Source-linked AI summary
BRACS: A Dataset for BReAst Carcinoma Subtyping in H&E Histology Images
Nadia Brancati, Anna Maria Anniciello, Pushpak Pati, Daniel Riccio, Giosuè Scognamiglio, Guillaume Jaume, Giuseppe De Pietro, Maurizio Di Bonito, Antonio Foncubierta, Gerardo Botti, Maria Gabrani, Florinda Feroce, Maria Frucci
TL;DR
Breast histology analysis is cumbersome and variable, while AI requires large, heterogeneous, well-annotated datasets. BRACS addresses this gap with annotated H&E images spanning breast-lesion categories, including atypical lesions, and is larger than existing datasets at WSI and ROI levels.
Problem
AI-based breast histology analysis requires large, heterogeneous annotated datasets, but existing datasets often cover limited lesion classes and standardized images lacking real-world artifacts.
Method
BRACS curates WSI- and ROI-level H&E images, derives WSI labels from the most severe detected lesion, and uses annotations reviewed by three board-certified pathologists.
Results
BRACS contains 547 WSIs and 4539 ROIs, with benign, malignant, and atypical lesion types further subdivided into seven categories; it is larger than existing datasets at WSI, ROI, and patient levels.
Takeaways & Limitations
Including atypical lesions and variable-size, heterogeneous images provides an opportunity to use AI to study challenging breast lesions in settings closer to pathology practice.
Takeaways & Limitations
ROI classification is performed without WSI-level context beyond the tissue lesion and its immediate micro-environment, complicating differentiation among morphologically similar subtypes.
Abstract
from arXiv · showhide
Breast cancer is the most commonly diagnosed cancer and registers the highest number of deaths for women with cancer. Recent advancements in diagnostic activities combined with large-scale screening policies have significantly lowered the mortality rates for breast cancer patients. However, the manual inspection of tissue slides by the pathologists is cumbersome, time-consuming, and is subject to significant inter- and intra-observer variability. Recently, the advent of whole-slide scanning systems have empowered the rapid digitization of pathology slides, and enabled to develop digital workflows. These advances further enable to leverage Artificial Intelligence (AI) to assist, automate, and augment pathological diagnosis. But the AI techniques, especially Deep Learning (DL), require a large amount of high-quality annotated data to learn from. Constructing such task-specific datasets poses several challenges, such as, data-acquisition level constrains, time-consuming and expensive annotations, and anonymization of private information. In this paper, we introduce the BReAst Carcinoma Subtyping (BRACS) dataset, a large cohort of annotated Hematoxylin & Eosin (H&E)-stained images to facilitate the characterization of breast lesions. BRACS contains 547 Whole-Slide Images (WSIs), and 4539 Regions of Interest (ROIs) extracted from the WSIs. Each WSI, and respective ROIs, are annotated by the consensus of three board-certified pathologists into different lesion categories. Specifically, BRACS includes three lesion types, i.e., benign, malignant and atypical, which are further subtyped into seven categories. It is, to the best of our knowledge, the largest annotated dataset for breast cancer subtyping both at WSI- and ROI-level. Further, by including the understudied atypical lesions, BRACS offers an unique opportunity for leveraging AI to better understand their characteristics.
Background & Summary
Breast histology analysis is difficult because images are complex, ambiguous, and large, while existing datasets incompletely represent diagnostic diversity and real-world artifacts. BRACS addresses these gaps with heterogeneous, realistic WSI- and RoI-level images spanning typical and atypical breast lesions.
- Histology images are complex, ambiguous, and large, making robust, reproducible, and efficient analysis challenging.These properties make manual and computational analysis cumbersome and time-consuming.
- Existing breast cancer datasets often use limited patient cohorts and mainly classify lesions as benign or malignant.Such class sets do not capture the complete diagnostic spectrum.
- Many public datasets lack clinical artifacts such as staining anomalies, ink marks, tissue folds, blurred regions, and tears.This limits their representation of real-world breast cancer diagnosis.
- BRACS provides heterogeneous, realistic H&E images at both WSI and RoI levels, with variable-sized RoIs that retain entire diagnostic lesions.Images are acquired from many patients and include typical as well as atypical lesion categories.
- BRACS covers seven lesion categories: Normal, Pathological Benign, UDH, FEA, ADH, DCIS, and IC.The categories span normal, benign, atypical, and malignant breast lesions.
Methods
BRACS was assembled as a heterogeneous breast histology dataset with WSI- and RoI-level annotations, multiple lesion subtypes, and atypical lesions. Its curation combines expert consensus annotation with patient- and slide-aware dataset organization.
- Annotation procedure: Three board-certified pathologists independently annotated representative RoIs, then resolved disagreements through consensus re-annotation.RoIs were labeled as normal tissue or one of six lesion subtypes.
- Annotation procedure: WSI-level labels were derived from the most severe cancerous lesion detected within each slide.A WSI can contain several lesions of different subtypes, so the slide label summarizes the most aggressive annotated tumor subtype.
- Subtype scheme: BRACS includes normal, pathological benign, UDH, FEA, ADH, DCIS, and IC lesion categories.Benign lesions are separated into PB and UDH, atypical lesions into FEA and ADH, and malignant lesions into DCIS and IC.
- Subtype scheme: The dataset includes atypical lesions because ADH and FEA are important categories that are often ignored in other public datasets.These lesions may indicate neighboring abnormalities or elevated future carcinoma risk and are not detectable by mammography or clinical examination according to the paper.
- Dataset characteristics: BRACS contains 547 WSIs from 189 patients and 4,539 RoIs from 387 WSIs collected on 151 patients.Slides were scanned at 0.25 µm/pixel with 40× magnification.
- Dataset organization: BRACS provides variable-sized RoIs that include the complete diagnostic lesion and pre-defined WSI- and RoI-level splits.Patient-derived WSIs remain in one split, and RoIs from a WSI are assigned to the same split to avoid correlated information across sets.
Data organization
BRACS separates whole-slide images, regions of interest, and annotations into organized, downloadable folders. The structure uses dataset splits and lesion-subtype labels for both WSIs and RoIs.
- Registered users can freely download BRACS under a Creative Commons CC0 license via the dataset website and FTP server.
- WSIs are divided into train, validation, and test folders, each partitioned into Benign, Atypical, and Malignant categories with subtype folders.
- WSIs use .svs files whose folder names define labels, while RoIs use .png files encoding the source slide, subtype, and RoI number.
- The WSI annotation folder follows the same organization and stores annotation files in .qpdata format.
- Figure 3 depicts the organization of BRACS dataset folders.
Author contributions statement
The project involved clinical, computational, and industrial research institutions, with authors contributing to conceptualization, execution, management, clinical databases, and annotation.
- The National Cancer Institute–IRCCS-Fondazione Pascale contributed authors from its Naples, Italy institution.
- The Institute for High Performance Computing and Networking–CNR contributed authors from Naples, Italy.
- IBM Zurich Research Lab contributed authors from Zurich, Switzerland.
- All authors provided critical revision feedback and participated actively in manuscript preparation.
- G.B., G.D.P., and M.G. handled conceptualization, while A.F. and M.F. handled conceptualization, execution, and project management.
- M.D.B. managed institutional clinical databases and contributed to annotation, while A.M.A. and F.F. contributed to clinical design.