Clinical Evidence

MammoScreen provides radiologists with reliable and actionable assessments of mammograms to help them improve their performance and accuracy. Some users find very distinct advantages to using AI with reading. When using MammoScreen, it was found that 100% of users increased performance1. 

Clinical Relevance

MammoScreen has been trained on more cancer images than any radiologist will see in their entire lifetime. The software has been provided with cases from the United States and Europe to ensure that it has access to a diversity of images and doesn’t become biased. MammoScreen detects, characterizes, and classifies suspicious architectural distortions, calcification, masses, asymmetries, and focal asymmetries.

Each exam is provided with a unified score (1-10) for each finding, breast, and overall case. These numbers are traceable, actionable, and easy-to-understand. This scoring system helps radiologists know what they are looking at as they open cases. Our scoring system (link to white paper) has been described as having a fellow by your side as you do your review. 

When assisted with MammoScreen, radiologists increased by eight percentage points in sensitivity, and a two percentage point increase in specificity. We also found that about 84% of users lower their false negative rate, and more than 80% of readers decrease their false positive rate1. 

Evidence

MammoScreen has advanced significantly since becoming in 2017 the winner of the DREAM Challenge2. Our unique use of 2D, 3D, and prior images, enables users to benefit from findings in the combination of images, whereas using only a 2D or 3D does not provide the same quality of interpretation.  

You may wonder why we provide the algorithm with three sets of images. After testing, it was found that providing 2D and 3D images helps MammoScreen detect cancer more effectively than when using just one of the images3. In addition, adding priors reduced the false positive rate, and we also found that it doubled the number of true positive cases at the same specificity4. 

CAD, and some AI are perceived as extra steps in the workflow that can increase reading time, with MammoScreen, the results are delivered concurrently, and it can also assist with reporting. In a study we found reading/reporting time decreased by 24% in average1! 

1. S. Pacilè, P. Germaine, C. Sclafert, T. Bertinotti, P. Fillard, and S. Singla Long, “Evaluation of a Multi-Instant Multimodal Artificial Intelligence System Supporting Interpretive and Noninterpretive Functions,” J. Breast Imaging, doi: 10.1093/jbi/wbae062
2. T. Schaffter et al., “Evaluation of Combined Artificial Intelligence and Radiologist Assessment to Interpret Screening Mammograms,” JAMA Netw. Open, vol. 3, no. 3, pp. e200265–e200265, Mar. 2020, doi: 10.1001/jamanetworkopen.2020.0265
3. S. Pacilè , C.Aguilar, A.Iannessi , P.Fillard (2023, May 4-7). “APPLICATION OF ARTIFICIAL INTELLIGENCE TO MAMMOGRAPHY TOMOSYNTHESIS COMBINED IMAGES FOR BREAST CANCER SCREENING” [conference presentation]. SBI 2023, National Harbor, MD
4. S. Pacilè, C. Aguilar, S. Chambon, and P. Fillard, “Including temporal changes information to an AI system for breast cancer detection to reduce false positive rate,” in 16th International Workshop on Breast Imaging (IWBI2022), SPIE, Jul. 2022, pp. 153–160. doi: 10.1117/12.2624098.

A novel triaging model for screening mammograms based on a density AI and a cancer detection AI

Simulation triage model based on MammoScreen density and cancer detection assessments.

• The proposed two-step triage strategy could safely select patients to be
  removed from radiologists’ workload without affecting the cancer detection rate.
• Using two AI systems (one for density and one for cancer detection) to triage
  mammograms could potentially:
  • Reduce radiologist workload at constant cancer detection rate and
  • Decrease the rate of women recalled back for further examinations.

When incorporating AI into your workflow, this study finds that there is an opportunity to safely reduce the radiologists workload while not impacting the cancer detection rate. It also can help decrease the rate of women recalled for further exams.

“A novel triaging model for screening mammograms based on a density AI and a cancer detection AI,” (2023, Nov. 26-30). [poster presentation] RSNA 2023, Chicago, IL

Artificial Intelligence for Digital Breast Tomosynthesis: A Tool to Enhance Radiologist’s Performance and Efficiency.

MammoScreen for Digital Breast Tomosynthesis: Boosting Radiologists’ Accuracy and Efficiency.

• MS can save up to 35% of reading time and up to 50% on low suspicion cases;
• Up to 50% of missed cancers by primary reader caught

Reducing reading time is one of the many benefits of using an AI, in this case, MammoScreen showed that there was up to 35% savings in reading time. Additionally, MammoScreen helped to find up to 50% of missed cancers.

“Artificial Intelligence for Digital Breast Tomosynthesis: A Tool to Enhance Radiologist’s Performance and Efficiency.” (2021, Nov. 28 – Dec. 2). [poster presentation]. RSNA 2021, Chicago, IL

Monitoring Methodology for an AI Tool for Breast Cancer Screening Deployed in Clinical Centers

Monitoring the MammoScreen Score distribution in a clinical setting.

In order to ensure accurate numbers, we monitored MammoScreen’s performance in a clnical setting. This study examines the score distubution in a clinical setting.

C. Aguilar, S. Pacilè, N. Weber, and P. Fillard, “Monitoring Methodology for an AI Tool for Breast Cancer Screening Deployed in Clinical Centers,” Life, vol. 13, no. 2, p. 440, Feb. 2023, doi: 10.3390/life13020440

Application of Artificial Intelligence to Mammography-Tomosynthesis Combined Images for Breast Cancer Screening.

Using MammoScreen on 2D and tomosynthesis images combined

• MammoScreen combining 2D and 3D analysis together is better in breast cancer
  detection compared to MammoScreen 3D, itself better than MammoScreen 2D.

Two heads are better than one, just like two images work better than one. This study discovered that MammoScreen’s performance improves when the analysis was provided with both 2D and 3D images simultaneously.

“APPLICATION OF ARTIFICIAL INTELLIGENCE TO MAMMOGRAPHY TOMOSYNTHESIS COMBINED IMAGES FOR BREAST CANCER SCREENING” (2023, May 4-7). [conference presentation]. SBI 2023, National Harbor, MD

Evaluation of a Multi-Instant Multimodal Artificial Intelligence System Supporting Interpretive and Noninterpretive Functions.

An evaluation of MammoScreen from Screening interpretation to Reporting.

• MammoScreen significantly improves radiologists’ performance when reviewing screening 2D+ 3D mammograms with prior: +8% in sensitivity and +2% in specificity

When using MammoScreen’s Combo Mode (2D+3D mammograms with prior results in an +8% increase in sensitivity, and +2% increase in specificity. This study also finds that MammoScreen helps reduce reading time by up to 24%.

S. Pacilè, P. Germaine, C. Sclafert, T. Bertinotti, P. Fillard, and S. Singla Long, “Evaluation of a Multi-Instant Multimodal Artificial Intelligence System Supporting Interpretive and Noninterpretive Functions,” J. Breast Imaging, doi: 10.1093/jbi/wbae062

Reducing False-Positive Recalls by Adding Temporal Changes Information to an AI System for Breast Cancer Detection.

Reducing False Positives in Breast Cancer Detection by Adding Priors

• Addition of prior reduces FP rate.
• Doubled number of cancer cases
  detected without any FP.
• About 20% of cancer found without
  creating FP

The addition of priors to MammoScreen reduces the false positive rate. Most impressively, the number of cancer cases doubled without any false positive.

S. Pacilè, C. Aguilar, S. Chambon, and P. Fillard, “Including temporal changes information to an AI system for breast cancer detection to reduce false positive rate,” in 16th International Workshop on Breast Imaging (IWBI2022), SPIE, Jul. 2022, pp. 153–160. doi: 10.1117/12.2624098.

Improving Breast Cancer Detection Accuracy of Mammography with the Current Use of an Artificial Intelligence Tool.

Increasing radiologists accuracy when using MammoScreen

• About 80% of readers lowered their FN rate
• > 50% half of the readers decreased their FP rate

There are a lot of numbers that radiologists concern themselves with – CDR, recall rate, etc. In this study MammoScreen helped lower 80% of readers false negative rate, and over half of readers decreased their false positive rate.

S. Pacilè, J. Lopez, P. Chone, T. Bertinotti, J. M. Grouin, and P. Fillard, “Improving Breast Cancer Detection Accuracy of Mammography with the Concurrent Use of an Artificial Intelligence Tool,” Radiol. Artif. Intell., vol. 2, no. 6, p. e190208, Nov. 2020, doi: 10.1148/ryai.2020190208.

Exploring the Ability of an AI System to Differentiate Breast Microcalcifications.

Understanding how MammoScreen differentiates microcalcifications

Within this study, we examined how MammoScreen performed on microcalcifications. The result was that using MammoScreen for microcalcifications would help biopsy requirements to be lowered.

“Exploring the Ability of an AI System to Differentiate Breast Microcalcifications.” (2021, Apr. 9-11). [poster presentation]. Virtual

Breast Screening and Artificial Intelligence: An Independent Evaluation of Two Different Software Caried Out at Valenciennes Hospital.

Comparing two separate mammography AI’s effectiveness

In this paper, you can read about the comparison between MammoScreen and a former competitor. This was an independent study performed in France.

A. L. Vourch, P. Edouard, and N. Laurent, “Breast screening and artificial intelligence: an independent evaluation of two different software carried out at Valenciennes hospital,” in 15th International Workshop on Breast Imaging (IWBI2020), International Society for Optics and Photonics, May 2020, p. 1151321. doi: 10.1117/12.2564129.

Impact of Artificial Intelligence in Breast Cancer Screening with Mammography.

MammoScreen’s impact on radiologists BI-RADS assignment

• Reduced number of “obvious”
• FP by 34% (breasts categorized as BI-RADS 4-5 instead of BI-RADS 1-2)
  and overall FP by 15%.
• Improved breast cancer detection
• Performance without increasing reading time

When using AI, the goal is to reduce false positives and increase the cancer detection. This study found that false positives are reduced by 15% overall, an improved cancer detection, and no increase in reading time.

L.-A. Dang et al., “Impact of artificial intelligence in breast cancer screening with mammography,” Breast Cancer, Jun. 2022, doi: 10.1007/s12282-022-01375-9.

Evaluation of Combined Artificial Intelligence and Radiologist Assessment to Interpret Screening Mammograms.

DREAM Challenge: Assessing the Use of AI and Radiologists Together to Analyze Screening Mammograms.

In a competition with over 100 other breast algorithms, MammoScreen came out on top. This was during the DREAM Challenge, this paper covers the process and challenge.

T. Schaffter et al., “Evaluation of Combined Artificial Intelligence and Radiologist Assessment to Interpret Screening Mammograms,” JAMA Netw. Open, vol. 3, no. 3, pp. e200265–e200265, Mar. 2020, doi: 10.1001/jamanetworkopen.2020.0265

Potential Benefits of an AI System in the Early Detection of Breast Cancer.

How MammoScreen can help detect breast cancer earlier

• With MS, radiologists can detect breast cancer 1 or 2 screening cycles earlier.
• MS can flag up to 27% of cancer cases 1 year earlier and up to 21%
  2 years earlier.

In this study, radioloigsits benefited from MammoScreen by finding cancer 2 years earlier 21% of the time, and 27% of the time 1 year earlier.

“Potential Benefits of an AI System in the Early Detection of Breast Cancer.” (2021, Apr. 9-11). [poster presentation]. Virtual

A novel approach for the evaluation of artificial intelligence on consecutive screening mammograms.

A method to evaluate MammoScreen on large dataset

• MammoScreen can detect visible cancer:
  • 42% one year before
  • 38.5% two years before

The importance of early detection cannot be understated. When using MammoScreen, 42% of cancer was found one year prior, and 38.5% was found two years prior.

“A novel approach for the evaluation of artificial intelligence on consecutive screening mammograms.” (2022, Nov. 27 – Dec. 1). [conference presentation]. RSNA 2022, Chicago, IL

Regional disparities in visual assessment of breast density: implications for risk stratification in breast cancer detection

Disparities in Visual Breast Density Assessment Between France and the US: Implications for Breast Cancer Screening.

In this study, we examined how radiologists annotated cases using the breast density assessment from MammoScreen. The goal was to not have a biased algorithm from MammoScreen, specifcally with breast density.

G. Operto, S. Pacilè, J. Guillaumin, and P. Fillard, “Regional disparities in visual assessment of breast density: implications for risk stratification in breast cancer detection,” in 17th International Workshop on Breast Imaging (IWBI 2024), SPIE, May 2024, pp. 132–140. doi: 10.1117/12.3025328

Learning General Cancer Distribution: Generalization of AI Models to Diagnostic Images

How MammoScreen would perform on diagnostic examination

• MammoScreen, when applied to non-standard screening views and diagnostic views, has overall comparable performance to the regular screening mammograms.
• Specifically, it showed higher accuracy and better specificity at a given sensitivity level, indicating that the model generalizes well to diagnostic images.
• MammoScreen could help reduce false positive rate and avoid about 20% of unnecessary biopsy procedure.

This study showcases how MammoScreen performs the same on diagnostic screenings as it does on screening images.

S. Pacilè, Y. Nikulin, P. Fillard, and F. Chammings, “Learning general cancer distribution: generalization of AI models to diagnostic images,” in 17th International Workshop on Breast Imaging (IWBI 2024), SPIE, May 2024, pp. 506–511. doi: 10.1117/12.3026769.

Can a Screening Mammography Teaching File with AI Improve Trainees’ Interpretation Skills?

Do Trainees interpretation skills improve with MammoScreen

• Reviewing a screening mammography teaching file with MS
• Improves trainees’ ability to detect lesions suspicious for malignancy.
• An AI teaching file can be a valuable educational tool.

AI should help when using, but can it help when you’re not using it? This study found that MammoScreen helped trainees know how to detect suspicious lesions even when not using the software.

“Can a Screening Mammography Teaching File with AI Improve Trainees’ Interpretation Skills?” (2022, Jul. 13-17). [poster presentation]. ECR 2022, Vienna

Time-to-event learning paradigm as a generalized approach to estimate risk of breast cancer using image-based deep learning models

Using Deep Learning on Medical Images to Predict Breast Cancer Risk Over Time

• Light architechtural changes were applied to MammoSreen, to develop an AI-based model for breast cancer risk predition over 1 to 5 years.
• Performance was compared against the MIRAI model, a state-of-the-art AI-based risk assessment tool.
• Our early risk model showed promising performance, nearly on par with MIRAI (0.790 for our risk model compared tto 0.801 for MIRAI for the dynamic cumulative AUC).

This study examines the proof of concept risk product from MammoScreen. Within this research, we compared MammoScreen’s performance to that of MIRAI.

T. Louis, S. Pacile, and P. Fillard, “Time-to-event learning paradigm as a generalized approach to estimate risk of breast cancer using image-based deep learning models,” in 17th International Workshop on Breast Imaging (IWBI 2024), SPIE, May 2024, pp. 524–536. doi: 10.1117/12.3027038