular, the available weights correspond to 3 sequentially connected TransUNets and normalization using FiLM (please refer to the corresponding publication and the GitHub repository for additional details). We are also including the corresponding logs that show the network's performance during training.