Background & Summary

Breast cancer is the most common cancer in women globally with 11.7% of all cases and the fourth most common cause of cancer deaths in women with 6.9% of all cancer deaths1. Biomarker assessment through IHC staining, particularly of the hormone receptors ER and PGR, for oestrogen and progesterone, respectively, as well as for the receptor for human epidermal growth factor 2, HER2, has become an essential component of the routine pathology workflow where available2,3. Another biomarker that is routinely assessed through IHC staining in some countries is KI67. The International KI67 in Breast Cancer Working Group (IKWG) currently recommends KI67 scoring at least in patients classified as ER-positive and HER2-negative based on IHC scores4.

Automated IHC biomarker scoring with image analysis software can enhance its validity and reproducibility. The IKWG found that e.g. for KI67 scoring, automated scoring with QuPath5 shows outstanding reproducibility6. In recent years, image analysis in the context of computational pathology has advanced for a whole range of applications. This has at least in part been facilitated by large, publicly available WSI data sets, such as the resources provided by the TCGA research network. Publicly available data does not only provide development data to the research community, but perhaps even more importantly, it also allows for a comparable benchmarking of novel methods on the same test data.

There are several application areas where multi-modal (e.g. multiple stains) WSI image data is required. This includes development of high performing WSI registration (the spatial alignment of corresponding tissue in two or more WSIs) methods, which is an enabling technology both for research and diagnostics. It can allow clinicians to fuse information from different IHC-stains in WSI viewers. Combining information from H&E-stained tissue with corresponding IHC-stained tissue regions can e.g. be of critical importance when investigating resection borders with respect to malignancy, which can be very time consuming without tissue alignment. Some commercial IHC scoring softwares also align H&E and IHC-stained tissue to enable pathologists to contextualise automated scoring results. In research projects, WSI registration can facilitate stain-guided learning7,8,9, virtual staining10,11,12, 3D reconstruction13,14 and the transfer of annotations between different WSIs and stains. However, there is currently a lack of publicly available data sets that include WSIs from H&E-stained tissue sections with matched IHC-stained tissue from the same tumour, despite the importance of IHC for pathological diagnosis.

To promote and enable further research in this domain, we have published the ACROBAT (AutomatiC Registration Of Breast cAncer Tissue) data set15, which consists of 4,212 WSIs from 1,153 female primary breast cancer patients. For each patient, the data set15 contains one WSI of H&E stained tissue and up to four WSIs with tissue that was stained with the routine diagnostic IHC markers ER, PGR, HER2 and KI67. An example of a case from the data set15 with all four IHC antibodies available is depicted in Fig. 1.

Fig. 1
figure 1

Example of an H&E-stained tissue region with corresponding IHC-stained tissue with all four routine diagnostic stains available in this data set. (a) shows H&E, (b) ER, (c) HER2, (d) Ki67 and (e) PGR. (f) shows an example of a WSI that was excluded since it contains multiple tissue sections.

The data set15 was initially collected as part of the CHIME study ( at Karolinska Institutet (Stockholm, Sweden). The primary purpose of the CHIME study is to advance precision medicine through computational pathology, based on population representative patient cohorts. Histopathology slides came from the routine clinical diagnostic workflow, with WSIs generated using high-throughput histopathology slide scanners at Karolinska Institutet.

The primary purpose of publishing this data set15 was to enable the ACROBAT WSI registration competition (, which took place in the scope of the MICCAI (Medical Image Computing and Computer Assisted Intervention) 2022 conference.

While the primary purpose of this data set15 is the development of WSI registration methods, we believe that there could also be other use cases for the data. These may include the development of digital staining and stain transfer methods, as well as novel methods in stain-guided learning. Further applications may include the development of tissue segmentation and classification algorithms, the development of artefact detection or correction methods and unsupervised pre-training e.g. of convolutional neural networks (CNNs), which are then fine-tuned for specific tasks. We therefore hope that this data set15 can contribute towards the advance of WSI registration methods but also other research activities in the domain of computational pathology.


Data selection and splits for ACROBAT data set

The CHIME breast cancer study is based upon a retrospective cohort study design. Consecutive female breast cancer cases diagnosed between 2012 and 2018 at Södersjukhuset (Stockholm, Sweden) were included. The study has approval by the regional ethics review board (Etiksprövningsmyndigheten, Stockholm, Sweden, ref. 2017/2106-31 and amendments 2018/1462-32, 2019-02336). Due to the retrospective nature of the study, consent was not required. The archived histopathology slides were retrieved and scanned. The training and validation set are a randomly selected subset of the data generated in terms of the CHIME study. The test set WSIs were chosen as a subset of the CHIME study data that has previously been reviewed by a pathologist specialising in breast pathology in the context of another research project. Cases were excluded for the ACROBAT data set15 only if one of the WSIs contains multiple sections of the same tissue, which occurs in approximately 1% of H&E WSIs in this data set15, as depicted in Fig. 1f. These cases were excluded since it would be unclear to which of the multiple sections tissue from the corresponding IHC WSI should be aligned to. In this case, a new case was randomly selected and included after the corresponding quality control for multiple tissue sections.

The training set consists of 3,406 WSIs from 750 patients. Each patient has one H&E WSI and up to four associated IHC WSIs from the routine diagnostic IHC antibodies ER, PGR, HER2, KI67, as depicted in Fig. 1. The validation data set consists of 200 WSIs from 100 patients and the test set consists of 606 WSIs from 303 patients. Each case in the validation and test sets consists of one H&E WSI per case and one IHC WSI, which was randomly selected stratifying for IHC antibody. The test set was furthermore selected by stratifying for clinical covariates by balanced sampling from the three different WSI scanners. Table 1 indicates the distributions of scanners and IHC antibodies in the respective subsets.

Table 1 Distribution of cases and WSIs to training, validation and test set, as well as respective distributions of stains, IHC antibodies and scanner models.

Whole slide image scanning

WSIs in the CHIME study that were available at the time of data selection were generated from archived histopathology slides with three Hamamatsu WSI scanners, consisting of one NanoZoomer S360 and two NanoZoomer XRs. Slides were digitised by a trained scanning technician using an automated scanning workflow, with manual rescanning of slides where automated focusing was not successful. Slides were scanned at a resolution of approximately 0.23 µm/pixel and a JPEG compression quality level of 80.

Image processing

The 40X NDPI WSIs were first anonymized and then converted to pyramidal TIFF files with 10X and lower resolutions. Macro images and other identifying information in the WSI metadata were removed with code available from Then, file names were generated consisting of a random case ID, the stain or antibody of the WSI, as well as the name of the respective set out of training, validation or test. TIFF files were then extracted using the libvips16 command im_vips2tiff at 10X and lower magnifications, with 7 to 9 magnification levels depending on the WSI available and a downsampling factor of 2 between these levels. This reduces the storage requirements of the data set15 from 10.13 TB to 482 GB, likely without impacting the performance of image registration algorithms as registration is typically performed at fairly low resolutions with diminishing to no improvements at higher resolutions17.

Annotation workflow

Members of the ABCAP research consortium ( were enrolled as annotators to generate landmarks, including 13 individuals in total. All annotators have previous experience from working directly with WSIs in a research context and have received corresponding training. Two of the annotators have pathologist training. Landmark annotations were generated using a customised version of TissUUmaps18. Image pairs in the validation data were annotated by one annotator, whereas each image pair in the test data was annotated by two annotators in two annotation phases. No annotations were generated for the training data. All annotations were conducted with the original NDPI files at 40X magnification. Counting landmarks from both annotation phases independently, annotators generated 35,760 landmark pairs in total.

In the first phase of the annotation process, which is the same for the validation and test data, annotators were shown one H&E and one IHC stained section side-by-side and were asked to mark 50 corresponding landmarks in both images, inserting first the IHC point and then the H&E point. In order to ensure the precision of annotations, annotators were asked to zoom the field of view to a level such that the TissUUmaps scale bar indicated 250 µm or less.

In the second phase, which was only applied to the test data, annotators were provided with modified annotation files from the first round of annotations. Landmark coordinates in the IHC image were fixed in place, while for the H&E random uniform noise of [−500, 500) pixels (±115 μm) was added to both the X and Y coordinates. Annotators were then asked to move the H&E landmarks to match the corresponding ones in the IHC WSI. Annotators were chosen randomly such that phase one and phase two annotations were created by different observers for each WSI. Detailed annotation guidelines for both phases are available from

Data Records

We published 4,212 WSIs of breast cancer resection specimens stained with H&E or IHC (ER, PGR, HER2, KI67) originating from 1,153 patients on the Swedish National Data Service SND15. All WSIs are provided as pyramidal TIFF files, starting at 10X resolution (ca. 0.92 µm/pixel) and lower resolutions. The naming convention of all WSIs follows the pattern caseid_stain_set.tiff where caseid indicates a randomly generated case ID, stain either H&E or the IHC antibody used, and set whether the file belongs to train, valid or test. Furthermore, there is a CSV table that indicates the microns-per-pixel at the first level in the respective TIFF files, the stain, the IHC antibody and the data split for each file name. This table is summarised in Table 1.

The data set15 is available for download in seven separate ZIP archives, five for the training data ( (71.47 GB), (70.59 GB), (75.91 GB), (71.63 GB) and (69.09 GB)), one for the validation data ( 21.79 GB) and one for the test data ( 68.11 GB). Lists of files and checksums in SHA1 format are available in order to ascertain completeness and integrity of the data set15 after download. Furthermore, a CSV file (df_acrobat_meta.csv) with a table with additional information for each WSI is available, alongside a ReadMe file (df_acrobat_meta_readme.txt) that explains the columns in this table. Table 2 contains these explanations. No additional clinical information beyond these columns is provided.

Table 2 Overview of WSI-wise information that is available for download alongside the ZIP archives that contain the WSIs as a CSV file.

Technical Validation

All WSIs originate from slides that were used in the routine diagnostic workflow. The tissue samples have therefore each been reviewed by at least one specialty pathologist using a microscope during the initial diagnosis. The macro images of all WSIs in the data set15 were reviewed by at least one observer in order to exclude WSIs with multiple tissue sections of the same resection specimen and in order to confirm that H&E and IHC tissue sections show corresponding tissue. All WSIs in the validation set were reviewed by at least one and all WSIs in the test set were reviewed by at least two human annotators during the landmark generation at 40X resolution. All WSIs in the test data were furthermore reviewed by a specialty pathologist while generating annotations for a research project that is independent from the ACROBAT challenge at 40X resolution, confirming their usability. The majority of WSIs in this study have also been used in other research studies, which further supports the validity of the data set15,19. There are several studies that use WSIs that were generated using the same scanners and workflow20,21,22. Some of the WSIs included in the data set15 contain artefacts. These WSIs were deliberately left in the data set15, in order to be able to assess the robustness of suggested registration methods.

The quality of landmarks in the test data can be assessed by computing the distances between the two human annotators. Landmarks with a distance between annotators of more than 115 µm were excluded, which was chosen as a threshold in correspondence to the noise added for the second annotation phase.

Usage Notes

Pyramidal TIFF files are compatible with OpenSlide23 and can e.g. be inspected with QuPath. Registration algorithms typically align WSIs iteratively starting at low resolutions. With OpenSlide, lower resolution versions of the WSI can be obtained through the different levels of the TIFF files, which makes additional computations for downsampling obsolete. ACROBAT Github repository at provides code to visually inspect landmarks, either only in IHC or also paired H&E landmark after registration by the user.

The data set15 is split into a training, a validation, and a test set. To evaluate the performance of registration methods, landmarks for the IHC WSIs in the validation and test data, are released publicly, whereas the target H&E landmarks can be used to quantify registration performance through an automated evaluation tool. Registered validation set landmarks can be submitted at to receive performance metrics based on these landmarks.