RPS 1402 - Hot Topic: personalised imaging

Author Block: L. A. Incardona¹, S. Molinaro², S. Pieroni², B. Di Nubila³, G. M. Giuseppetti⁴, P. Belli⁵, E. Montrucchio¹, M. Franchini²; ¹La Spezia/IT, ²Pisa/IT, ³Milan/IT, ⁴Ancona/IT, ⁵Rome/IT
Purpose: Breast density is a major limitation of mammographic screening, reducing sensitivity and delaying cancer diagnosis. The P.I.N.K. Study was designed to evaluate the contribution of multimodality imaging in breast cancer diagnosis. This sub-analysis quantified mammography (MX) performance according to density and assessed the additional detection (AD) provided by ultrasound (US), digital breast tomosynthesis (DBT), and magnetic resonance imaging (MRI).
Methods or Background: The P.I.N.K. database included 29,360 women who underwent 60,270 integrated diagnostic exams between 2017 and 2025. A total of 1,246 breast cancers were surgically confirmed. MX sensitivity and the AD of US, DBT, and MRI were calculated across density categories (A–D).
Results or Findings: Of 1,246 cancers, 422 occurred in fatty breasts (A–B) and 824 in dense breasts (C–D). In A–B breasts, MX detected 389 of 422 cancers (92.2%), with 33 additional cancers (7.8%) identified by other modalities. In C–D breasts, MX detected 575 of 824 cancers (69.8%), while integrated imaging identified 249 additional cancers (30.2%), increasing detection by up to 25%. The benefit was greatest in younger women and in those with extremely dense tissue.
Conclusion: Integrated imaging adds little in fatty breasts but significantly improves detection in dense breasts. On a large multicentre scale, P.I.N.K. demonstrates that tailoring imaging strategies to breast density enhances early diagnosis while avoiding unnecessary procedures in low-density women, supporting more effective and resource-efficient screening pathways.
Limitations: This was a retrospective analysis of clinical 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: G. Zanetti, E. B. Nikolova, J. K. Weber, J. Wieler, A. Boss, T. Frauenfelder, M. Marcon; Zürich/CH
Purpose: To investigate the value of supplemental breast ultrasound (SBU) in women with a personal history of breast cancer (PHBC) undergoing mammography surveillance
Methods or Background: In this retrospective study asymptomatic women with PHBC and undergoing SBU between January 2015- May 2016 with a follow-up of at least 24 months or classified BI-RADS 4/5 and histological evaluation were included. The number of additional malignant lesions and PPV of BI-RADS category were compared between mammography and SBU.
Results or Findings: Of the included 1089 examinations (992 patients, mean age SD 62.9 ± 11.9 years) 75.7% underwent breast conserve surgery and 24.3% underwent mastectomy. The follow-up mean time was 6.7 ± 5.1 years for BCS and 8.0 ± 6.7 years for mastectomy. All malignant breast lesions diagnosed at SBU were also visible in mammography (0.2% in the BCS and 1.0% in the mastectomy group). SBU identified two (0.2%) and one (0.3%) additional axillary lymphnode metastasis respectively in the BCS and in the mastectomy group . A PPV of 0% was found for all BI-RADS 3 cases on SBU, mammography or combined imaging. Lesions classified BIRADS 4 at SBU showed PPVs of 33.3% and 50% in the BCS and mastectomy group, respectively. In 0.5% of BCS and 0.3% of mastectomy examinations, additional biopsies of BI-RADS 4 lesions with benign results were performed.
Conclusion: SBU provided added value only for detecting axillary recurrences but not for identifying additional breast cancers. Its use should be balanced against increased rates of biopsy and of short-term follow-up.
Limitations: This is a retrospective study in a single academic center with a relatively small sample size. The included patients underwent different treatments at different periods and this could have influenced the recurrence rate.
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: Ethics approval KEK 2016-00064

Author Block: E. D’Ascoli, C. Depretto, G. Rossini, C. De Berardinis, G. Irmici, G. Della Pepa, L. Corradini, G. P. Scaperrotta, F. Sardanelli; Milan/IT
Purpose: To evaluate the correlation between background parenchymal enhancement (BPE) on contrast-enhanced mammography (CEM) and breast cancer (BC) risk, in order to identify a useful indicator for risk stratification and the adoption of personalised screening strategies.
Methods or Background: We retrospectivly evaluated CEMs performed at our Institution between 2021 and 2022, including patients diagnosed with BC and patients who underwent CEM for problem solving, with at least 2-year negative follow-up. For each examination, breast density (BD) and BPE (none/minimal = 1, mild = 2, moderate = 3, severe = 4) were assessed according to the BI-RADS lexicon by two independent readers with at least 3 years of CEM experience. For BC cases, BD and BPE were assessed by evaluating the contralateral breast. For each BD class, BPE scores were compared between patients with BC and healthy controls in order to assess the association between BPE and increased cancer risk.
Results or Findings: We performed a preliminary analysis on 179 patients. For each BD category, 25 patients with BC and 25 healthy controls were included, with the exception of category A, for which only 4 healthy subjects were identified. In all BD categories, the average BPE was higher in patients with BC than in controls. Specifically, in category A, the average BPE was 1.36 in cancer patients and 1.25 in healthy subjects (increase of 8.8%); in category B, 1.68 vs 1.56 (+7.7%); in category C, 2.20 vs. 1.64 (+34.1%); in category D, 2.92 vs. 2.32 (+25.9%).
Conclusion: The analysis showed an association between high BPE and BC, regardless of density; BPE could therefore be integrated into risk assessment models and represent a useful indicator for population stratification, promoting personalised screening.
Limitations: Retrospective monocentric study with limited number of patients.
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by the Independent Ethics Committee at the Fondazione IRCCS Istituto Nazionale dei Tumori, Milano.

Author Block: G. Barcaroli, F. Galati, R. Maroncelli, C. De Nardo, V. Rizzo, G. Moffa, F. Pediconi; Rome/IT
Purpose: Digital breast tomosynthesis (DBT) is a valuable imaging modality for breast cancer detection; however, its interpretation remains time-consuming and subject to inter- reader variability. This study aimed to develop and evaluate two deep learning (DL) models based on transfer learning for the binary classification of breast lesions (benign vs.malignant) using DBT images to support clinical decision-making and risk stratification.
Methods or Background: In this retrospective monocentric study, 184 patients with histologically or clinically confirmed benign (107 cases, 41.8%) or malignant (77 cases, 58.2%) breast lesions were included. Each case underwent DBT with a single lesion manually segmented for radiomic analysis. Two convolutional neural network (CNN) architectures—ResNet50 and DenseNet201—were trained using transfer learning from ImageNet weights. A 10-fold cross-validation strategy with ensemble voting was applied. Model performance was evaluated through ROC–AUC, accuracy, sensitivity, specificity, PPV, and NPV.
Results or Findings: The ResNet50 model outperformed DenseNet201 across most metrics. On the internal testing set, ResNet50 achieved a ROC–AUC of 63%, accuracy of 60%, sensitivity of 39%, and speci- ficity of 75%. The DenseNet201 model yielded a lower ROC–AUC of 55%, accuracy of 55%, and sensitivity of 24%. Both models demonstrated relatively high specificity, indicating potential utility in ruling out malignancy, though sensitivity remained suboptimal.
Conclusion: This study demonstrates the feasibility of using transfer learning-based DL models for lesion classification on DBT. While the overall performance was moderate, the results highlight both the potential and current limitations of AI in breast imaging. Further studies and approaches are warranted to enhance model robustness and clinical applicability.
Limitations: Retrospective and monocentric study, small cohort
Funding for this study: None
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: C. M. Weiss, E. Di Gaetano, E. Cattarin, R. Cerniato, I. Vinci, G. Soppelsa, G. Morana; Treviso/IT
Purpose: Recent studies show that AI-algorithms for breast cancer (BC) diagnosis have potential applications in risk-assessment, specifically in utilizing exam risk scores (EXRS) to predict the likelihood of developing the disease. We examined whether ExRS on baseline negative screening mammograms (SM) could predict BC detected in the subsequent round in an AI-supported-screening (AISS) using human-double-reading, and evaluated if these results were consistent across different BI-RADS breast-density categories (BD).
Methods or Background: A retrospective analysis was conducted on 135,372 SMs from two consecutive AISS rounds, examining 67,686 women between November 2021 and July 2025, with an average interval of 777 days between rounds. AI assigned an ExRS (0–100) to each SM. ExRS values were compared between women who developed BC (451/67,686) and those who did not (67,235/67,686), with subgroup analyses by BD. Statistical tests included chi-square or z-tests, t-test or Mann–Whitney U-test, and McNemar’s test.
Results or Findings: Among 451 BC cases, mean ExRS rose from 15.4 at baseline to 73.9 in the subsequent round (median 6.8vs83.0;p<0.001). Women negative at both rounds (n=67,235) had little change: means were 6.7 and 6.4 (medians 2.1 and 2.3;p<0.001). ExRS was significantly higher in BC cases than negatives at both baseline (15.4vs6.7;p<0.001) and subsequent (73.9vs6.4;p<0.001). Baseline ExRS were higher in dense (C–D) than non-dense (A–B) breasts. For BC cases, mean ExRS increased from 13.8 to 73.7 (A–B) and 17.6 to 74.1 (C–D). In negatives, ExRS stayed low: 6.0 to 5.5 (A–B) and 7.6 to 7.9 (C–D).
Conclusion: The AI-derived ExRS is able to differentiate between women with varying levels of risk for developing BC at baseline, and this ability was consistent across all BD categories. The results show that ExRS can be also used for risk-based stratification in screening
Limitations: No limitations
Funding for this study: No fundings
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: J. Peters¹, D. Van Der Waal¹, M. Schmidt², C. Van Gils³, M. Broeders¹; ¹Nijmegen/NL, ²Amsterdam/NL, ³Utrecht/NL
Purpose: Tailoring breast cancer screening to individual risk can improve its harm–benefit ratio versus a ‘one-size-fits-all’ approach. We externally validated four widely used risk models—Gail, BCSC, BOADICEA, and IBIS—in a large Dutch screening cohort. Recent model updates include breast density using different approaches (categorical or continuous), so we evaluated performance with and without density.
Methods or Background: The PRISMA study is a prospective cohort (2014–2019) embedded in the Dutch biennial breast cancer screening program, which invites women aged 50-75 years. Breast cancer risk was predicted for 38,767 participants using four models (Gail, BCSC, BOADICEA, IBIS) based on questionnaire data (personal, lifestyle, hormonal, family history) and breast density. Gail does not include density; BCSC uses density categories; BOADICEA and IBIS allow categories, a continuous measure, or no density. Volumetric percent density and Volpara Density Grades (Volpara version 1.5.0) were measured on raw mammograms. Breast cancers were ascertained via linkage with the Netherlands Cancer Registry until October 2023. Model performance for 5-year risk was evaluated using the concordance index (C-index), observed–expected (O/E) ratio, and calibration slope.
Results or Findings: During a median 4.3 years follow-up, 609 breast cancers occurred. Discrimination was poor for Gail (C-index 0.56) and modest for BCSC, BOADICEA and IBIS (C-indices 0.60-0.61). Calibration was good for BCSC (O/E 1.03, slope 0.81) and BOADICEA (O/E 0.96, slope 0.78), but IBIS (O/E 0.71, slope 0.66) and Gail (O/E 0.79, slope 0.59) overpredicted risk. Adding continuous breast density improved discrimination (ΔC-index +0.02–0.04) and calibration most.
Conclusion: Traditional breast cancer risk models show at most moderate performance; despite small improvements from continuous breast density, overall accuracy remains limited for personalized screening.
Limitations: The limitations of the study are incomplete family history and genetic data, which may particularly underestimate BOADICEA and IBIS performance.
Funding for this study: Dutch Cancer Society (KWF7626) and Dutch Research Society (ZonMw 200500004)
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: CMO Arnhem-Nijmegen reference no. 2014/177

Author Block: M. Keupers, W. Sarkol, S. Nijssen, Y-K. Wang, L. Cockmartin, H. Bosmans, C. Van Ongeval; Leuven/BE
Purpose: For women with clinically elevated breast cancer risk due to personal risk factors, family history or breast density, screening guidelines remain ambiguous. Epidemiological risk-models like CanRisk and IBIS, are currently used to guide supplemental screening. Recent studies demonstrated that deep learning models might be superior. This study compares risk assessment of two epidemiological models with two deep learning models.
Methods or Background: In this single-center cohort 401 women with clinically elevated breast cancer risk were included (2014) with ten-year follow-up info. CanRisk v2.4.2. and IBIS v8.0b assessed lifetime and short-term risk based on patient files. Deep learning models i.e. MIRAI and Transpara were used to predict short-term risk. Model performance was evaluated using receiver operating characteristic (ROC) analysis; areas under the curve (AUCs) were calculated and compared using paired DeLong’s test.
Results or Findings: Of the 401 women (mean age=49 ±7) 25 developed breast cancer after a ten-year period; 11 cancers were detected after five years. The AUCs for risk assessment within five and ten years for CanRisk were 0.58 (95% confidence interval (CI) [0.37; 0.79]) and 0.61 (95% CI [0.49; 0.73]) resp, and for IBIS lifetime risk it was 0.61 (95% CI [0.49; 0.72]). Deep learning models were evaluated to predict breast cancer risk after five years. Transpara (0.73 (95% CI [0.52; 0.95]) showed superior performance over Mirai (0.57 (95% CI [0.38; 0.77]) (p=0.042).
Conclusion: In this single-center cohort with elevated breast cancer risk, one deep learning model outperformed the epidemiological models for short term risk prediction. Generalizability of deep learning models for increased-risk populations should be tested. Further evaluation in larger cohorts is needed.
Limitations: Number of patients in this single-center, increased-risk cohort.
Still some missing data despite thorough review of electronic patient files.
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 Ethical Committee of the Catholic University Leuven, Belgium (MP029292 and S70006).