Artificial Intelligence (AI) for the early detection of breast cancer: a scoping review to assess AI’s potential in breast screening practice

Figures & data

Table 1. Summary of the characteristics of studies reporting on artificial intelligence (AI) in breast cancer detection.

Study first author(s)	Purpose or aim (s)	Study design	Data Source for AI development or evaluation	Type of data (input variables where stated)	Number of subjects or images (N)	Subjects’ mean or median age	Breast cancer proportion in study	Included consecutive subjects or screens?	Reference Standard
Rodriguez-Ruiz [9]	To compare stand-alone performance of a commercially available AI system to that of radiologists in detecting BC on DM	Retrospective (image databases)	DM exams collected during previous reader studies	DM	2652 DMs	Various ranges across datasets	24.6%	No	Datasets verified with pathology or ≥1 year follow-up
Al-Masni [10]	Describe and evaluate a CAD system based on ROI (region of interest) deep learning techniques, using a CNN referred to as You Only Look Once (YOLO), to detect and classify breast masses on DM.	Retrospective (image database)	DDSM	DM (lesion feature, mammographic mapping)	600 DM from 150 women	NR	50.0%	No	NS, sourced classified database
Ribli [11]	Propose a CAD system based on deep CNNs to detect and classify lesions on mammograms	Retrospective (image database)	DDSM, Inbreast image datasets plus institutional images	Digital and digitized film mammograms	2949 mammograms	NR	NR	No	NS, several image databases
Chougrad [12]	To develop a CAD system by refining existing models using CNN, exploring a transfer learning method, to help classify mammography mass lesions.	Retrospective (image databases)	Merged subsets from several image databases: DDSM, BC Digital Repository, and Inbreast databases.	Digital, including digitized film, mammograms (pixel information)	6116 images from 1529 subjects (selected benign or cancer cases, excluded normal)	NR	51.0%	No	NS, sourced classified database (many biopsy-proven)
Bandeira-Diniz [13]	To develop a model to classify breasts on mammography into dense and non-dense breasts, and to develop another model to classify regions of the breast into mass and non-mass regions based on symmetry in dense and non-dense breasts separately using CNN.	Retrospective (imaging database)	DDSM	DM (pixel information, some demographic information)	2,482 images from 1241 women (80% used for model training/ 20% for testing)	NR	NR	No (only images with at least one mass were chosen)	NS, sourced classified database.
Becker [14]	To evaluate the diagnostic accuracy of a multipurpose image analysis software based on deep learning with artificial neural networks for the detection of BC in an independent, dual-centre mammography data set.	Retrospective (image database)	Mammograms from an institution during one year; and BC Digital Repository data set	DM (lesion feature, mammographic mapping)	1146 subjects (images NR)	59.6	7.0%	No	Histology or clinical follow-up
Lotter [15]	To develop a CAD system using CNN to classify whether BC is present on the mammogram	Retrospective (image database)	DDSM	DM	10,480 images from 2620 women	NR	NR	No	Sourced classified database, includes pathology outcomes
Becker [16]	To train a generic deep learning software to classify breast cancer on ultrasound images, and to compare its performance to human readers with variable breast image-reading experience.	Retrospective (examinations done in one year)	Still images from ultrasound examinations in one institution (not real-time ultrasound)	Ultrasound images	632 patients (70% used for training/30% for validation)	53 (SD: 15)	13.0%	No (18% of initial cohort included); those with normal US or benign lesions excluded	Histology or 24-month clinical follow-up
De Oliviera Silva [17]	To describe and test use of independent component analysis (compared with principal component analysis) in CAD for detecting lesions in dense breasts	Retrospective (image database)	Selected images from DDSM and Mini-MIAS	DM (pixel information)	216 images from 322 patients	NR	NR	No	NS, sourced classified database
Kooi [18]	To develop a model using deep CNN to detect malignant lesions on mammograms and to compare this to a mammography CAD system (based on manually designed features); to assess to what extent clinical information (location, context features, and age) improves model accuracy; and to compare its performance against human readers.	Retrospective (imaging database)	A large-scale screening program in the Netherlands.	DM (pixel information and some clinical information)	40,506 images from 6729 subjects (training), 4303 images from 745 subjects (validation), 18,453 images from 2188 subjects (test set)	NR (women >50 years)	4.7%	NR	Biopsy for cancers, or follow-up
Samala [19]	To translate knowledge learned from non-medical images to medical diagnostic tasks using a multi-task transfer learning DCNN through supervised training of DCNNs (comparing multi-task approach to single-task transfer learning classification methods), and apply this approach in CAD of breast cancer	Retrospective (image database)	DDSM, and mammograms from University of Michigan Health System	Digitized screen-film mammograms (SFM) and DMs	2,242 images (containing 1,057 malignant, 1,397 benign findings) (number of subjects NR)	51.7	43.0% (lesion-based estimate)	No	NS, sourced classified database
Carneiro [20]	To develop a model using deep CNN to assess whether unregistered mammographic images (cranio-caudal and medio-lateral oblique views from each breast) can be classified as containing malignant lesions, benign lesions or only normal tissue.	Retrospective (imaging databases)	Inbreast, DDSM and ImageNet (ImageNet used for model training) databases	DM (BI-RADS class plus lesion delineation (shape and features))	410 + 680 images (Inbreast + DDSM) from 287 subjects (ImageNet has non-medical imaging data)	NR	NR	NR	NS, sourced classified databases
Teare [21]	To present a machine-learning based algorithm which utilizes novel techniques (deep CNN) plus false colour-enhancement technique) to detect malignant lesions in digital mammographic images, and to achieve accuracy similar to that of expert radiologists	Retrospective (imaging database)	DDSM and Zebra Mammography Dataset (ZMDS)	DM	761 images from 586 women	NR	32.0%	NR^#	Histology or 2-year imaging follow-up
Dhungel [22]	To present an integrated methodology for detecting, segmenting and classifying breast masses from mammograms with minimal user intervention.	Retrospective (image database)	Inbreast dataset	DM	410 images from 115 subjects	NR	28.3%	NR^#	NS, sourced classified database
Sun [23]	To propose a semi-supervised learning (SSL) scheme using deep CNN for BC diagnosis, that only requires a small portion of labelled data in training set rather than a large amount of labelled data for training and fine-tuning.	Retrospective (image database)	In-house full-field digital mammography image database	DM (lesion features)	1874 paired mammographic images (subjects NR)	51	45.0% (lesion-based estimate)	NR^#	NR
Saraswathi & Srinivasan [24]	To evaluate a fully complex-valued relaxation neural network (FCRN) based system to identify normal, benign and malignant lesions in digital mammographic images, to improve classification accuracy.	Retrospective (imaging database)	MIAS	DM	322 images (number of subjects NR)	NR	NR; 55% (validation sample)	NR^#	NS, sourced classified database
Velikova [25]	To obtain a balanced view on the role and place of expert knowledge and learning methods in building Bayesian networks for medical image interpretation, using interpretation of mammograms as the example	Retrospective (image database)	Imaging data from the Dutch breast cancer screening program	DM	795 subjects (images NR)	NR	43.3%	NR^#	Cancers verified by pathology reports
Dheeba & Tamil-Selvi [26]	Propose a supervised machine learning algorithm (DEOWNN) for automatic detection of cancerous masses in mammograms.	Retrospective (image database)	MIAS	DM (texture features)	322 images (paired breast images) from 161 subjects	NR	16.0%	NR^#	NS, sourced classified database
Dheeba & Tamil-Selvi [27]	To develop a CAD system to detect microcalcification clusters in digital mammograms using Swarm Optimization Neural Network (SONN).	Retrospective (imaging database)	MIAS; clinical mammogram images for validation	DM (texture features)	322 images (paired breast images) from 161 subjects; (validation: 216 images from 54 subjects)	NR	16.0%	NR^#	NS, sourced classified database
Parmeggiani [28]	To test an Artificial Neural Network (ANN) system developed to detect BC in mammographic and echographic images on a cohort of women being screened for BC, to potentially develop autonomous AI systems.	Prospective cohort	A screening program in a university-affiliated hospital in Naples, Italy.	Mammograms and ultrasound scans	550 subjects (Images NR)	NR (> 40 years)	13.30%	NR^#	Surgery or follow-up
Lesniak [29]	To reduce false positives in mammographic breast cancer screening by using computer aided detection with a support vector machine (SVM) based system	Retrospective (image database)	A BC screening database	Scanned film mammograms (lesion descriptors)	10,064 images from 1539 subjects	NR	32%	NR^#	sourced classified database (cancers biopsy-proven)
Huang [30]	To compare the performance of three different hybrid algorithms, particle swarm optimizer (PSO)-based artificial neural network (ANN), adaptive neuro-fuzzy inference system (ANFIS) and a case-based reasoning (CBR) classifier (latter with a logistic regression (LR) or a decision tree (DT) model), in BC diagnosis.	Retrospective (image database)	Machine Learning Repository Mammographic Mass Data Set	DM (BI-RADS score, lesion descriptors), age	815 subjects (images NR)	NR (range 18–96)	46% (lesion-based estimate)	No	NS, sourced classified database
Ayer [31]	To assess whether an artificial neural network (ANN) trained on a large dataset of consecutive screening and diagnostic mammograms can discriminate between benign and malignant disease, and accurately predict the probability of BC	Retrospective, using prospectively collected and interpreted mammogram database	All mammograms performed at Medical College of Wisconsin Breast Care Centre (1999–2004)	Mammographic features (BI-RADS descriptors), routinely collected demographic risk factors	48,744 mammograms from 18,269 subjects	56.5	0.80%	Yes (consecutive mammographic exams)	Biopsy or ascertained by matching with State Cancer Reporting System

BC (breast cancer); CAD (computer-aided diagnosis); CNN (Convolutional Neural Network); DM (digital mammograms); DCNN (Deep Convolutional Neural Network); MIAS (Mammographic Image Analysis Society database); DDSM (Digital Database for Screening Mammography); NR (not reported); NS (not specified).

#Not reported however not likely to have included consecutive subjects or screens based on other information regarding the data source or the proportion with cancer.

Table 2. Summary of the findings of studies reporting on artificial intelligence (AI) in breast cancer detection.

First author	Category of AI	Model validation (specify type if reported)	Was there external validation using an independent dataset?	Area under the curve (AUC) %	Accuracy (other than AUC) %	Sensitivity%	Specificity%
Rodriguez-Ruiz [9]	Deep learning CNNs, feature classifiers, and image analysis algorithms	AI system trained and validated in earlier work	Present study represents a validation in an independent dataset	84.0%[vs average for 101 radiologists = 81.4%, statistically non-inferior based on difference in AUC of 2.6% (95%CI: −0.3 to 5.5)]*	NR	75%-85% across various datasets at radiologists’ specificity	NR
Al-Masni [10]	CNN, FC-NN	5-fold cross validation	No, however a modified (augmented image data) version used for testing	87.7%	85.5% (mass location 99.7%)	93.2%	78.0%
Ribli [11]	Faster R-CNN (based on a CNN with additional components)	NR	Participated in the DREAMǂ challenge which was based on an independent dataset (AUC 85%)	95%	NR	90%	NR
Chougrad [12]	Deep CNN	Stratified 5-fold cross validation	yes, however based on selected benign/cancer cases from the MIAS.	Range of estimates using various datasets: 96.0–99.0%	Range of estimates using various datasets: 95.5–98.2%	NR	NR
Bandeira-Diniz [13]	Deep CNN	Not done	No, however subset of 20% of database images (not used for training) saved for testing developed model	NR	Mass/non-mass region classification 91.0% in non-dense breasts, 94.8% in dense breasts.	Mass/non-mass region classification: 91.5% in non-dense, and 90.4% in dense breasts.	Mass/non-mass region classification 90.7% in non-dense, 96.4% in dense breasts
Becker [14]	ANN	Not done	Yes, tested using an independent image dataset	81–85% (training data) [vs radiologists 83–94%]*; 79%-81% (test data) [vs radiologists 77–87%]*	NR	59.8% (training data); 71.6%,73.7% (study 1, 2)	84.4% (training data); 69.6%, 72.0% (study 1, 2)
Lotter [15]	Multi-scale CNN	Cross-validation by patient (8% of data)	No, however subset of 5% of database used to validate the developed model	92%	NR	NR	NR
Becker [16]	Deep CNN	Validated	No, however subset of 30% of images (not used for training) saved for model validation	96% (training set); 84% (validation set) [vs experienced/intermediate human readers 89% (P < 0.05), or inexperienced readers 79% (P < 0.05)]*	NR	84.2% [experienced/intermediate readers 84.2%]	80.4% [experienced/intermediate readers 89/83%]
De Oliviera Silva [17]	Independent component analysis	NR	No	NR	92.7% in non-dense, 79.2% in dense breasts	95.7% in non-dense breasts, 66.7% in dense breasts	90.0% in non-dense breasts, 91.7% in dense breasts.
Kooi [18]	Deep CNN	8-fold cross validation	No, however subset of images (not used for training) used for testing developed model	CNN 92.9% (CAD system 91%); CNN with other variables 94.1%; test set: CNN 85.2% [vs radiologists 91.1% (p = 0.001 for mean AUC]*	_	NR	NR
Samala [19]	Deep CNN	4-fold cross-validation	Yes, independent test set (multi-task DCNN)	78% (single-task transfer learning), 82% (multi-task transfer learning), 76% (lesion-based), 79% (view-based)	NR	NR	NR
Carneiro [20]	Deep CNN	5-fold cross validation (Inbreast dataset)	No	90% (semi-automated approach), 86% (fully automated approach)	NR	69.0–94.0%	66.0–92.0%
Teare [21]	Dual deep CNN	Validated	No	92.2%	84%	91%	80.4%
Dhungel [22]	CNN, RF, BO	5-fold cross validation	No, however subset of images (not used for training) used for testing model	69–76% (minimal user intervention), 80–91% (manual set-up)	84–95% (range for various methods)	98%	70%
Sun [23]	Deep CNN	No	No	88.2% (using mixed labelled and unlabelled data)	82.4% (using mixed labelled and unlabelled data)	79.2–81.0% (for different data weighing functions)	69.4–72.3% (for different data weighing functions)
Saraswathi & Srinivasan [24]	FCRN	10-fold cross validation	No	67.9% (without cross-validation), 94.7% (with cross-validation)	_	90% (cross-validation)	91% (cross-validation)
Velikova [25]	BN	2-fold cross validation	No	62.8–75.5% (range for models learnt with discretised or continuous data)	NR	NR	NR
Dheeba & Tamil-Selvi [26]	DEOWNN	Not done	No	97.80%	NR	96.9%	92.9%
Dheeba & Tamil-Selvi [27]	SONN	Not done	Yes, 216 clinical mammogram images (54 patients) from screening centres	97.6% (for detecting microcalcification clusters); validation 91.3%	NR	NR	NR
Parmeggiani [28]	ANN	Not done	Yes – the ANN was pre-trained using external data and was validated in the study in a different population	NR	NR	80% [73.6–74.2% radiologists alone]*; 88.5% for AI and readers combined	70.9% [54–57.4% radiologists alone]*; 69.5% for AI and readers combined
Lesniak [29]	SVM, RF and CNN	10-fold cross validation.	No, however independent dataset used for normalisation of data	NR	NR	Mean true positive fraction 68.6 (higher than other CAD systems)	NR
Huang [30]	PSO, ANN, ANFIS, CBR with DT	10-fold cross validation	No, subset of images (not used for training) used for testing models	91.1% (PSO-based ANN), 92.8% (ANFIS), 83.6% (CBR-DT), 79.9% (CBR-LR)	NR	NR	NR
Ayer [31]	ANN	10-fold cross validation	No, subset of images from same dataset not used for training used for testing	96.5 [vs radiologists 93.9 (P < 0.001) aggregate level analysis] *	NR	90.7 [vs radiologists 82.2 (P < 0.001) aggregate level analysis]*	(sensitivity shown at a fixed specificity of 90%)

*Indicates results for radiologists in squared brackets.

ǂ DREAM refers to the Digital Mammography challenge (information via https://www.synapse.org/Digital_Mammography_DREAM_Challenge)

CNN (Convolutional Neural Network); SNN (Spiking Neural Network); BN (Bayesian Networks); FC-NN (Fourier Convolutional Neural Networks); ANN (Artificial Neural Network); DCNN (Deep Convolutional Neural Network); SONN (Swarm Optimization Neural Network); SVM (Support Vector Machine); DEOWNN (Differential Evolution Optimized Wavelet Neural Network); FCRN (Fully complex-valued relaxation neural network); PSO (Particle Swarm Optimizer); ANFIS (Adaptive Neuro-Fuzzy Inference System); MIAS (Mammographic Image Analysis Society database); CBR (Case-Based Reasoning); CAD (computer-aided diagnosis); BO (Bayesian optimisation); DT (Decision Trees); RF (Random Forests).

Appendix 1. Literature search and study identification strategy – Artificial Intelligence (AI) for breast cancer detection

Appendix 2. Database search terms

Database: Ovid MEDLINE(R) <1946 to October Week 4 2018> 1 exp Breast Neoplasms/ (268387) 2 limit 1 to (English language and humans and yr = “2010 -Current”) (90643) 3 artificial intelligence.m_titl. (659) 4 machine.m_titl. (8965) 5 neural.m_titl. (61500) 6 deep learning.m_titl. (366) 7 3 or 4 or 5 or 6 (71321) 8 2 and 7 (172)

Rodriguez-Ruiz A, Lång K, Gubern-Merida A, et al. Stand-alone artificial intelligence for breast cancer detection in mammography: comparison with 101 radiologists. J Natl Cancer Inst. 2019;111(9):djy222.

 Google Scholar

Al-Masni MA, Al-Antari MA, Park J-M, et al. Simultaneous detection and classification of breast masses in digital mammograms via a deep learning YOLO-based CAD system. Comput Methods Programs Biomed. 2018;157:85–94.

 PubMed Web of Science ®Google Scholar

Ribli D, Horváth A, Unger Z, et al. Detecting and classifying lesions in mammograms with deep learning. Sci Rep. 2018;8(1):4165.

 PubMed Web of Science ®Google Scholar

Chougrad H, Zouaki H, Alheyane O. Deep convolutional neural networks for breast cancer screening. Comput Methods Programs Biomed. 2018 Apr 01;157: 19–30.

 PubMed Web of Science ®Google Scholar

Bandeira Diniz JO, Bandeira Diniz PH, Azevedo Valente TL, et al. Detection of mass regions in mammograms by bilateral analysis adapted to breast density using similarity indexes and convolutional neural networks. Comput Methods Programs Biomed. 2018;156:191–207.

 PubMed Web of Science ®Google Scholar

Becker AS, Mueller M, Stoffel E, et al. Classification of breast cancer in ultrasound imaging using a generic deep learning analysis software: a pilot study. Br J Radiol. 2018;91(1083):20170576. PubMed PMID: 29215311.

 PubMed Web of Science ®Google Scholar

Lotter W, Sorensen G, Cox D. A multi-scale CNN and curriculum learning strategy for mammogram classification. In: Deep learning in medical image analysis and multimodal learning for clinical decision support. Springer; 2017. p. 169–177.

 Google Scholar

Becker AS, Marcon M, Ghafoor S, et al. Deep learning in mammography: diagnostic accuracy of a multipurpose image analysis software in the detection of breast cancer. Invest Radiol. 2017;52(7):434–440. PubMed PMID: 00004424-201707000-00007.

 PubMed Web of Science ®Google Scholar

de Oliveira Silva LC, Barros AK, Lopes MV. Detecting masses in dense breast using independent component analysis. Artif Intell Med. 2017;80:29–38.

 PubMed Web of Science ®Google Scholar

Kooi T, Litjens G, van Ginneken B, et al. Large scale deep learning for computer aided detection of mammographic lesions. Med Image Anal. 2017;35:303–312.

 PubMed Web of Science ®Google Scholar

Samala RK, Chan H-P, Hadjiiski LM, et al. Multi-task transfer learning deep convolutional neural network: application to computer-aided diagnosis of breast cancer on mammograms. Phys Med Biol. 2017;62(23):8894–8908. PubMed PMID: 29035873.

 PubMed Web of Science ®Google Scholar

Carneiro G, Nascimento J, Bradley AP. Automated analysis of unregistered multi-view mammograms with deep learning. IEEE Trans Med Imaging. 2017;36(11):2355–2365.

 PubMed Web of Science ®Google Scholar

Teare P, Fishman M, Benzaquen O, et al. Malignancy detection on mammography using dual deep convolutional neural networks and genetically discovered false color input enhancement. J Digit Imaging. 2017;30(4):499–505. PubMed PMID: 28656455.

 PubMed Web of Science ®Google Scholar

Dhungel N, Carneiro G, Bradley AP. A deep learning approach for the analysis of masses in mammograms with minimal user intervention. Med Image Anal. 2017 Apr 01;37: 114–128.

 PubMed Web of Science ®Google Scholar

Sun W, Tseng T-L, Zhang J, et al. Enhancing deep convolutional neural network scheme for breast cancer diagnosis with unlabeled data. Computerized Med Imaging Graphics. 2017;57:4–9.

 PubMed Web of Science ®Google Scholar

Saraswathi D, Srinivasan E. A CAD system to analyse mammogram images using fully complex-valued relaxation neural network ensembled classifier. J Med Eng Technol. 2014 Oct 01;38(7):359–366.

 PubMedGoogle Scholar

Velikova M, Lucas PJF, Samulski M, et al. On the interplay of machine learning and background knowledge in image interpretation by Bayesian networks. Artif Intell Med. 2013 Jan 01;57(1):73–86.

 PubMed Web of Science ®Google Scholar

Dheeba J, Tamil Selvi S. An improved decision support system for detection of lesions in mammograms using differential evolution optimized wavelet neural network [journal article]. J Med Syst. 2012 Oct 01;36(5):3223–3232.

 PubMed Web of Science ®Google Scholar

Dheeba J, Selvi ST. A swarm optimized neural network system for classification of microcalcification in mammograms [journal article]. J Med Syst. 2012 Oct 01;36(5):3051–3061.

 PubMed Web of Science ®Google Scholar

Parmeggiani D, Avenia N, Sanguinetti A, et al. Artificial intelligence against breast cancer (A.N.N.E.S-B.C.-Project). Ann Ital Chir. 2012 Jan-Feb;83(1):1–5. PubMed PMID: 22352208; eng.

 PubMed Web of Science ®Google Scholar

Lesniak JM, Hupse R, Blanc R, et al. Comparative evaluation of support vector machine classification for computer aided detection of breast masses in mammography. Phys Med Biol. 2012;57(16):5295.

 PubMed Web of Science ®Google Scholar

Huang M-L, Hung Y-H, Lee W-M, et al. Usage of case-based reasoning, neural network and adaptive neuro-fuzzy inference system classification techniques in breast cancer dataset classification diagnosis [journal article]. J Med Syst. 2012 Apr 01;36(2):407–414.

 PubMed Web of Science ®Google Scholar

Ayer T, Alagoz O, Chhatwal J, et al. Breast cancer risk estimation with artificial neural networks revisited: discrimination and calibration. Cancer . 2010;116(14):3310–3321. PubMed PMID: 20564067.

 PubMed Web of Science ®Google Scholar