Normal breast tissue classifiers assess large-scale tissue compartments with high accuracy
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Cancer research emphasises early detection, yet quantitative methods for analysing normal tissue remain limited. Hematoxylin and eosin (H&E)-stained tissues in digitised whole slide images (WSIs) enable computational histopathology; however, artificial intelligence (AI)-based analyses of normal breast tissue (NBT) remain scarce.
Methods
We curated 70 WSIs of NBTs from multiple sources with pathologist-guided manual annotations of epithelium, stroma, and adipocytes, and developed robust convolutional neural network (CNN)-based, patch-level classification models, named NBT-Classifiers , to tessellate and classify NBTs at different scales. Data and code are available at https://github.com/cancerbioinformatics/OASIS and https://github.com/cancerbioinformatics/NBT-Classifier .
Findings
Across three external cohorts, NBT-Classifiers trained on 128 x 128 µm and 256 x 256 µm patches achieved AUCs of 0·98–1·00. Two explainable artificial-intelligence (AI)- visualisation techniques confirmed the biological relevance of tissue class predictions. An end-to-end WSI pre-processing framework was then integrated, capable of localising lobules and peri-lobular stroma. The outputs are compatible with the QuPath v0.3.0 platform and enable downstream digital image analysis, such as texture and nuclei morphology assessment, at the patch level.
Interpretation
NBT-Classifiers represent a robust, generalisable, end-to-end deep learning framework, which will enable broader application in studies of normal tissues, in the context of breast and breast cancer.
Funding
The Breast Cancer Research Trust, Breast Cancer Now (and their legacy charity Breakthrough Breast Cancer), the Medical Research Council (MRC) [MR/X012476/1], Cancer Research UK [CRUK/07/012, KCL-BCN-Q3], and CRUK City of London Centre Award [CTRQQR-2021/100004].
Research in context
Evidence before this study
We searched PubMed from Jan 1, 2013, to Mar 20, 2025, using the query (("normal breast" OR "terminal duct lobular unit" OR "TDLU" OR "acini") AND ("artificial intelligence" OR "deep learning" OR "convolutional neural network" OR "CNN" OR "machine learning" OR "automation" OR "classification" OR "segmentation" OR "detection")) AND (journal article[pt] NOT review[pt]) NOT ("computed tomography" OR "CT" OR "ultrasound" OR "sonography" OR "US" OR "mammography" OR "mammogram" OR "Raman" OR "single-cell" OR "MRI"). This yielded 264 initial hits and after a thorough manual review, only eight studies were found to be relevant for using deep learning or machine learning to automate tissue recognition or quantification of normal breast tissues (NBTs) on digitised histological images. All these methods are highly specialised for detecting and quantifying lobules or quantifying tissue composition, and none of them are specifically tailored for providing patch-level tissue classifications on WSIs of NBTs. Moreover, NBTs are heavily underrepresented in large annotated WSI databases. A recent literature review summarised publicly available breast hematoxylin and eosin (H&E) WSI datasets from 2015 to 2023, identifying 17 datasets comprising a total of 10,385 breast H&E WSIs. Of these, two datasets contain female normal breast WSIs with pathology-guided manual annotations.
Added value of this study
Motivated by the research gap, we collected a comprehensive WSI dataset that captures real-world variability present across NBTs and provided by expert ground-truth manual annotations. Based on extensive cross-validation and external testing, we developed robust patch-level classification models to classify three major tissue compartments within NBTs, at different scales.
Implications of all the available evidence
With high AUCs of 0·98–1·00 across three external cohorts, our artificial-intelligence (AI) tool can be used to automatically classify tissue compartments and localise lobular regions for downstream normal breast research. These approaches have the potential to enhance our understanding of how various NBT components contribute to both benign and malignant breast pathology and lay the groundwork for the development of more advanced deep learning models and spatial-defined molecular large-scale analyses in the future.
