text stringlengths 0 5.79k |
|---|
WSI and its corresponding RoIs. All the RoIs were extracted and annotated with the help of QuPath17 . The number of extracted |
RoIs per WSI ranges from 0 to 119, with an average of 11 RoIs per WSI. Multiple RoIs corresponding to different lesion |
subtypes were selected per WSI such that the collected RoIs across all WSIs collectively encompass the lesion heterogeneity. |
This aspect is very crucial for representing pathogenesis and disease progression, and consequently allowing for the inclusion |
of sufficiently variable data for DL model training. Subsequently, the WSI-level labels were trivially defined as the most |
aggressive tumor subtype annotated in the image. |
Subtype annotations |
An important characteristics of BRACS is the inclusion of atypical lesions, ADH and FEA. While ignored in other public |
datasets, these categories remain important as they might be indicators of either (i) the presence of abnormalities in neighboring |
breast tissue that could go undetected, e.g., due to the extraction of small tissue samples, or (ii) a high risk of onset of future |
carcinoma, i.e., development of DCIS and IC. In addition, these lesions cannot be detected by mammography or other breast |
imaging techniques, nor can they be felt during a clinical breast examination. When detected in a core biopsy, more frequent |
imaging follow-up and often surgical excision are recommended18 . BRACS also includes lesion subtypes belonging to benign |
and malignant types. In particular, benign lesions are subtyped as either non-cancerous lesions (PB) or inflammatory responses |
2/10 |
Figure 1. Example of a WSI and its associated RoIs. The color of each contour indicates the tumor subtype of the lesion |
according to a fixed color palette. |
(UDH). Malignant lesions are categorized as either DCIS or IC. Finally, histology images representing normal tissue sample are |
classified in the Normal (N) category. |
In order to clarify the description of the different tissue subtypes, a brief description of the mammary gland should be |
considered. The breast is a modified apocrine sweat gland, made up of 15-25 independent glandular units called lobes, each of |
which is formed by a compound tubulo-acinar gland. The lobes are comprised of adipose tissue and divided by connective |
tissue septa. Inside each lobe, the main ducts branch into terminal ducts, each of which leads to a lobule that is made up of |
many berries to form the ductulo-lobular terminal unit. Detailed information on lesions included in BRACS can be found |
in19 . The specific features of the different sample tissue subtypes are shortly summarized in the following and a representative |
example for each of them is shown in Figure 2. |
Normal Tissue (N) |
In normal mammary glandular tissue, there are two types of epithelial cells (the luminal layer and the basal myoepithelial layer) |
and two types of stromal cells (interlobular stroma and intralobular stroma). Differently from PB, the ratio between epytelial |
component and stroma is preserved. |
Pathological Benign (PB) |
Benign breast lesions can be grouped according to the risk of developing invasive carcinoma and include several groups |
of histological entities classified in relation to morphology. In our study, for reasons of differential diagnosis, in the PB |
category we included both non-proliferative lesions and proliferative lesions with the exception of UDH, FEA and ADH, which |
were considered as three independent subtypes. Therefore, PB includes cyst, apocrine metaplasia, ductal ectasia, squamous |
metaplasia, atrophy, stromal fibrosis, mastitis, sclerosing adenosis, papilloma, radial scar, simple and complex fibroadenoma. |
Usual Ductal Hyperplasia (UDH) |
UDH is characterized by an increase in the epithelial layers. It is a cohesive proliferation of disorderly distributed but oriented |
cells. It can have different architectural aspects (solid pattern, fenestrated pattern, micropapillary pattern). Even if UDH shares |
some architectural features with ADH and DCIS, it does not show atypia. |
Flat Epithelial Atypia (FEA) |
FEA is a proliferative lesion characterized by low grade cytological atypia, cell monomorphism, loss of polarity and orientation |
with respect to the basement membrane, presence of apical snout, endoluminal secretion and frequent calcifications. |
Atypical Ductal Hyperplasia (ADH) |
ADH is a proliferation of monomorphic cells, which only partially fill the ductal spaces. Architectural aspects include a solid |
pattern, a cribriform pattern and a papillary pattern. The cytologic atypia is similar to that of low-grade DCIS, but the lesion |
3/10 |
(a) (b) (c) |
(d) (e) |
(f) (g) |
Figure 2. Examples of different tissue samples: (a) Normal (N), (b) Pathological Benign (PB), (c) Usual Ductal Hyperplasia |
(UDH), (d) Flat Epithelial Atypia (FEA), (e) Atypical Ductal Hyperplasia (ADH), (f) Carcinoma in Situ (DCIS) and (g) |
Invasive Carcinoma (IC). |
4/10 |
spans no more than 2mm or has an insufficient architectural atypia involving only partially ducts and / or lobules. |
Ductal Carcinoma in Situ (DCIS) |
In situ carcinoma is a malignant proliferation of epithelial cells that fills the entire duct, without evidence of stroma invasion. |
Typically it involves multiple adjacent ductal spaces. It can have cribriform, solid, papillary and micropapillary patterns. |
Invasive Carcinoma (IC) |
IC is characterized by the invasion of tumor cells infiltrating the breast stroma with loss of peripheral myoepithelial cells. The |
presence of the myoepithelial cell layer is an important distinction of DCIS from IC. |
Detecting certain subtypes is particularly challenging as some morphological patterns can be shared by several classes. For |
instance, ADH shares morphological similarities with DCIS. In certain cases, it even includes all the features of DCIS, but is |
simply limited in size. Also, UDH, ADH and DCIS are all characterized by an intraductal growth pattern, which makes these |
classes difficult to classify and differentiate in H&E-stained sections. This is particularly true when pathologists cannot utilize |
significant WSI-level context. We emphasize that when classifying BRACS RoIs, no context information is provided beside the |
tissue lesion and immediate tissue micro-environment. |
BRACS dataset characteristics |
The BRACS dataset contains 547 WSIs extracted on 189 different patients. It also includes 4’539 RoIs extracted from 387 WSI |
collected on 151 patients. All slides were scanned with an Aperio AT2 scanner at 0.25 µm/pixel using a magnification factor |
of 40×. Table 1 and Table 2 report the number of WSIs (with and without RoIs) and RoIs according to the lesion type and |
subtypes, respectively. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.