AutomatiC Registration Of Breast cAncer Tissue 2023

The ACROBAT challenge aims to advance the development of whole-slide-image (WSI) registration algorithms that can align WSIs of tissue sections from the same tumour that were stained with immunohistochemistry (IHC) or haematoxylin and eosin (H&E). In order to evaluate usefulness of developed methods on real-world data, all slides in this challenge originate from routine diagnostic workflows. Registration across multiple stains can be an enabling technology both for research and diagnostics. There are e.g. several studies [1]–[4] that use IHC as a local label to train subsequent models to predict from H&E WSIs. In diagnostic settings, it can be of critical importance to overlay WSIs of differently stained tissue e.g. to evaluate resection borders for invasive cancer. In the case of residual tumor in resection margins, additional treatment such as another surgery might be appropriate. Unfortunately, WSI registration is particularly challenging due to the gigapixel scale of these images. Furthermore, a few micrometers thin tissue sections easily deform during sample preparation and it is possible that parts of the tissue in WSIs of the same sample are not contained in other WSIs of that same sample, particularly if sections are not consecutive. The appearance of tissue can also differ considerably depending on the stain. Multi-stain WSI registration is therefore an active field of research and has previously been addressed in the ANHIR[5] challenge and more recently the ACROBAT challenge (acrobat.grand-challenge.org/), which was held in conjunction with MICCAI 2022. In the first ACROBAT challenge, the objective was to register WSIs of tissue sections that were stained with one of the four routine diagnostic IHC stains ER, PGR, HER2 and KI67 to corresponding WSIs of tissue sections stained with H&E. Details of the data set are available in [6]. The first round of the ACROBAT challenge was successful in establishing the state-of-the-art in WSI registration on WSIs from routine diagnostic slides. The top-performing methods exceeded our expectations both with regards to accuracy, as well as to robustness. The code for most of these methods is available online. A particularly interesting finding from the first round of the ACROBAT challenge was that conceptually very different registration methods can perform equally well in this task. Now, we would like to propose a second round of the challenge that aims to investigate how well these different methods perform under domain shifts. More specifically, we would like to estimate how well methods generalise to unseen IHC stains that were not part of the previously published data. To facilitate the ACROBAT 2023 challenge, we will annotate a new test set consisting of undisclosed and unpublished IHC stains for 200 test cases, each consisting of two WSIs that are to be registered, including IHC-IHC and IHC-HE image pairs. Instead of registered landmarks, participants will this time be asked to submit docker containers of their methods. Participants will therefore be completely blinded to the test data, which will enable an investigation of robustness of the registration methods under domain shift and unknown conditions. [1] W. Bulten et al., “Epithelium segmentation using deep learning in H&E-stained prostate specimens with immunohistochemistry as reference standard,” Sci. Rep., vol. 9, no. 1, p. 864, Jan. 2019. [2] D. Tellez et al., “Whole-Slide Mitosis Detection in H&E Breast Histology Using PHH3 as a Reference to Train Distilled Stain-Invariant Convolutional Networks,” IEEE Trans. Med. Imaging, Mar. 2018, doi: 10.1109/TMI.2018.2820199. [3] P. Gamble et al., “Determining breast cancer biomarker status and associated morphological features using deep learning,” Communications Medicine, vol. 1, no. 1, pp. 1–12, Jul. 2021. [4] M. Valkonen et al., “Cytokeratin-Supervised Deep Learning for Automatic Recognition of Epithelial Cells in Breast Cancers Stained for ER, PR, and Ki-67,” IEEE Trans. Med. Imaging, vol. 39, no. 2, pp. 534–542, Feb. 2020. [5] J. Borovec et al., “ANHIR: Automatic Non-Rigid Histological Image Registration Challenge,” IEEE Trans. Med. Imaging, vol. 39, no. 10, pp. 3042–3052, Oct. 2020. [6] P. Weitz et al., “ACROBAT -- a multi-stain breast cancer histological whole-slide-image data set from routine diagnostics for computational pathology,” arXiv [eess.IV], Nov. 24, 2022. [Online]. Available: http://arxiv.org/abs/2211.1362