Background

Breast cancer is the most commonly diagnosed tumour in women, and its incidence worldwide has seen important increases over time. Most of our understanding of mammary gland development and breast carcinogenesis has derived from experiments in animal models, with mice being the most commonly used model. Nevertheless, a high proportion of potential oncology drugs fail in clinical trials, partly because of probable differences and incompatibilities between the rodent models and humans, raising concerns over the suitability of rodent models for some questions of cancer research. To address this potential gap, patient derived xenografts (PDXs) have emerged as an interesting tool. PDXs are preclinical models developed by transplanting human-derived cells into immunosuppressed or humanized mice, and various studies have shown that they successfully recapitulate the complexity of human tissues, which has resulted in this model gaining momentum for translational research. A further advance overcoming some caveats in PDXs is the Mouse IntraDuctal Model (MIND), which consists on the injection of primary human-derived breast cells (HDBCs) directly into the mouse mammary ductal tree via cleaved teat. In the intraductal microenvironment HDBCs and breast cancer cells grow independently of any hormone supplementations, while retaining hormone responsiveness and hormone receptor signature. A caveat of PDXs and MIND models is that it is important to discriminate between xenografted and host cells, but histological analyses of the resulting tissues are usually confounded by the presence of cells of the different species, which can give rise to data misinterpretation. To circumvent this hurdle, Juppet et al (1) developed a Single Cell Classifier (SCC), a data-driven deep learning-based computational tool which allows performing automated species classification of individual cells in H&E stained sections. It is available as a plugin in ImageJ/Fiji, and can be downloaded at the link: https://github.com/Biomedical-Imaging-Group/SingleCellClassifier.

.

Key findings and developments

Method development for cell classification

Slides of tissue sections prepared for H&E staining were scanned using an Olympus VS120-L100 slide scanner with a 20x/0.75 objective connected to a colour camera. The resulting images were loaded into QuPath, and are publicly available in Zenodo (https://zenodo.org/record/3960270).

The authors then developed a method for data extraction, whereby a version of each image was extracted to define areas likely to contain ducts- defined as large, densely packed cell regions having a dense Eosin signal. This information was used to extract them using Fiji’s Color Deconvolution with the built-in “H&E DAB” vectors. The extracted signal was filtered before thresholding with ImageJ’s Default method. Finally, a connected components analysis (AnalyzeParticles) was used to obtain ROIs. The bounding boxes of these ROIs were extracted and enlarged so that no ducts were touching the edge of the image.

To define boundaries between neighbouring cells, nuclei segmentation was used as a first step, using StarDist. A StarDist model was trained with multiple images extracted from H&E-stained sections of normal human breast xenografted mammary glands. Following this, the estimation of the cells was computed according to two criteria (which can be edited according to user needs: a) cells should not overlap with each other, and b) their thickness (i.e. distance to the nuclei) should not exceed 2µm). These criteria can be represented as masks that can be integrated to delineate each cell. Following this, 484 features (were extracted from the detected nuclei and cells. These included 47 cell-intrinsic features such as shape and size, area, ratio between nuclei and cells, colour-related features (eg. mean variation of each channel), and texture-related features (eg. entropy and contrast). It also included 376 contextual features, (i.e. features of the neighbouring cells), and 61 features related to the organization of the neighbours. Altogether these features included position of a given cell within a cluster given their mean and variance of their distance, defining borders, alignment, orientation, whether the cells were connected to a chain of cells, tortuosity, and size, and shape of the cluster. What motivated the extraction of contextual features was evidence that cells that are physically connected share common features.

To classify the cells based on their features, the authors trained a neural network model by supervised learning using Python. Multiple images were extracted from H&E-stained xenografted mouse mammary glands, containing either only human or only mouse cells. A human-specific E-cadherin antibody was used to uniquely probe for xenografted non-malignant cells. Tumour cells were uniquely detected using a human-specific cytokeratin 7 (CK7) antibody.

An ImageJ/Fiji plugin names “Single Cell Classifier” was implemented to allow users to a) perform classifications using the methods previously described and b) adopt other models in addition to the built-in ones, and/or edit the parameters of the methods in a manner that is user-specific. The plugins depend on 3 other plugins: StarDist, MorphoLibJ and CSBDeep (the latter of which executes the classification neural network).

Classification of normal and malignant tissues

The authors went on to characterize the impact of each feature on the neural network, to assess the morphology of the engrafted human cells in the intraductal environment. Shape and size of individual cells poorly discriminated between human and mouse cells, however, the contextual features for shape and size achieved a better correlation. Moreover, texture features seemed to offer better help for discriminating cell species than shape and size. Yet, contextual features without any texture information still allowed for classification with almost 90% accuracy (while shape and size alone reached almost 67% accuracy). The SCC reached up to 96% classification accuracy between both species in both normal and malignant tissues, across different breast cancer subtypes. They overall conclude that SCC will facilitate the interpretation of H&E stained histological sections in an automated manner, and that it has the advantage that it allows the implementation of new in-house built models for further classification tasks for these or other applications.

 

What I like about this preprint

I like that the authors provide a novel tool to bridge a gap within histology, which is a very broad field with impact on many scientific disciplines. I like that the authors made everything open, which allows for other scientists to apply and/or adapt the model to their own needs. This is consistent with the principles of open access and open science, which I (like many other scientists), favour.

 

Open questions

1. You mention that texture-specific features allowed reaching the highest accuracy. What was the limitation of colour-specific features, particularly if using species-specific antibodies?
2. Your SCC development seems really an excellent tool. What would be further limitations you identified, that authors should be aware of? You made the tool open and editable to adapt for user-specific needs, therefore, what would be specific points you would suggest users to be aware of while considering editing different parameters?
3. You tested and trained the SCC in your work on breast cancer models, allowing you to differentiate between cancer subtypes, and between normal and malignant tissue. How easily applicable are the same parameters to other types of xenograft and other types of cancer modelled in a similar manner in mice?
4. Being a very useful tool, it wouldn’t be surprising if it were also useful to other model organisms. Would you expect much changes would be needed in the adaptation of the SCC for different organisms (than the mouse/human combination you developed it originally for)?

References

1. Juppet et al, Deep learning enables individual xenograft cell classification in histological images by analysis of contextual features, bioRxiv,2020.

 

doi: https://doi.org/10.1242/prelights.26466

Read preprint (No Ratings Yet)