iCAD Deep Learning AI for Breast Cancer Detection and Risk
Peer-Reviewed Journal Articles and Scientific Meeting Presentations

Deep Learning AI for Breast Cancer Detection Peer-Reviewed Journal Articles

1. Graewingholt A, Rossi PG. Retrospective analysis of the effect on interval cancer rate of adding an artificial intelligence algorithm to the reading process for two-dimensional full-field digital mammography. J Med Screen. 2021 Jan 12:969141320988049. doi: 10.1177/0969141320988049. Epub ahead of print. PMID: 33435812.
   • Sensitivity of ProFound AI for lesions judged to be present in screening mammograms was 93% and sensitivity for interval cancers was 48%
   • Using ProFound AI as a reading tool has the potential to reduce interval cancers
2. Graewingholt A, Duffy S. Retrospective comparison between single reading plus an artificial intelligence algorithm and two-view digital tomosynthesis with double reading in breast screening. J Med Screen. 2021 Jan 5:969141320984198. doi: 10.1177/0969141320984198. Epub ahead of print. PMID: 33402033.
   • Sensitivity of a single radiologist reading digital breast tomosynthesis with ProFound AI is non-inferior to that of standard of care using 2D mammography with double reading
3. Conant EF, Toledano AY, Periaswamy S, Fotin SV, Go J, Boatsman JE, Hoffmeister JW. Improving accuracy and efficiency with concurrent use of artificial intelligence for digital breast tomosynthesis. Radiol Artif Intell. 2019 Jul 31;1(4):e180096. doi: 10.1148/ryai.2019180096. PMID: 32076660; PMCID: PMC6677281.
   • Reading times were significantly reduced (52.7%), and sensitivity, specificity, recall rate, and area under the receiver operating characteristic curve (AUC) improved (8.0%, 6.9%, 7.2%, 0.057, respectively) in a nonclinical reader study when ProFound AI was utilized concurrently with image interpretation for digital breast tomosynthesis
   • The results of this study suggest that both improved efficiency and accuracy could be achieved in clinical practice by using ProFound AI
4. James JJ, Giannotti E, Chen Y. Evaluation of a Computer-Aided Detection (CAD)-enhanced 2D synthetic mammogram: comparison with standard synthetic 2D mammograms and conventional 2D digital mammography. Clin Radiol. 2018 Oct;73(10):886-892. doi: 10.1016/j.crad.2018.05.028. Epub 2018 Jun 30. PMID: 29970247.
   • CAD-enhancement by PowerLook Tomo Detection blending information from the digital breast tomosynthesis data set onto a synthetic 2D image significantly improves performance of synthetic 2D mammograms and also exhibits improved diagnostic accuracy compared to conventional 2D digital mammography
5. Benedikt RA, Boatsman JE, Swann CA, Kirkpatrick AD, Toledano AY. Concurrent computer-aided detection improves reading time of digital breast tomosynthesis and maintains interpretation performance in a multireader multicase study. AJR Am J Roentgenol. 2018 Mar;210(3):685-694. doi: 10.2214/AJR.17.18185. Epub 2017 Oct 24. PMID: 29064756.
   • A pivotal reader study with 6 radiologists and 80 digital breast tomosynthesis studies demonstrated that concurrent reading with CAD-enhancement by PowerLook Tomo Detection resulted in 23.5% faster reading time with non-inferiority of radiologist interpretation performance
6. Balleyguier C, Arfi-Rouche J, Levy L, Toubiana PR, Cohen-Scali F, Toledano AY, Boyer B. Improving digital breast tomosynthesis reading time: a pilot multi-reader, multi-case study using concurrent Computer-Aided Detection (CAD). Eur J Radiol. 2017 Dec;97:83-89. doi: 10.1016/j.ejrad.2017.10.014. Epub 2017 Oct 24. PMID: 29153373.
   • A pilot reader study with 20 radiologists and 240 digital breast tomosynthesis studies demonstrated that concurrent reading with CAD-enhancement by PowerLook Tomo Detection resulted in 29.2% faster reading time, while maintaining reader interpretation performance

Deep Learning AI for Breast Cancer Detection Scientific Meeting Presentations

1. RSNA 2021: Ice MF, Periaswamy S, Shisler J, Hoffmeister JW, Nguyen M, Sardina AJ, Rose SL. Use of artificial intelligence software to Identify clinically aggressive lesions on prior digital breast tomosynthesis exams based on molecular subtypes.
   • In 32.7% of 505 cancer cases, AI identified visible and actionable lesions on the prior exams, indicating that breast cancer could have been detected 12 to 18 months earlier with the use of AI.
   • An AI system is correctly able to identify 77.8% luminal B and 62.5% triple negative cancers 12-18 months prior to clinical diagnosis.  This provides an opportunity to detect more aggressive cancers earlier, potentially leading to better outcomes.
2. RSNA 2021: Letter HP, Toledano AY, Hoffmeister JW, Nishikawa RM, Periaswamy S, Sharma S, Conant EF. Effect of artificial intelligence software on digital breast tomosynthesis screening cancer detection and abnormal interpretation rates among subspecialized breast radiologists: a real world experience.
   • While we are unable to show statistically significant changes in CDR and AIR outcomes in the two groups given the low numbers of total cancers and relatively short time period, the results are consistent with prior reader studies. There are modest, yet noticeable, improvements in CDR with AI, without significant increases in AIR with AI. We expect to obtain more accurate estimates of the impact of DBT-AI as we accrue more patients.
   • To our knowledge, our study is the first to assess the real-world application of a DBT-based AI algorithm to assist radiologists in the detection of breast cancer in screening exams.
3. RSNA 2021: Conant EF, Periaswamy S, Elahi A, Hoffmeister JW, Gastounioti A. Tomosynthesis-driven AI-derived breast cancer scoring among a racially diverse cohort of women undergoing screening: potential for triage and precision cancer screening.
   • Preliminary data across a racially diverse screening population suggests there may be differences in the distribution of case scores and performance of a DBT-AI algorithm based on race.
   • Artificial intelligence algorithms for CAD-driven triaging and precision breast cancer screening paradigms may require incorporation of racial data for further refinement of prediction models.
4. SBI 2021: Conant EF, Toledano AY, Periaswamy S, Hoffmeister JW, Nishikawa RM. Feasibility of automated identification of low-likelihood of cancer in digital breast tomosynthesis screening exams.
   • ProFound AI triaging cases at a 0% false negative rate identifies 33.4% of digital breast tomosynthesis screening exams
   • Combining ProFound AI case scores with age and breast density, the triaged percent could potentially increase to 58.6%.
5. NCBC 2021: Traill M, McGoff T, Neal C, Jacobs C, Corser W, Hoffmeister J. Correlation between BI-RADS Assessment Categories and Artificial Intelligence Case Scores.
   • Results from 890 consecutive screening DBT exams showed a strong positive correlation of ProFound AI case score of <60% and patients assessed as BI-RADS 1 or 2
   • A ProFound AI case score >60% is an indicator of increased chance of malignancy on screening DBT
6. ECR 2021: Conant EF, Toledano AY, Periaswamy S, Hoffmeister JW, Nishikawa RM. Use of an artificial intelligence software to identify low-likelihood of cancer exams collected across four commercial digital tomosynthesis screening systems. [RPS 605-3]
   • ProFound AI triaging cases at a 0% false negative rate identifies 33.4% of digital breast tomosynthesis screening exams
   • Combining ProFound AI case scores with age and breast density, the triaged percent could potentially increase to 58.6%.
7. SIIM 2020: Periaswamy S, Conant EF, Toledano AY, Go J, Boatsman JE, Hoffmeister JW. Breast cancer detection with stand-alone artificial intelligence compared to radiologists in digital breast tomosynthesis. [Course ID: 4012]
   • Stand-alone ProFound AI has higher sensitivity and lower specificity than the average of radiologists without ProFound AI, with a slightly higher AUC for ProFound AI that is not statistically significant
   • Understanding ProFound AI performance may help radiologists more efficiently and effectively use ProFound AI for detecting breast cancer with digital breast tomosynthesis
8. ECR 2020: Conant EF, Toledano AY, Periaswamy S, Fotin SV, Haldankar H, Go J, Boatsman J, Hoffmeister J. Breast cancer detection by mammographic view with artificial intelligence in digital breast tomosynthesis. [RPS 605b-6]
   • The estimated sensitivity of ProFound AI for craniocaudal (CC) and mediolateral oblique (MLO) views were 0.74 and 0.77, respectively, while the estimated specificity with CC images was 0.64 which is significantly higher than that of MLO images at 0.54
   • Overall, ProFound AI detected 91% (59/65) of cancers and ruled out 41% (79/195) non-cancer cases
   • While ProFound AI was not trained to maximize performance within mammographic view, these results suggest that unbalanced conspicuity across mammographic views may have implications for ProFound AI performance with single view exams
9. ECR 2020: Heywang-Köbrunner SH, Jänsch A, Mieskes C, Hertlein M, Hacker A. The value of 2D-AI-based CAD for second or third reading tested on 17,910 screening mammograms. [RPS 702-4]
   • Using a case score threshold of 30, ProFound AI for digital mammography achieved a sensitivity of 91.5% and specificity of 80.2%, while two radiologists participating in double reading screening had lower sensitivities and higher specificities: 84.6% and 91.6% for reader 1; 89.7% and 91.5% for reader 2
   • Achieved results justify hopes to use ProFound AI for digital mammography for second reading (e.g. in countries with a shortage of readers) or for 3rd reading in the near future
10. RSNA 2019: Conant EF, Toledano AY, Periaswamy S, Fotin SV, Go J, Pike J, Boatsman JE, Hoffmeister JW. Improved breast cancer detection and reading time with concurrent use of deep learning-based artificial intelligence for digital breast tomosynthesis when interpreted with digital mammography versus synthetic mammography. [SSA01-07]
    • The use of ProFound AI with digital breast tomosynthesis improved AUC, sensitivity, specificity and reading time when reading digital breast tomosynthesis with digital 2D or with synthetic 2D
11. RSNA 2019: Conant EF, Toledano AY, Periaswamy S, Fotin SV, Go J, Pike J, Boatsman JE, Hoffmeister JW. How artificial intelligence may help improve accuracy and reading times in the interpretation of digital breast tomosynthesis screening studies. [BR204-ED-TUA9]
    • This exhibit demonstrates concurrent use of ProFound AI for digital breast tomosynthesis that detects soft tissue and calcific lesions in digital breast tomosynthesis slices and provides lesion outlines and calibrated confidence scores at the lesion-level and case-level
    • Example cases are from a reader study with 24 radiologists each reading 65 cancer and 195 non-cancer cases both with and without ProFound AI showing significant improvements on average in AUC, sensitivity, specificity, recall rate and reading time
12. SIIM 2019: Conant EF, Toledano AY, Periaswamy S, Fotin SV, Go J, Pike J, Boatsman JE, Hoffmeister JW. Improved breast cancer detection and reading time with digital breast tomosynthesis for breast subspecialists and general radiologists with concurrent use of artificial intelligence. [Machine Learning: Other]
    • Concurrent use of ProFound AI with digital breast tomosynthesis improves AUC, reading time, case-level sensitivity and specificity for both breast subspecialists and general radiologists
    • Improvements in breast cancer detection and reading time with concurrent use of ProFound AI with digital breast tomosynthesis apply to both breast subspecialists and general radiologists
13. SBI 2019: Conant EF, Toledano AY, Periaswamy S, Fotin SV, Go J, Pike J, Boatsman JE, Hoffmeister JW. Concurrent use of deep learning based artificial intelligence improves detection of breast cancer and reading time with digital breast tomosynthesis in women with dense and non-dense breasts. [Abstract ID: 582552]
    • The use of ProFound AI with digital breast tomosynthesis improved breast cancer detection, AUC, sensitivity and specificity and shortened reading time in dense breasts and non-dense breasts
    • ProFound AI with digital breast tomosynthesis benefits all women, those with dense and with non-dense breasts
14. ECR 2019: Conant EF, Periaswamy S, Fotin S, Go J, Pike J, Boatsman J, Hoffmeister J. Impact of breast cancer characteristics on reader performance with concurrent use of artificial intelligence with digital breast tomosynthesis. [C-1648]
    • Sensitivity and reading time improved for all lesion types and histopathologies when ProFound AI was used concurrently with digital breast tomosynthesis
    • Specificity improved for mammographic soft tissue cases and cases without any suspicious lesions, but not for calcifications-only cases
15. ECR 2019: Conant EF, Toledano AY, Periaswamy S, Fotin S, Go J, Pike J, Boatsman J, Hoffmeister J. Case examples to demonstrate positive and negative impacts of a deep learning based concurrent artificial intelligence system for digital breast tomosynthesis. [C-2151]
    • The case examples provided may help radiologists more effectively use ProFound AI incorporated with digital breast tomosynthesis
    • ProFound AI has high standalone performance with 93% sensitivity for all case types and 44-67% specificity
    • Since ProFound AI misses 7% of cancers, it’s important not to overly rely on ProFound AI when you detect a suspicious lesion that is not marked by ProFound AI
    • Since ProFound AI marks 33-56% of non-cancers, it’s important not to overly rely on ProFound AI when it detects a region that you may not consider suspicious
16. RSNA 2018: Conant EF, Toledano AY, Periaswamy S, Fotin SV, Go J, Hoffmeister JW, Boatsman JE. Improved accuracy and efficiency with concurrent use of artificial intelligence for digital breast tomosynthesis screening. [RC215-14]
    • Concurrent use of ProFound AI with digital breast tomosynthesis improves cancer detection with increases of 0.057 in AUC, 8.0% in sensitivity, 6.9% in specificity; and decreases of 7.2% in recall rate and 52.7% in reading time
    • Radiologist’s concurrent use of ProFound AI for digital breast tomosynthesis with certainty of finding scores increases detection of breast cancer with significant reduction in reading time while improving sensitivity and specificity
17. ECR 2018: James J, Giannotti E, Chen Y. Evaluation of a CAD-enhanced 2D synthetic mammogram: comparison with a standard synthetic mammogram & FFDM. [B-1636]
    • CAD enhancement with PowerLook Tomo Detection for digital breast tomosynthesis offers an additional approach to improve the performance of synthetic 2D mammography
    • Blending information from the digital breast tomosynthesis data set onto a synthetic image with PowerLook Tomo Detection can potentially produce an image which is superior to 2D digital mammography
18. ECR 2017: Benedikt RA, Toledano AY, Boatsman J, Arfi Rouche J, Boyer B, Hoffmeister J, Balleyguier C. Concurrent CAD with digital breast tomosynthesis improves reading time and maintains performance for dedicated breast radiologists and general radiologists. [C-1177]
    • The impact of concurrent use of PowerLook Tomo Detection for digital breast tomosynthesis by breast radiologists and general radiologists was similar
    • Average reading time improved about 30% while maintaining sensitivity, specificity, and AUC with concurrent use of PowerLook Tomo Detection for digital breast tomosynthesis
19. RSNA 2016: Benedikt RA, Swann CA, Kirkpatrick AD, Toledano A, Periaswamy S, Boatsman JE, Go J, Hoffmeister JW. Concurrent CAD for digital breast tomosynthesis. [SSE02-01]
    • Concurrent use of PowerLook Tomo Detection for digital breast tomosynthesis results in a 29.2% faster reading time with non-inferiority of radiologist performance compared to reading without PowerLook Tomo Detection
    • Concurrent use of PowerLook Tomo Detection maintains high performance of digital breast tomosynthesis with a significant reduction in reading time
20. RSNA 2016: Balleyguier C, Arfi-Rouche J, Levy L, Toubiana PR, Cohen-Scali F, Toledano A, Periaswamy S, Go J, Hoffmeister JW, Boyer B. Pilot reader study of concurrent CAD for digital breast tomosynthesis. [BR231-SD-MOA6]
    • Concurrent use of PowerLook Tomo Detection for digital breast tomosynthesis results in a 23.5% faster reading time with non-inferiority of radiologist performance compared to reading without PowerLook Tomo Detection
    • A pivotal reader study with 20 readers, 60 cancers, and 180 non-cancers is planned to more robustly evaluate these endpoints
    • Concurrent use of PowerLook Tomo Detection maintains high performance of digital breast tomosynthesis with a significant reduction in reading time thus improving workflow even for very experienced radiologists
21. EUSOBI 2016: Arfi-Rouche J, Balleyguier C, Levy L, Toubiana P, Cohen-Scali F, Boyer B, Benedikt R, Boatsman J, Kirkpatrick A, Swann C, Toledano A, Periaswamy S, Go J, Hoffmeister J. Determining efficacy of concurrent CAD for digital breast tomosynthesis. [Poster 19]
    • PowerLook Tomo Detection for digital breast tomosynthesis has greater than 90% sensitivity in detecting malignant soft tissue densities with more than 40% of malignant soft tissue densities are visible on synthetic images enhanced with PowerLook Tomo Detection versus standard synthetic images
    • Concurrent use of PowerLook Tomo Detection maintains performance of digital breast tomosynthesis with significant 29.2% reduction in reading time
22. JFR 2015: Arfi-Rouche J, Boyer B, Levy L, Toubiana P, Cohen-Scali F, Toledano A, Periaswamy S, Go J, Hoffmeister J, Balleyguier C. Evaluation of the contribution of a diagnostic support system (CAD) in tomosynthesis. [11:10 Oct 16]
    • The use of PowerLook Tomo Detection for digital breast tomosynthesis saves 23.5% on time interpreting tomosynthesis exams without altering the performance of the radiologist
    • This technique appears promising for future application of digital breast tomosynthesis screening

Deep Learning AI for Breast Cancer Risk Peer-Reviewed Journal Articles

1. Eriksson M, Czene K, Strand F, Zackrisson S, Lindholm P, Lång K, Förnvik D, Sartor H, Mavaddat N, Easton D, Hall P. Identification of women at high risk of breast cancer who need supplemental screening. Radiology. 2020 Nov;297(2):327-333. doi: 10.1148/radiol.2020201620. Epub 2020 Sep 8. PMID: 32897160.
   • The ProFound AI short-term risk model based on three mammographic features, with optional lifestyle factors and a polygenic risk score, identified women at high risk of breast cancer who need supplemental screening
   • Women identified as high-risk by ProFound AI Risk were more likely to be diagnosed with stage II cancers and with tumors 20 mm or larger and were less likely to have stage I and estrogen receptor–positive tumors
   • ProFound AI Risk reached an AUC of 0.73 for predicting breast cancer, while traditional risk models showed lower AUCs: 0.62 for Tyrer-Cuzick with mammographic density and 0.61 for Gail with mammographic density

Deep Learning AI for Breast Cancer Risk Scientific Meeting Presentations

1. SABCS 2020: Eriksson M, Czene K, Hall P. Identification of women at high risk of breast cancer and in need of supplementary screening – A cohort study. [RPS 602b-1]
   • By combining mammographic features, their left-right asymmetries, and optionally lifestyle factors, family history, and a polygenic risk score we generated a model that identifies women at high likelihood of being diagnosed with breast cancer within two year of a negative screen and in possible need of supplemental screening or preventative intervention
   • The image-based ProFound AI Risk model reached an AUC of 0.73 (95% CI 0.71,0.74)
   • The lifestyle and genetic extended model AUCs were 0.74 (95% CI 0.72,0,75) and 0.77 (95% CI 0.75,0.79) respectively
2. ECR 2020: Eriksson M, Czene K, Zackrisson A, Hall P. Identification of women at high risk of breast cancer and in need of supplementary screening. [PS8-01]
   • By combining mammographic features, lifestyle factors, family history, and a polygenic risk score, we generated a model that identifies women with a high likelihood of being diagnosed with breast cancer within two years and in need of supplemental screening
   • The full model reached an area under the curve of 0.77 (95% CI 0.76, 0.79) with good model fit (Hosmer-Lemeshow = 0.2)
   • There was an 8-fold difference in risk between the 8% of women at high risk and at general risk
   • Women identified as high risk were more likely to be diagnosed with more aggressive cancers
   • The image-based ProFound AI Risk model was validated in two independent cohorts