Terabyte-scale supervised 3D training and benchmarking dataset of the mouse kidney

C57BL/6J mice were purchased from Janvier Labs (Le Genest-Saint-Isle, France) and kept in individually ventilated cages with ad libitum access to water and standard diet (Kliba Nafag 3436, Kaiseraugst, Switzerland) in 12 h light/dark cycles. Dataset 1 derives from the left kidney of a male mouse, 15 weeks of age with a body weight of 28.0 g. Dataset 2 is the right kidney of the same mouse. Dataset 3 derives from the right kidney of a female mouse, 15 weeks of age with a body weight of 22.5 g. All animal experiments were approved by the cantonal veterinary office of Zurich, Switzerland, in accordance with the Swiss federal animal welfare regulations (license numbers ZH177/13 and ZH233/15). Mice were anaesthetized with ketamine/xylazine. A blunted 21G butterfly needle was inserted retrogradely into the abdominal aorta and fixed with a ligation. The abdominal aorta and superior mesenteric artery above the renal arteries were ligated, the vena cava opened as an outlet and the kidneys were flushed with 10 ml, 37 °C phosphate-buffered saline (PBS) to remove the blood, then fixed with 50 ml 37 °C 4 % paraformaldehyde in PBS (PFA) solution at 150 mmHg hydrostatic pressure. 2.4 g of 1,3-diiodobenzene (Sigma-Aldrich, Schnelldorf, Germany) were dissolved in 7.5 g of 2-butanone (Sigma-Aldrich) and mixed with 7.5 g PU4ii resin (vasQtec, Zurich, Switzerland) and 1.3 g PU4ii hardener. The mixture was filtered through a paper filter and degassed extensively in a vacuum chamber to minimize bubble formation during polymerization, and perfused at a constant pressure of no more than 200 mmHg until the resin mixture solidified. Kidneys were excised and stored in 15 ml 4 % PFA. For scanning, they were embedded in 2 % agar in PBS in 0.5 ml polypropylene centrifugation tubes. Kidneys were quality-checked with a nanotom® m (phoenix|x-ray, GE Sensing & Inspection Technologies GmbH, Wunstorf, Germany). Samples showing insufficient perfusion or bleeding of resin into the renal capsule or sinuses were excluded. Kidneys were scanned at the ID19 tomography beamline of the European Synchrotron Radiation Facility (ESRF, Grenoble, France) using pink beam with a mean photon energy of 19 keV. Radiographs were recorded at a sample-detector distance of 28 cm with a 100 µm Ce:LuAG scintillator, 4× magnification lens and a pco.edge 5.5 camera with a 2560 × 2160 pixel array and 6.5 µm pixel size, resulting in an effective pixel size of 1.625 µm. Radiographs were acquired with a half-acquisition scheme in order to extend the field of view to 8 mm. Six height steps were recorded for each kidney, with half of the vertical field of view overlapping between each height step, resulting in fully redundant acquisition of the inner height steps. 5125 radiographs were recorded for each height step with 0.1 s exposure time, resulting in a scan time of 1 h for a whole kidney. 100 flat-field images were taken before and after each height step for flat-field correction. Images were reconstructed using the beamline’s in-house PyHST2 software, using a Paganin-filter with a low δ/β ratio of 50 to limit loss in resolution and appearance of gradients close to large vessels. Registration for stitching two half-acquisition radiographs to the full field of view was performed manually with 1 pixel accuracy. Data size for the reconstructed datasets was 1158 GB per kidney. Outliers in intensity in the recorded flat fields were segmented by noise reduction with 2D continuous curvelets, followed by thresholding to calculate radius and coordinates of the ring artefacts. The redundant acquisition of the central four height steps allowed us to replace corrupted data with a weighted average during stitching. The signals of the individual slices were zeroed in the presence of the rings, summed up and normalized by counting the number of uncorrupted signals. In the outer slices, where no redundant data was available, and in locations where rings coincided in both height steps, we employed a discrete cosine transform-based inpainting technique with a simple iterative approach, where we picked smoothing kernels progressively smaller in size and reconstructed the signal in the target areas by smoothing the signal everywhere at each iteration. The smoothed signal in the target areas was then combined with the original signal elsewhere to form a new image. In the next iteration, in turn, the new image was then smoothed to rewrite the signal at the target regions. The final inpainted signal exhibits multiple scales since different kernel widths are considered at different iterations. The alignment for stitching the six stacks was determined by carrying out manual 3D registration and double checking against pairwise stack-stack phase-correlation analysis. The stitching process reduced the dataset dimensions per kidney to 4608 × 4608 × 7168 pixels, totaling 567 GB. We performed image enhancement based on 3D discretized continuous curvelets, in a similar fashion as Starck et al., but with second generation curvelets (i.e., no Radon transform) in 3D. The enhancement was carried out globally by leveraging the Fast Fourier Transform with MPI-FFTW, considering about 100 curvelets. The “wedges” (curvelets in the spectrum) have a conical shape and cover the unit sphere in an approximately uniform fashion. For a given curvelet, a per-pixel coefficient is obtained by computing an inverse Fourier transform of its wedge and the image spectrum. We then truncated these coefficients in the image domain against a hard threshold, and forward-transformed the curvelet again into the Fourier space, modulated the curvelets with the truncated coefficients and superposed them. As a result, the pixel intensities were compressed to a substantially smaller range of values, thus helping to avoid over- and under-segmentation of large and small vessels, respectively. A threshold-based segmentation followed the image enhancement. The enhancement parameters and threshold were manually chosen by examining six randomly chosen regions of interest. Spurious islands were removed by 26-connected component analysis, and cavities were removed by 6-connected component analysis. The bulk of the processing workload, required to transform data into an actionable training set, was carried out at the Zeus cluster of the Pawsey supercomputing centre. Zeus consisted of hundreds of computing nodes featuring Intel Xeon Phi (Knights Landing) many-core CPUs, together with 96 GB of ``special’’ high-bandwidth memory (HBM/MCDRAM), as well as 128 GB of conventional DDR4 RAM. The final training and assessments were carried out at the Euler VI cluster of ETH Zurich, with two-socket nodes featuring AMD EPYC 7742 (Rome) CPUs and 512 GB of DDR4 RAM. A machine learning-based approach relying on invariant scattering convolution networks was employed to segment the glomeruli and remove perirenal fat from the blood vessel segment. For the glomerular training data, three selected regions of interest of 512 × 256 × 256 voxels in size were selected from the cortical region of one kidney (dataset 2) and segmented by a single annotator by fully manual contouring in all slices. For the fat, manual work was reduced by providing an initial semiautomatic segmentation, which the manual annotation then corrected. The training data were supplemented by additional regions of interest that contained no glomeruli or fat at all, and thus did not require manual annotation. The manual annotations were then used to train a hybrid algorithm that relied on a 3D scattering transform convolutional network topped with a dense neural network. The scattering transform relied upon ad-hoc designed 3D kernels (Morlet’s wavelet with different sizes and orientations) that uniformly covered all directions at different scales. In the scattering convolutional network, filter nonlinearities were obtained by taking the magnitude of the filter responses and convolving them again with the kernels in a cascading fashion. These nonlinearities are designed to be robust against small Lipschitz-continuous deformations of the image. In contrast to our curvelet-based image enhancement approach, we decomposed the image into cubic tiles, then applied a windowed (thus local) Fourier transform on the tiles by considering regions about twice their size around them. While it would have been possible to use a convolutional network based upon a global scattering transform, this would have produced a very large number of features that would have had to be consumed at once, leading to an intermediate footprint in the petabyte-scale, exceeding the available memory of the cluster. The scattering transform convolutional network produced a stack of a few hundred scalar feature maps per pixel. If considered as a “fiber bundle”, the feature map stack is equivariant under the symmetry group of rotations (i.e., the stack is a regular representation of the 3D rotation group SO(3)). This property can be exploited by further processing the feature maps with a dense neural network with increased parameter sharing across the hidden layers, making the output layer-invariant to rotations.

Kuo W, Rossinelli D, Schulz G, Wenger RH, Hieber S, Müller B, Kurtcuoglu V

Browse Data in IDR

idr0147-kuo-kidney3d/experimentA ()

BioFile Finder

BioFile Finder is a tool for filtering, sorting and grouping tabular data. Each table below is loaded from IDR as a Parquet file, with each row representing an image. Images can be selected in BioFile Finder and the "Download" link will open them in the IDR viewer.

NB: large tables may take some time to load in BioFile Finder.

experimentA ( , parquet: -)

Download

To download original image files in your browser, you can access original data.

Download as JSON.

For more download options, including bulk download using Globus or FTP clients, see the IDR Data download page.