Table 4 Dataset summary

Set	Usage	Source	Private	Year	#Subjects	Subject categories	Clinico-pathology	Number of images	Manufacturer	View
1	SSL,SL	Lund	Yes	2009–2012	778	Cancer	Yes	4790	99% GE	CC, MLO ML
2	SSL, SL	Malmö	Yes	2010–2017	750	Cancer	Yes	10,783	99% Siemens	CC, MLO ML
3	Test	Kristianstad	Yes	2017	103	Cancer	Yes	473	100% GE	CC, MLO ML
4	SSL	CBIS-DDSM^a	No	1999	1566	Cancer, normal	No	3032	Unknown	CC, MLO
5	SSL	MIAS^b	No	2003	161	Cancer, normal	No	322	Unknown	MLO
6	SSL	INBreast^c	No	2011	115	Cancer, normal	No	410	Unknown	CC, MLO
7	SSL	MBTST-DM^d	No	2010–2015	14,669	Normal	No	59,365	100% Siemens	CC, MLO

Seven datasets from three acquired and four open sources were collected for this study. Supervised learning (SL) was developed and tested on sets 1–2, where both mammograms and clinical predictors and outcomes were available. Set 3 was the external test set to evaluate SL. Sets 4–7, with large number of unlabeled mammograms, were only used for self-supervised learning (SSL). In sets 4–6, cancer patients comprised the majority. For sets 1–2, the numbers of subjects and images refer to those used for SSL, while detailed patient selection for the SL cohort is presented in Fig. 1.
GE general electric, CC craniocaudal, ML mediolateral, MLO mediolateral oblique.
^aCurated Breast Imaging Subset of Digital Database of Screening Mammography⁵⁷.
^bMammographic Imaging Analysis Society database⁵⁸.
^cInBreast database⁵⁹.
^dMalmö Breast Tomosynthesis Screening Trial—Digital Mammography database⁵⁵.

Quick links

Search