RPS 1002 - Breast cancer screening: technology, tools, and trends

Author Block: S. Murphy, T. Mooney, N. Phelan, A. Smith, A. Connors, A. Larke, S. Mcnally, P. Fitzpatrick, M. Mullooly; Dublin/IE
Purpose: To examine recall patterns and characteristics within the population-based breast screening programme in Ireland.
Methods or Background: Breast cancer screening aims to reduce breast cancer mortality and morbidity through early detection and treatment. Recall rate is a key performance indicator of population-based breast screening, representing the proportion of women recalled for further evaluation. Guidance on acceptable recall rates vary internationally.

An anonymous aggregate retrospective study of 2,031,995 mammography screening examination results, was conducted between 2000 and 2019. Descriptive patterns of recall rates and characteristics were examined and stratified by prevalent and incident examinations. Differences across the time-periods (2000-2008, 2009-2017 and 2018-2019) were assessed using Chi-square tests.
Results or Findings: Recall rate for screening examinations conducted during the full study period was 4.05% (n=82,338/2,031,995). Across three time-periods examined, recall rates among the prevalent screening examination group, increased, from 5.5% to 8.0% to 10.0% and within the incident group from 2.3%->2.8%->3.0%. Recalls due to calcifications and asymmetry increased over the time periods, most notably within the prevalent examinations where recalls due to calcification increased from 6.0/1,00->9.0/1,000->13.4/1,000(p<0.001),whilst recalls due to asymmetry increased from 17.1/1,000->31.3/1,000->41.0/1,000(p<0.001). Overall, among both prevalent and incident screening examinations, an increase in the cancer detection rate (CDR) was observed(p=0.005 and p<0.001 respectively). However, the overall positive predictive value(PPV) remained relatively stable.
Conclusion: This study highlights the upward trajectory of recall within Ireland’s national breast screening service. The findings highlight the need for discussions among a diverse range of stakeholders, including national and international screening networks, to determine the optimal recall rate to ensure the benefits of screening are maximised and all potential harms are minimised.
Limitations: Lack of information regarding detailed characteristics of the cancers detected and information regarding breast density.
Funding for this study: No
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: As this study utilises existing anonymised data, formal ethical approval was not required in accordance with local ethics review board guidelines.

Author Block: L. Kregting, L. Pennings, M. J. A. Smid-Geirnaerdt, A. Bluekens, J. Gommers, S. D. Verboom, I. Sechopoulos, M. Broeders; Nijmegen/NL
Purpose: To evaluate the short-term screening performance of Digital Breast Tomosynthesis (DBT) in the first round of the DBT with Advanced Reading Methods (STREAM) prospective screening trial being performed in the Dutch Breast Cancer Screening Programme.
Methods or Background: The Dutch screening programme includes biennial four-view digital mammography (DM) for women 50 to 75 years old. All examinations are independently double-read by screening radiologists with consensus reading or arbitration in case of disagreement. STREAM trial participants underwent the first of two rounds of four-view DBT screening instead of DM between July 2023 and May 2024. A contemporaneous control group received regular DM screening. Current analyses include first round results for recall rate (RR) , stratified by Breast Imaging Reporting and Data System (BI-RADS) category 0, 4, or 5, and cancer detection rate (CDR) .
Results or Findings: Out of 44,682 invited screenees, 18,186 participated in STREAM (41%). Among them, 488 were recalled (RR: 26.8 per 1,000), of whom 199 (1.1%) were assessed as BI-RADS 0, 257 (1.4%) as BI-RADS 4, and 32 (0.2%) as BI-RADS 5. Of the recalled women, 166 women were diagnosed with screen-detected breast cancer (CDR: 9.1 per 1,000). The control group consisted of 95,052 participants of DM screening. Among them, 2,252 were recalled (RR: 23.7 per 1,000), of whom 661 were diagnosed with breast cancer (CDR: 7.0 per 1,000). This results in an increase in RR of 2.9 per 1,000 (p=0.012) and an increase in CDR of 2.1 per 1,000 (p=0.002) for DBT screening compared to DM.
Conclusion: First round results suggest a modest increase in RR and an increase in CDR for DBT screening compared to DM in the Dutch breast cancer screening programme.
Limitations: Interval cancer and second round data not available yet.
Funding for this study: This study is funded by ZonMW (grantnr.:5550402130002) and KWF (grantnr.:13710).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Trial approved by the Dutch Minister of Health, Welfare and Sport in view of the Population Screening Act (WBO). (registrationnr.:3574709-1044141-PG).

Author Block: L. A. Incardona¹, M. Franchini², S. Molinaro², S. Pieroni², J. Nori³, G. P. Scaperrotta⁴, A. Nicolucci³, E. Cassano⁵, E. Montrucchio¹; ¹La Spezia/IT, ²Pisa/IT, ³Florence/IT, ⁴Milano/IT, ⁵Milan/IT
Purpose: Mammography (MX) is the cornerstone of breast cancer screening but has limited sensitivity in specific subgroups. The P.I.N.K. Study was designed to evaluate the role of multimodality imaging in breast cancer diagnosis. This sub-analysis aimed to characterise cancers missed by MX and detected only by additional imaging.
Methods or Background: The P.I.N.K. database included 29,360 women undergoing 60,270 integrated diagnostic exams between 2017 and 2025. Among 1,246 surgically confirmed breast cancers, 233 (18.7%) were not visible on MX and were diagnosed exclusively by ultrasound (US), digital breast tomosynthesis (DBT), or magnetic resonance imaging (MRI). Breast density, histology and biological subtype were analysed.
Results or Findings: Missed cancers accounted for 233 of 1,246 cases (18.7%). Most were invasive (80%, 186/233), multifocal in nearly one third (32%, 74/233), and predominantly found in dense breasts (75%, 175/233). Younger women under fifty years were affected (40%). Histologically, invasive lobular carcinoma represented about 20% (47/233), while invasive ductal carcinoma accounted for the majority. Regarding biological profile, Luminal A was the most frequent subtype (55%, 128/233), followed by Luminal B (25%, 58/233), with HER2-positive (12%, 28/233) and triple-negative tumours (8%, 19/233) less common. US detected 142 of 233 missed cancers (60.9%), DBT 56 (24.0%), and MRI 35 (15.0%). These findings indicate that missed cancers are clinically significant and not indolent.
Conclusion: In this large multicentre cohort, almost one in five breast cancers were missed by MX but detected through integrated imaging. Their profile (dense breasts, younger women, invasive histology, and distinct biological features) highlights the clinical relevance of multimodality approaches. P.I.N.K. shows that the value of integrated imaging is both quantitative and qualitative, preventing missed cancers from becoming interval cancers and supporting personalised diagnostic strategies.
Limitations: This was a retrospective analysis of observational data.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: S. V. Grinsven¹, E. Monninkhof¹, R. Mann², W. B. Veldhuis¹, C. Van Gils¹, F. T. D. T. S. G. -¹; ¹Utrecht/NL, ²Nijmegen/NL
Purpose: To study the effect of supplemental MRI screening for women with extremely dense breasts on advanced breast cancers.
Methods or Background: Dutch women with extremely dense breasts and a negative mammogram were pre-randomized to either the MRI-invitation (n=8,061) or control group (n=32,312, standard mammography). In the first (prevalent) round, supplemental MRI led to higher cancer detection and fewer interval cancers. The rate of advanced breast cancers in subsequent (incident) screening rounds serves as a further important measure to assess the impact on health outcomes. Since many women randomized to MRI-invitation did not participate, intention-to-treat analyses dilute the true effect. Therefore, our main analysis is a per-protocol approach, comparing advanced breast cancer (TNM stage II+) rates, adjusted for age and socioeconomic status using inverse probability weighting. Rate differences (RD) were calculated with 95% confidence intervals.
Results or Findings: As expected, in the first round, the advanced breast cancer rate per 1000 women was similar between MRI participants and controls (RD: 0.8 [95% CI: -0.6, 2.2]). In the second round, the rate was lower in the MRI group but not yet statistically significant (RD: 1.4 [95% CI: -0.3, 3.2]). By the third round, the rate in the MRI group was significantly lower than in the control group (RD: 2.6 [95% CI: 0.9, 4.3]).
Conclusion: From the second round, MRI participants had lower advanced cancer rates, reaching significance in the third round. These findings show that the health benefits of MRI screening likely extend beyond lowering interval cancers.
Limitations: The primary outcome used a per-protocol and not an intention-to-treat approach. As mortality could not be studied due to limited sample size and follow-up, advanced breast cancer was used as a surrogate. These results will inform mortality modelling.
Funding for this study: Supported by the University Medical Center Utrecht (project number, UMCU DENSE), the Netherlands Organization for Health Research and Development (project numbers, ZonMW-200320002-UMCU and ZonMW Preventie 50-53125-98-014), the Dutch Cancer Society (project numbers, DCS-UU-2009-4348, UU-2014-6859, and UU-2014-7151), the Dutch Pink Ribbon–A Sister’s Hope (project number, Pink Ribbon-10074), Bayer Pharmaceuticals (project number, BSP-DENSE), and Stichting Kankerpreventie Midden-West. For research purposes, Volpara Health Technologies provided Volpara Imaging Software, version 1.5, for installation on servers in the screening units.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The Dutch Minister of Health, Welfare and Sport, who
was advised by the Health Council of the Netherlands
(2011/2019 WBO, The Hague, The Netherlands),
approved the DENSE trial on November 11, 2011.

Author Block: S. Tyagi, M. M. Jabeer, J. Singh, A. Chandalia; Bengaluru/IN
Purpose: Breast cancer is the most commonly diagnosed cancer among women worldwide, accounting for 11.6% of all cases and 6.9% of cancer-related deaths (GLOBOCAN 2022). Despite advances in therapy, early detection remains the cornerstone of improving survival outcomes. Mammography is the standard screening tool, yet its effectiveness is limited by interpretive variability, radiologist shortages, and increasing case volumes that strain healthcare systems. This study aims to develop and evaluate an artificial intelligence (AI) model for automated detection and structured reporting of breast cancer in mammograms, with the goal of enhancing diagnostic accuracy, consistency, and accessibility.
Methods or Background: A retrospective dataset of 100,000 biopsy-confirmed mammographic images, including both benign and malignant cases, was used to train the AI model. The dataset was curated to capture diverse tumor types and breast densities. A convolutional neural network (CNN) architecture was employed for automated detection and classification. In addition to image analysis, the system generates structured reports summarizing suspicious findings and relevant diagnostic information. Model performance was assessed using sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). Results were benchmarked against experienced radiologists to determine clinical relevance.
Results or Findings: On an independent test set of 300 cases, the AI system achieved an AUC of 0.92, sensitivity of 96.4%, and specificity of 87.7%. The overlay and report results were independently reviewed and verified by an expert radiologist, confirming clinical reliability.
Conclusion: The AI system demonstrates strong potential for accurate and efficient breast cancer detection in mammography. Integration into clinical workflows could facilitate earlier diagnoses, reduce diagnostic delays, and alleviate radiologist workload, particularly in high-volume or resource-limited settings.
Limitations: Broader and multi-center prospective studies are required to confirm generalizability across diverse imaging environments.
Funding for this study: No funding was obtained for this work.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: M. R. López De La Torre Carretero, D. A. Zambrano, A. M. Delgado Brito, C. D. Solano, A. Elizalde, L. J. Pina Insausti; Pamplona/ES
Purpose: To evaluate the performance of an artificial intelligence (AI) tool in classifying mammograms from opportunistic screening versus clinical populations when compared with expert radiologist interpretation.
Methods or Background: This prospective multicenter study included 2,574 patients who underwent digital mammography, tomosynthesis, and targeted ultrasound when indicated. Patients were stratified into opportunistic screening (76.7%) or clinical setting (symptomatic, oncologic follow-up...) groups. Mammograms were analyzed by a commercially available AI tool, which categorized studies as “normal” or “suspicious”. Radiologists also classified studies as “normal”/”suspicious”, blinded to AI.
The reference standard was histopathology (when available) or the final radiologist assessment after ultrasound. Performance was compared using chi-square and McNemar tests, both globally and across patient subgroups.
Results or Findings: Mean age was 56.4 years (55.4 in screening, 59.7 in clinical).
The AI classified 38% of cases as suspicious (31.1% in screening and 58% in clinical patients; p < 0.0001). Meanwhile, radiologists classified 20% of cases as suspicious (22% in screening vs 13% in clinical; p < 0.0001).
In the global analysis, AI sensitivity was 87.9%, with limited specificity (63.9%), with 64.4% accuracy. Radiologists showed 91.3% sensitivity, with significantly higher specificity (86.3%), overall precision (86.5%), and agreement with the reference standard (kappa = 0.20 vs. 0.06 for AI).
Stratified analysis showed that AI performed better in screening (sensitivity 94.3%, specificity 70.1%) than in clinical settings (sensitivity 84.6%, specificity 43.4%). In clinical cases, radiologists achieved 98% sensitivity and 81.2% specificity (kappa = 0.27), outperforming AI across all measures. All comparisons between AI and radiologists were statistically significant (p < 0.0001).
Conclusion: AI demonstrated robust performance in screening settings. However, its limited specificity and lower accuracy in clinical scenarios highlight the need for substantial improvements before it can match expert radiologist diagnosis.
Limitations: Retrospective
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Research Ethics Committee - Clínica Universidad de Navarra

Author Block: E. L. Van Der Veer¹, A. Bluekens¹, A. M. P. Coolen-Janssen¹, W. Vreuls², A. Voogd³, L. Duijm²; ¹Tilburg/NL, ²Nijmegen/NL, ³Maastricht/NL
Purpose: The prevalence of breast lesions of uncertain malignant potential, commonly referred to as high-risk lesions (HRLs), detected at screening mammography is increasing, which stresses the importance of the recently published guidelines on their management. In this study, we present the incidence, upgrade rate, adherence, and follow-up outcomes in a screened population.
Methods or Background: A total of 17,809 recalled women, who took part in the Dutch screening program between January 2009 and July 2019, were included in this retrospective analysis of a prospectively obtained database. A HRL was identified in 537 recalled women and their incidence rate, diagnostic work-up and follow-up, and any upgrade to (pre)malignancy after surgical excision and during follow-up were examined.
Results or Findings: The incidence rate of HRLs was 0.87 per 1,000 screens and 3.02 per 100 recalls. The majority of HRLs consisted of papillary lesions (32.4%) and atypical ductal hyperplasia (ADH) (21.6%). Surgical excision was performed for 254 (47.3%) of the 537 HRLs, resulting in 59 (23.2%) HRLs being upgraded to (pre)malignancy. Invasive tumors were mainly found in papillary lesions with atypia (n=9/17) and ADH (n=6/17). Of the women diagnosed with an HRL, 60.7% had an indication for radiologic follow-up. Of these women, 79.1% actually underwent follow-up with 43.0% undergoing 5-year surveillance. During follow-up, 28 (10.9%) of the women with an HRL developed a (pre)malignancy.
Conclusion: Of women with screen-detected HRLs, 16.2% developed breast cancer, either after surgical excision or during follow-up. Less than half of the women completed the 5-year surveillance, highlighting the importance of continued attention to obtain maximal follow-up adherence.
Limitations: ADH diagnosis varies between observers due to its similarity to low-grade DCIS, with biopsy type and sample size affecting outcomes. Some subgroups were too small for comparison with guidelines.
Funding for this study: This research did not receive any funding.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: F. Mohsin¹, M. Khalid¹, A. Maryam¹, S. Chaudhry², K. Siddique²; ¹Karachi/PK, ²Lahore/PK
Purpose: BI-RADS category 3 lesions are typically benign (<2% malignancy risk), managed with follow-up imaging, but their upgrade to BI-RADS 4 may indicate malignancy, particularly in high-risk patients. This study evaluates malignancy risk and predictors among BI-RADS 3 lesions upgraded to BI-RADS 4 in a cohort with a high prevalence of breast cancer history.
Methods or Background: We retrospectively analyzed 352 patients undergoing breast ultrasound from June 2021 to June 2023, with 123 categorized as BI-RADS 3. Of these, 17 patients (18 lesions) were upgraded to BI-RADS 4 and biopsied. Data included age, prior breast cancer history (100/123, 81.3%), lesion characteristics (size, number, morphology), and histopathology. Descriptive statistics, binomial tests, Fisher’s exact tests, and logistic regression were used to assess upgrade rates, malignancy rates, and predictors, with significance at p=0.05.
Results or Findings: The upgrade rate was 13.8% (17/123; 95% CI: 8.5–21.0%), significantly exceeding the ACR’s <2% benchmark (p=0.05). Among upgraded lesions, 11.1% (2/18; 95% CI: 1.4–34.7%) were malignant (recurrent breast cancer, lymphomatous lymph node), also surpassing the <2% benchmark (p=0.05). Benign outcomes mainly included stromal fibrosis (38.9%) followed by benign lymphadenopathy (27.8%), and fibroadenoma (22.2%). Lesion size (OR=1.15 per mm, p=0.05) and multiple lesions (OR=3.2, p=0.05) predicted malignancy, while age, cancer history, and morphology were not significant (p>0.05).
Conclusion: The high upgrade and malignancy rates in this high-risk cohort suggest that BI-RADS 3 lesions with larger size or multiple foci, particularly in breast cancer survivors, may warrant earlier biopsy to detect malignancy, emphasizing the need for refined surveillance criteria.
Limitations: Not applicable.
Funding for this study: No funding was provided for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: None

Author Block: N. F. Clerkin¹, C. Ski¹, P. Brennan², R. M. Strudwick³; ¹Belfast/UK, ²Sydney/AU, ³Ipswich/UK
Purpose: Radiography Advanced Practitioners (RAPs) have interpreted mammograms in the United Kingdom since 1995 (1). Evidence on factors influencing RAP diagnostic performance remains limited. This study aimed to identify reader and image-based factors affecting RAP performance in mammography interpretation.
Methods or Background: The research comprised of three components. First, a systematic literature review, identifying 38 studies on performance variation in mammography. Second, experimental work which involved 18 UK based RAPs interpreting 60 cloud hosted mammograms with known truth; results were analysed against reader characteristics. Third, a comparative observer study focussed on factors that impact performance, RAPs (n=18) and radiologists (n=24) interpreted the same cases. Difficulty indices were calculated for all cases using an established methodology and causal agents influencing performance were explored.
Results or Findings: The review highlighted a paucity of RAP specific performance data. A range of findings included, higher ROC values with: less compared with more than 10 years’ experience (p=0.004); more compared with less than 100 cases read per week (p=0.036); consistent rather than sporadic use of prior mammograms (p=0.0231); recent rather than distant eye testing (p=0.021). Specificity was lower amongst RAPs who stated their emotional mindset impacted their reading (88) compared to those who did not (66.5) (p=0.034). Strong correlations were found between RAPs and radiologists for cases with specific levels of difficulty (cancer: r = 0.83; normal: r = 0.73), uniquely to RAPs, soft tissue cancers compared with calcifications as well as cases without prior images were more challenging.
Conclusion: For the first time, factors that impact performance of RAPs have been identified. Whilst some factors are common for radiologists and RAPs, some are unique, suggesting that extending findings to both must be done with caution.
Limitations: Limitations included questionnaire responses and gender representation.
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Ethical approval was gained from the University of Suffolk Post Graduate Research Committee on the 14th of January 2022 (RETH(P)21/006) .
NHS Ethical considerations were achieved through the NHS Health Research Authority (HRA). The HRA confirmed no Research Ethics Committee review was required.
Ethical exemption was also provided by the University of Sydney to include radiologist performance data in the study.

Author Block: P. D. (. Trieu, M. Dimigen, M. Barron, S. Lewis; Sydney/AU
Purpose: To explore the effectiveness of an intelligent, adaptive training approach based on Bayesian methodology for improving diagnostic performance among breast screening readers using training test sets.
Methods or Background: Seven breastscreen readers, including radiologists, breast physicians, and radiology trainees, participated in adaptive training sessions using the BREAST screening mammogram database which comprised 1,753 test set completions by 618 breast screening readers. Training was delivered on the BREAST (Breastscreen-REader-Assessment-STrategy) platform. The newly developed AI-driven Bayesian algorithm dynamically selected normal and cancer cases tailored to each reader’s current performance level. Training continued until each reader achieved a diagnostic threshold of ≥90% in both sensitivity and specificity, or ROC AUC. Readers' performance before and after training was compared using the Wilcoxon signed-rank test.
Results or Findings: All readers reached the performance goal after 2 to 5 training sessions over the period from 3 to 6 months. Statistically significant improvements were observed in key diagnostic metrics, including overall case sensitivity (from 0.822±0.079 to 0.943±0.044; P=0.015), lesion-level sensitivity (0.764±0.103 to 0.886±0.073; P=0.043), ROC AUC (0.861±0.029 to 0.935±0.034; P=0.016), and JAFROC (0.082±0.038 to 0.895±0.049; P=0.018). The improvement in sensitivity was consistent across various breast density levels and lesion types, including irregular masses, calcifications, and small lesion sizes (15%-to-18%; P<0.05).
Conclusion: This study demonstrates that an AI-based Bayesian adaptive training algorithm effectively enhances breast image interpretation performance. The personalized selection of training cases significantly improved diagnostic accuracy across multiple performance metrics and lesion characteristics. These findings support the integration of intelligent, adaptive training models into breast screening education to optimize reader performance and diagnostic outcomes.
Limitations: Although the number of participating readers is currently limited, the study is ongoing with additional readers being recruited. Expanded data and updated results will be presented at the conference.
Funding for this study: Seed Funding - The Sydney Cancer Institute
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The University of Sydney

Author Block: T. Jezek¹, A. Beck-Toelly¹, N. Miłecki¹, N. Ailec¹, N. Pötsch¹, A. D'Angelo², P. A. Baltzer¹, T. H. Helbich¹, P. Clauser¹; ¹Vienna/AT, ²Rome/IT
Purpose: The objective of the study was to assess the participation of the transgender (TG) population in breast and prostate cancer screening programs.
Methods or Background: This retrospective, monocentric study was approved from the local ethics committee. All transgender patients attending the dedicated local clinic were included. Data on breast imaging examination and prostate specific antigen (PSA) controls were retrieved by three investigators from the local information system. Exclusion criteria were lack of data in the system, no hormonal therapy. Data retrieved were type and duration of the hormonal therapy, family history for breast and prostate cancer, personal history of breast or prostate cancer and previous PSA and mammography exams. Descriptive statistics was used to present the data.
Results or Findings: To date, 750 TG individuals have been evaluated (mean age 34.7, standard deviation SD 16.9; 389 TG male and 361 TG female). 68 had a positive family history for breast or ovarian cancer (9.1%), and 16 (2.1%) for prostate cancer. Only one individual with a personal history of breast cancer was identified (TG male). Of the 209 individuals above 40 years old (mean age 53.6, SD 9.5), only 13 (one TG male, 12 TG female) participated in mammography screening (6.2%). Among the 126 TG females, 50 underwent a PSA test (39.7%). No breast or prostate cancers were diagnosed during or after the transition.
Conclusion: Our findings show a very low participation in breast cancer screening among the TG population, even in individuals with a positive family history. Participation in prostate cancer testing was higher, most likely because it is conducted through a simple blood test. This indicates a widespread lack of awareness regarding available screening programs and their relevance.
Limitations: Retrospective study involving many patients with incomplete data.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was approved by the local ethics committee (1183/2023).

Author Block: G. Tiberi¹, D. Álvarez Sánchez-Bayuela², N. Ghavami¹, R. Loretoni¹, T. Frauenfelder³, C. Alves⁴, M. A. B. Álvarez Benítez⁵, M. Calabrese⁶, A. S. Tagliafico⁶, J. Lubinski⁷, C. Romero Castellano²; ¹Perugia/IT, ²Toledo/ES, ³Zürich/CH, ⁴Lisbon/PT, ⁵Córdoba/ES, ⁶Genoa/IT, ⁷Szczecin/PL
Purpose: To present interim results of the study “A Clinical Investigation to Evaluate Microwave Imaging via MammoWave in a Population-based Screening Program for Early Breast Cancer (BC) Detection” (ClinicalTrials.gov ID:NCT06291896), activated within HORIZON-MISS-2021-CANCER-02-01 scheme.
Methods or Background: Women undergoing routine screening mammograms were invited to join the study. After providing informed consent, participants underwent microwave breast imaging using MammoWave. MammoWave’s output was automatically labelled by an artificial intelligence (AI) model (based on hierarchical machine-learning approach), classifying each breast as “NSF” (no-suspicious findings, no lesion or lesion with low suspicion) or “WSF” (with-suspicious findings, indicating presence of a suspicious lesion). MammoWave’s results are compared to a reference standard, defined as the outcome of the conventional breast examination pathway—including histological confirmation—with a two-year follow-up. Reference standard will be classified as ‘positive’ when BC is confirmed by histology, and ‘negative’ otherwise. Primary outcomes are sensitivity/specificity of MammoWave’s AI model in BC detection.
Results or Findings: We report findings from the first 3,000 volunteers enrolled at 9 hospitals across 5 European countries. The interim analysis was performed on 5,896 breasts (12 ‘positive’ cases), from 2,967 subjects included in full analysis set population (mean age 57.2 years±7.5 [SD]). This interim analysis allowed us to evaluate AI model’s performance, with 42% sensitivity [95%CI:13.8-69.6] and 75% specificity [95%CI:74.1-76.3]. We proceeded to update AI model by fusing information of hierarchical machine-learning approach and statistical classifiers. Updated AI model allowed us to (retrospectively) reach: 67% sensitivity [95%CI:29–99] (80% in dense breasts [95%CI:45–99]); 81% specificity 95%CI:79.2-83.2] (80% in dense breasts [95%CI:77–83]).
Conclusion: MammoWave’s AI model demonstrates promise in breast cancer screening, particularly in terms of specificity; sensitivity requires further optimization. Final results of this trial on 10,000 volunteers will provide more insights.
Limitations: Non-randomized design; small cancer sample size.
Funding for this study: This work was supported by the funding received by the MammoScreen project, co-funded by the European Union’s Horizon research and innovation framework programme, Grant agreement 101097079.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by the Research Ethics Committee of the Liguria Region (CET), Italy, on 13 November 2023 (approval number: CET - Liguria: 524/2023 - DB id 13399), the Research Ethics Committee of Complejo Hospitalario de Toledo (CEIC), Spain, on 29 November 2023 (approval number: CEIC - 1094), the National Ethics Committee for Clinical Research (CEIC), Portugal on 12 January 2024 (approval number: CEIC - 2311KC814), the Bioethical Committee of Pomeranian Medical University in Szczecin, Poland (approval number: KB-006/23/2024) on 13 March 2024, and the Zurich Cantonal Ethics Commission, Swiss (BASEC 2023-D0101) on 7 June 2024.

Author Block: V. Batra, J. Harvey; Rochester, New York, NY/US
Purpose: Breast arterial calcifications (BAC) on mammography are frequently underreported. Growing evidence suggests that BAC correlate with coronary artery disease (CAD) and may serve as a surrogate marker of cardiovascular risk in women undergoing routine breast cancer screening. In women without a known history of cardiovascular disease, consistent recognition and reporting of BAC could facilitate early risk stratification and preventive intervention.
Methods or Background: Using the TriNetX database, we conducted a retrospective analysis of women aged 40–50 years who underwent screening mammography, identified through ICD-10 codes. We quantified the total number of screening exams, the proportion with reported calcifications, and the subset of patients with documented CAD. We also examined the potential underestimation of BAC prevalence due to their frequent classification as BI-RADS 1 (negative) rather than BI-RADS 2 (benign findings).
Results or Findings: Across 109 U.S. healthcare organizations, 6,783,908 screening mammograms were performed in women aged 40–50 years. Of these, 2,396,182 examinations were reported as abnormal or inconclusive. A total of 6,070,184 women in this age group had a diagnosis of CAD. The intersection of women with both BAC and CAD included only 4,420 patients—substantially lower than expected—suggesting significant underreporting likely related to BI-RADS coding variability.
Conclusion: Coronary artery disease remains the leading cause of death among women in the United States. The inconsistent reporting of vascular calcifications on mammography may obscure an important opportunity for early cardiovascular risk detection. Standardizing the reporting of BAC as BI-RADS 2 could enhance identification of subclinical atherosclerosis and support earlier, targeted cardiovascular prevention in women without known heart disease.
Limitations: This study is limited by its retrospective design and reliance on ICD-10 coding within the TriNetX database, which may lead to underreporting or misclassification of vascular calcifications and coronary artery disease.
Funding for this study: None
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: J. Sänger¹, J. Happe¹, A. Dudle², C. Ruppert², A. Weber³, A. Ciritsis², T. Frauenfelder¹, A. Boss³; ¹Zürich/CH, ²Zurich/CH, ³Wetzikon/CH
Purpose: To determine the prevalence of breast arterial calcifications (BAC) in a large cohort of opportunistic mammography breast cancer screening using artificial intelligence (AI) and to highlight both their potential diagnostic overlap with grouped microcalcifications and their role as an opportunistic biomarker for cardiovascular risk assessment.
Methods or Background: This retrospective study included 4692 women who underwent a total of 6,061 mammographic examinations with a total of 24,519 mammographic images between 2011 and 2017 at a tertiary centre. A YOLO-based object detection algorithm was trained on standard mammographic projections (CC and MLO) to detect and differentiate BAC and non-vascular calcifications, assigning BI-RADS categories to the latter.
Results or Findings: BAC was found with AI-based analysis in 652 of 4,692 women (13.9%, 95% CI 12.9–14.9). Among 6,061 individual examinations, BAC were present in 800 (13.2%, 95% CI 12.4–14.1). Considering all 24,519 mammographic images, BAC was detected in 1,644 images (6.7%, 95% CI 6.4–7.0). At the time of the first mammogram, BAC-positive woman had a mean age of 66.6 ± 10.9 years, whereas those without BAC had a mean age of 56.3 ± 11.3 years (p < 0.001).
Conclusion: Artificial intelligence applied to mammography revealed that one in seven women has BAC. Early BAC can be challenging to distinguish from suspicious grouped microcalcifications, and systematic AI-based screening during routine breast imaging may aid differentiation. Moreover, BAC recognition is clinically relevant as it correlates with CAC and cardiovascular risk. Embedding automated BAC detection into established breast cancer screening programmes may therefore provide a low-cost, scalable opportunity for opportunistic cardiovascular risk assessment, with the potential to enable earlier identification of women at risk and contribute to risk reduction.
Limitations: The limitations of the study are the retrospective design and the single-centre setting.
Funding for this study: Funding was provided by the Swiss Cancer League (KFS-5524-02-2022) and the Clinical Research Priority Program “Artificial Intelligence in oncological Imaging” of the University of Zurich.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was approved on 6 July 2021 by the Institutional Review Board (No. 2021-01095).