OCELOT Dataset

Download Link

The dataset is available on Zenodo! Before downloading the dataset, please make sure to carefully read and agree to the Terms and Conditions.

Introduction

The OCELOT dataset is a histopathology dataset designed to facilitate the development of methods that utilize cell and tissue relationships. The dataset comprises both small and large field-of-view (FoV) patches extracted from digitally scanned whole slide images (WSIs), with overlapping regions. The small and large FoV patches are accompanied by annotations of cells and tissues, respectively. The WSIs are sourced from the publicly available TCGA database and were stained using the H&E method before being scanned with an Aperio scanner.

Each sample of the OCELOT dataset is composed of six components, $$\mathcal{D} = \{\left(x_{s}, y_s^{c}, x_l, y_l^{t}, c_x, c_y\right)_{i}\}_{i=1}^{N}$$ where $x_s, x_l$ are the small and large FoV patches extracted from the WSI, $y_s^{c}, y_l^{t}$ refer to the corresponding cell and tissue annotations, respectively, and $c_x, c_y$ are the relative coordinates of the center of $x_s$ within $x_l$. The below figure shows the visualization of a sample.

Dataset Details

We collect 304 Whole Slide Images (WSIs) obtained from the publicly available TCGA database¹. Then, the dataset is divided into three subsets: training, validation, and test, following a 6:2:2 ratio. To prevent information leaking among the data subsets, we randomly split the dataset per WSI, so that different patches from the same WSI are not included in different splits. We maintain consistent cancer-type ratios in each subset.

File Structure

images/{train,val,test}/{cell,tissue}/{uuid:03d}.jpg

annotations/{train,val,test}/cell/{uuid:03d}.csv

annotations/{train,val,test}/tissue/{uuid:03d}.png

metadata.json

Note. the validation and test splits will be released after August 11, 2023.

Metadata

The metadata of each sample is stored in a JSON file (metadata.json), which contains the following information:

slide_name

patch_x_offset

patch_y_offset

mpp_x

mpp_y

organ

subset

x_start

y_start

x_end

y_end

resized_mpp_x

resized_mpp_y

We also provide a script to parse the metadata into python class. You can find the script in utils/metadata.py.

Statistics

Annotation

Dataset size per organ and data subset

Sources of error

The extraction and processing of the specimens is a complex procedure depending on multiple factors that can cause various artifacts. As a result, the annotation of cells and tissue patches can be non-trivial for specific samples. Furthermore, pathologists show high intra- and inter-rater variability when assessing the different tissues and cells; i.e., there is a considerable amount of subjectivity. Both issues potentially lead to high annotation-wise discrepancies across pathologists. Especially for the cell patches, we adopt the 2+1 annotation strategy in order to reduce annotation variability.

Data Collection

We collect 304 Whole Slide Images (WSIs) obtained from the publicly available TCGA database¹. We provide the list of case IDs required to query further patient data that is available via the NCI GDC Database². These WSIs are gathered from 6 different organs: bladder, endometrium, head-and-neck, kidney, prostate, and stomach. From each WSI, two types of patches (image regions) are extracted at a resolution of approximately ~0.2 micronper-pixel, as follows:

• Tissue patch (large FoV): Inside each WSI, 1~3 patches with size 4096x4096 pixels were selected. The annotations consist of pixel-wise tissue categories: Background (BG), Cancer Area (CA), and Unknown (UNK).
• Cell patch (small FoV): Inside each tissue patch (as defined above), a 1024x1024 pixel region is randomly selected. Cells are annotated with the location and category into two categories: Background cell (BC), and Tumor Cell (TC).

In total, we acquire a total of 667 cell and tissue patch pairs with the corresponding annotations.

Patch Configuration

Cell detection tasks benefit from fine-grained spatial information to better capture detailed cell properties (e.g. border, shape, color, and opacity). In contrast, tissue segmentation requires a larger context to enable a better understanding of the overall structural information. Therefore, we define the FoV sizes of $x_s$ (cell detection) and $x_l$ (tissue segmentation) as 1024×1024 and 4096×4096 pixels, respectively, at a resolution of 0.2 Microns-per-Pixel (MPP). Finally, the large FoV patches and tissue annotations ($x_l$, $y_l^t$) are down-sampled by a factor of 4, resulting in a size of 1024×1024 pixels.

Annotation procedure

All cell-tissue pairs of patches are annotated by board-certified pathologists. A total of 67 certified pathologists conducted the manual annotations as follows:

cell annotations

Pathologists pinpoint the center position of each cell and assign the corresponding cell category. Annotations consider two categories of cells: Tumor Cell (TC) and Background Cell (BC). Background cells include all other cells in the patch that are not TCs.

tissue annotations

Pathologists annotated pixel-wise tissue categories by marking cancer and non-cancer regions in the tissue patch. Finally, the patch is reviewed and submitted. Three categories are considered: Cancer Area (CA), Background (BG), and Unknown (Unk). The Unknown class corresponds to uncertain annotated pixels that can be excluded during training and evaluation.

startegy

Each tissue patch is annotated by a single board-certified pathologist while for the cell patches we adopt a 2+1 annotation strategy with a final validation step. We find that the cell annotations require a more rigorous annotation procedure due to the higher intra- and inter-rater variability. First, 2 independent expert pathologists annotate the same patch. Afterward, both annotations are passed to the third annotator. The task of the third annotator is to merge both annotations. Finally, a fourth annotator validates the merged annotation to finally ensure the quality of the annotation and samples.

References

• [1] Carolyn Hutter and Jean Claude Zenklusen. “The cancer genome atlas: creating lasting value beyond its data.” Cell, 173(2):283–285, 2018.
• [2]. https://portal.gdc.cancer.gov/