Page History

Data Access

Click the Versions tab for more info about data releases.

Additional Resources for this Dataset

• The Python code for the deep-learning models, and for 3D glandular segmentations based on synthetic-CK8 datasets, are available on GitHub at https://github.com/WeisiX/ITAS3D.

Detailed Description

Our 3D imaging method is entirely slide-free and non-destructive.  We image intact tissue specimens (prostate biopsies) that have been labeled with a fluorescent analog of H&E staining and optically cleared to make them transparent to light. The imaging is performed with a custom-developed in-house open-top light-sheet (OTLS) microscope [A.K. Glaser et al., Nature Communications, 2019].  

Our OTLS microscope uses an sCMOS camera to collect images of optically cleared tissues in a slice-by-slice manner as we translate the specimens through an illumination “light sheet.”  The sCMOS cameras generate raw 16-bit TIFF files (grayscale) that are assembled into volumetric imaging “tiles” within the RAM of the acquisition computer.  These imaging tiles are like volumetric bricks of data, as shown in Fig. 1G of [A.K. Glaser et al., Nature Communications, 2019].  Each tile is then saved to the hard drive in the form of a multi-resolution HDF5 file.  As adjacent imaging tiles are scanned within the specimen, they are appended into the same HDF5 in the hard drive. The HDF5 file is similar to a 3D TIFF stack except it contains multiple downsampled versions of the dataset so that users can quickly visualize or access the 3D data at whatever resolution they need.  The HDF5 file is also “chunked” so that smaller volumetric regions of interest can be quickly accessed.  The HDF5 file has an associated XML file that contains microscope metadata (e.g. stage coordinates for each tile, acquisition parameters, etc.).

Before performing computational analysis of the imaging data, we “fuse” the 3D imaging data contained in the raw HDF5 file.  This is done with an open-source program called “BigStitcher” that finely aligns all of the volumetric imaging tiles, shears the dataset, and blends the edges of those imaging tiles such that a “seamless” 3D image is generated.  In this project, we also downsample the imaging data by a factor of 2X in all three dimensions during this fusion process.  The fused dataset is saved as an uncompressed HDF5 file and is used as an input for our downstream ITAS3D segmentation pipeline [W. Xie et al., Cancer Research, 2022]. 

The fused HDF5 datasets (2X downsampled compared to the original data) are the rawest form of the imaging data that is provided in this TCIA collection.  These are two-channel grayscale volumetric datasets in which one channel is the fluorescent nuclear stain (TO-PRO-3) and the second channel is the eosin stain.  The TCIA collection also contains synthetic CK8 immunofluorescence datasets (2X downsampled compared to the original data) that are generated through a deep-learning image-translation model [W. Xie et al., Cancer Research, 2022].  The CK8 stain labels the luminal epithelial cells of all prostate glands (benign or cancerous).  Finally, we provide the segmentation masks of the lumen, epithelial, and stromal compartments of each biopsy (4X downsampled compared to the original data).

Citations & Data Usage Policy

Other Publications Using This Data

TCIA maintains a list of publications which leverage TCIA data. If you have a manuscript you'd like to add please contact the TCIA Helpdesk.

Version 1 (Current): Updated 2023/03/07