johnseenun/thesis-data · Data Studio · Datasets at Hugging Face

text stringlengths 0 5.79k
Data Benign Atypical Malignant Total
WSIs with RoIs 149 75 163 387
WSIs without RoIs 116 14 30 160
WSIs 265 89 193 547
RoIs 1837 1263 1439 4539

Table 1. BRACS data distribution according to lesion type.


Data N PB UDH FEA ADH DCIS IC
WSIs with RoIs 17 77 55 34 41 51 112
WSIs without RoIs 27 70 19 7 7 10 20
WSIs 44 147 74 41 48 61 132
RoIs 484 836 517 756 507 790 649

Table 2. BRACS data distribution according to lesion subtype.

Comparison with existing datasets
In recent years, several data sets have been proposed, with more and more samples and classes20–25 . Table 3 details existing
public datasets of histology images for breast lesion classification. Datasets that are (i) either no more accessible21 , (ii)
subsets of already mentioned datasets22 , or (iii) targeting specific tasks, e.g., lesion proliferation scores prediction24 , pN-stage
prediction25 , are not mentioned in Table 3.
The IDC12 and Camelyon1623 datasets are focusing on the detection of the presence of a given lesion. In particular,
IDC provides RoIs at small spatial resolution (50 × 50 pixels) extracted from large areas of Invasive Ductal Carcinoma. We
emphasize that even if the number of RoIs in BRACS is lower than in IDC, BRACS images are on average much larger allowing
the inclusion of whole glandular areas. BRACS includes a larger number of WSIs than Camelyon16 and more subtypes.
BreakHis26 and BACH20 datasets are devoted to multi-classification tasks, but remain significantly smaller than BRACS, both
in terms of image size and number of samples, and include less subtypes than BRACS.
In conclusion, the number of WSIs, RoIs and associated patients, is significantly larger in BRACS compared to the existing
public datasets. Moreover, BreakHis and BACH include fixed-size RoIs, while BRACS images are of arbitrary size. Assuming
fixed-size tumor regions is a strong assumption that does not apply in real-life scenarios. Limiting the size of RoIs imposes to
either (i) partially cut the lesion, hence producing a loss of information that could be pivotal for a correct diagnosis, or (ii)
manually curating RoIs such that they all have similar sizes, which does not encompass tumor heterogeneity. By proposing


5/10
Lesion Classes Data Type Resolution
Dataset, Year n. Pat.
Benign Malignant Size (Magnification) in pixels
Invasive Ductal Carcinoma RoI
IDC negative IDC positive 162 50x50
(IDC)12 , 2014 277.524 (40x)
RoI
Adenois Carcinoma
1.995 (40x)
Fibroadenoma Lobular carcinoma
BreakHis13 , 2015 2.081 (100x) 82 700x460
Phyllodes tumor Mucinous carcinoma
2.013 (200x)
Tubular adenona Papillary carcinoma
1.820 (400x)
Normal In-situ carcinoma WSI
Camelyon1623 , 2016 400 Variable Size
Benign invasive carcinoma 400 (20x and 40x)
RoI
39 2048×1536
Normal In-situ carcinoma 400 (200x)
BACH20 , 2018
Benign invasive carcinoma WSI
10 Variable Size
10(20x)

Table 3. Popular publicly available breast histopathology image dataset.

Lesion Classes Data Type Resolution
Dataset, Year n. Pat.
Benign Atypical Malignant Size (Magnification) in pixels
RoI
Normal 151 Variable size
FEA In-situ carcinoma 4537 (40x)
BRACS27 , 2021 Benign
ADH invasive carcinoma WSI
UDH 189 Variable size
547 (40x)

Table 4. General information on BRACS dataset


samples of varying sizes, BRACS promotes the development of DL algorithms that need to be able to operate on inputs of
different dimensionality. BRACS is also the first dataset to include pre-cancerous atypical lesions. BRACS and BACH share
the same malignant lesion subtypes, while BreakHis refines this class by partitioning it into four specific subtypes. All the
benign lesion subtypes defined in BACH and BreakHis are also included in BRACS (normal and benign). In addition, BRACS
includes UDH lesion subtype that is not considered in BACH and BreakHis.
In summary, BRACS characteristics are unique, as it allows for multi-class classification task of breast cancer lesions,
including challenging atypical lesions. BRACS is also larger than the existing datasets in terms of number RoIs, WSIs and
patients.

Dataset splitting
To foster reproducibility and following ML best practices, we provide pre-defined WSI- and RoI-level splits in training,
validation and test sets. Data split was generated such that all the WSIs extracted from a patient belong to the same set.
Similarly, all the RoIs extracted in a given WSI are assigned to the same split. By following this approach, we avoid that
different sets include correlated patient- and slide-level information, that could be leading to overly optimistic prediction
results28 . Table 5 and Table 6 present the number of WSIs included in the train, validation, and test splits. Information about
the number of patients in each set is provided in Table 5. Equivalent information for RoIs is shown in Table 7 and Table 8. At
RoI-level, the ratio between the most common subtype (PB with 714 samples) and the least common one (N with 357) is around
two. At WSI-level, the most common subtype is PB with 120 samples and the least common one is FEA with 24 samples,
hence of ratio of around four. Considering the constraint of patient- and WSI-level split, and the fact that atypical lesions are
more rare than malignant ones, BRACS offers a rather balanced set, that can directly be used for training DL systems.

Benign Atypical Malignant Total WSIs Total Patients

Data Benign Atypical Malignant Total

WSIs with RoIs 149 75 163 387

WSIs without RoIs 116 14 30 160

WSIs 265 89 193 547

RoIs 1837 1263 1439 4539

Table 1. BRACS data distribution according to lesion type.

Data N PB UDH FEA ADH DCIS IC

WSIs with RoIs 17 77 55 34 41 51 112

WSIs without RoIs 27 70 19 7 7 10 20

WSIs 44 147 74 41 48 61 132

RoIs 484 836 517 756 507 790 649

Table 2. BRACS data distribution according to lesion subtype.

Comparison with existing datasets

In recent years, several data sets have been proposed, with more and more samples and classes20–25 . Table 3 details existing

public datasets of histology images for breast lesion classification. Datasets that are (i) either no more accessible21 , (ii)

subsets of already mentioned datasets22 , or (iii) targeting specific tasks, e.g., lesion proliferation scores prediction24 , pN-stage

prediction25 , are not mentioned in Table 3.

The IDC12 and Camelyon1623 datasets are focusing on the detection of the presence of a given lesion. In particular,

IDC provides RoIs at small spatial resolution (50 × 50 pixels) extracted from large areas of Invasive Ductal Carcinoma. We

emphasize that even if the number of RoIs in BRACS is lower than in IDC, BRACS images are on average much larger allowing

the inclusion of whole glandular areas. BRACS includes a larger number of WSIs than Camelyon16 and more subtypes.

BreakHis26 and BACH20 datasets are devoted to multi-classification tasks, but remain significantly smaller than BRACS, both

in terms of image size and number of samples, and include less subtypes than BRACS.

In conclusion, the number of WSIs, RoIs and associated patients, is significantly larger in BRACS compared to the existing

public datasets. Moreover, BreakHis and BACH include fixed-size RoIs, while BRACS images are of arbitrary size. Assuming

fixed-size tumor regions is a strong assumption that does not apply in real-life scenarios. Limiting the size of RoIs imposes to

either (i) partially cut the lesion, hence producing a loss of information that could be pivotal for a correct diagnosis, or (ii)

manually curating RoIs such that they all have similar sizes, which does not encompass tumor heterogeneity. By proposing

5/10

Lesion Classes Data Type Resolution

Dataset, Year n. Pat.

Benign Malignant Size (Magnification) in pixels

Invasive Ductal Carcinoma RoI

IDC negative IDC positive 162 50x50

(IDC)12 , 2014 277.524 (40x)

RoI

Adenois Carcinoma

1.995 (40x)

Fibroadenoma Lobular carcinoma

BreakHis13 , 2015 2.081 (100x) 82 700x460

Phyllodes tumor Mucinous carcinoma

2.013 (200x)

Tubular adenona Papillary carcinoma

1.820 (400x)

Normal In-situ carcinoma WSI

Camelyon1623 , 2016 400 Variable Size

Benign invasive carcinoma 400 (20x and 40x)

RoI

39 2048×1536

Normal In-situ carcinoma 400 (200x)

BACH20 , 2018

Benign invasive carcinoma WSI

10 Variable Size

10(20x)

Table 3. Popular publicly available breast histopathology image dataset.

Lesion Classes Data Type Resolution

Dataset, Year n. Pat.

Benign Atypical Malignant Size (Magnification) in pixels

RoI

Normal 151 Variable size

FEA In-situ carcinoma 4537 (40x)

BRACS27 , 2021 Benign

ADH invasive carcinoma WSI

UDH 189 Variable size

547 (40x)

Table 4. General information on BRACS dataset

samples of varying sizes, BRACS promotes the development of DL algorithms that need to be able to operate on inputs of

different dimensionality. BRACS is also the first dataset to include pre-cancerous atypical lesions. BRACS and BACH share

the same malignant lesion subtypes, while BreakHis refines this class by partitioning it into four specific subtypes. All the

benign lesion subtypes defined in BACH and BreakHis are also included in BRACS (normal and benign). In addition, BRACS

includes UDH lesion subtype that is not considered in BACH and BreakHis.

In summary, BRACS characteristics are unique, as it allows for multi-class classification task of breast cancer lesions,

including challenging atypical lesions. BRACS is also larger than the existing datasets in terms of number RoIs, WSIs and

patients.

Dataset splitting

To foster reproducibility and following ML best practices, we provide pre-defined WSI- and RoI-level splits in training,

validation and test sets. Data split was generated such that all the WSIs extracted from a patient belong to the same set.

Similarly, all the RoIs extracted in a given WSI are assigned to the same split. By following this approach, we avoid that

different sets include correlated patient- and slide-level information, that could be leading to overly optimistic prediction

results28 . Table 5 and Table 6 present the number of WSIs included in the train, validation, and test splits. Information about

the number of patients in each set is provided in Table 5. Equivalent information for RoIs is shown in Table 7 and Table 8. At

RoI-level, the ratio between the most common subtype (PB with 714 samples) and the least common one (N with 357) is around

two. At WSI-level, the most common subtype is PB with 120 samples and the least common one is FEA with 24 samples,

hence of ratio of around four. Considering the constraint of patient- and WSI-level split, and the fact that atypical lesions are

more rare than malignant ones, BRACS offers a rather balanced set, that can directly be used for training DL systems.

Benign Atypical Malignant Total WSIs Total Patients