A prospective study of HER3 expression pre and post neoadjuvant therapy of different breast cancer subtypes: implications for HER3 imaging therapy guidance.

PURPOSE: HER3, a member of the EGFR receptor family, plays a central role in driving oncogenic cell proliferation in breast cancer. Novel HER3 therapeutics are showing promising results while recently developed HER3 PET imaging modalities aid in predicting and assessing early treatment response. However, baseline HER3 expression, as well as changes in expression while on neoadjuvant therapy, have not been well-characterized. We conducted a prospective clinical study, pre- and post-neoadjuvant/systemic therapy, in patients with newly diagnosed breast cancer to determine HER3 expression, and to identify possible resistance mechanisms maintained through the HER3 receptor. EXPERIMENTAL DESIGN: The study was conducted between May 25, 2018 and October 12, 2019. Thirty-four patients with newly diagnosed breast cancer of any subtype (ER ± , PR ± , HER2 ±) were enrolled in the study. Two core biopsy specimens were obtained from each patient at the time of diagnosis. Four patients underwent a second research biopsy following initiation of neoadjuvant/systemic therapy or systemic therapy which we define as neoadjuvant therapy. Molecular characterization of HER3 and downstream signaling nodes of the PI3K/AKT and MAPK pathways pre- and post-initiation of therapy was performed. Transcriptional validation of finings was performed in an external dataset (GSE122630). RESULTS: Variable baseline HER3 expression was found in newly diagnosed breast cancer and correlated positively with pAKT across subtypes (r = 0.45). In patients receiving neoadjuvant/systemic therapy, changes in HER3 expression were variable. In a hormone receptor-positive (ER +/PR +/HER2-) patient, there was a statistically significant increase in HER3 expression post neoadjuvant therapy, while there was no significant change in HER3 expression in a ER +/PR +/HER2+ patient. However, both of these patients showed increased downstream signaling in the PI3K/AKT pathway. One subject with ER +/PR -/HER2- breast cancer and another subject with ER +/PR +/HER2 + breast cancer showed decreased HER3 expression. Transcriptomic findings, revealed an immune suppressive environment in patients with decreased HER3 expression post therapy. CONCLUSION: This study demonstrates variable HER3 expression across breast cancer subtypes. HER3 expression can be assessed early, post-neoadjuvant therapy, providing valuable insight into cancer biology and potentially serving as a prognostic biomarker. Clinical translation of neoadjuvant therapy assessment can be achieved using HER3 PET imaging, offering real-time information on tumor biology and guiding personalized treatment for breast cancer patients.

Feasibility of risk assessment for breast cancer molecular subtypes.

PURPOSE: Few breast cancer risk assessment models account for the risk profiles of different tumor subtypes. This study evaluated whether a subtype-specific approach improves discrimination. METHODS: Among 3389 women who had a screening mammogram and were later diagnosed with invasive breast cancer we performed multinomial logistic regression with tumor subtype as the outcome and known breast cancer risk factors as predictors. Tumor subtypes were defined by expression of estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2) based on immunohistochemistry. Discrimination was assessed with the area under the receiver operating curve (AUC). Absolute risk of each subtype was estimated by proportioning Gail absolute risk estimates by the predicted probabilities for each subtype. We then compared risk factor distributions for women in the highest deciles of risk for each subtype. RESULTS: There were 3,073 ER/PR+ HER2 - , 340 ER/PR +HER2 + , 126 ER/PR-ER2+, and 300 triple-negative breast cancers (TNBC). Discrimination differed by subtype; ER/PR-HER2+ (AUC: 0.64, 95% CI 0.59, 0.69) and TNBC (AUC: 0.64, 95% CI 0.61, 0.68) had better discrimination than ER/PR+HER2+ (AUC: 0.61, 95% CI 0.58, 0.64). Compared to other subtypes, patients at high absolute risk of TNBC were younger, mostly Black, had no family history of breast cancer, and higher BMI. Those at high absolute risk of HER2+ cancers were younger and had lower BMI. CONCLUSION: Our study provides proof of concept that stratifying risk prediction for breast cancer subtypes may enable identification of patients with unique profiles conferring increased risk for tumor subtypes.

Missed Screening Mammography Appointments: Patient Sociodemographic Characteristics and Mammography Completion After 1 Year.

OBJECTIVE: Patients who miss screening mammogram appointments without notifying the health care system (no-show) risk care delays. We investigate sociodemographic characteristics of patients who experience screening mammogram no-shows at a community health center and whether and when the missed examinations are completed. METHODS: We included patients with screening mammogram appointments at a community health center between January 1, 2021, and December 31, 2021. Language, race, ethnicity, insurance type, residential ZIP code tabulation area (ZCTA) poverty, appointment outcome (no-show, same-day cancelation, completed), and dates of completed screening mammograms after no-show appointments with ≥1-year follow-up were collected. Multivariable analyses were used to assess associations between patient characteristics and appointment outcomes. RESULTS: Of 6,159 patients, 12.1% (743 of 6,159) experienced no-shows. The no-show group differed from the completed group by language, race and ethnicity, insurance type, and poverty level (all P < .05). Patients with no-shows more often had: primary language other than English (32.0% [238 of 743] versus 26.7% [1,265 of 4,741]), race and ethnicity other than White non-Hispanic (42.3% [314 of 743] versus 33.6% [1,595 of 4,742]), Medicaid or means-tested insurance (62.0% [461 of 743] versus 34.4% [1,629 of 4,742]), and residential ZCTAs with ≥20% poverty (19.5% [145 of 743] versus 14.1% [670 of 4,742]). Independent predictors of no-shows were Black non-Hispanic race and ethnicity (adjusted odds ratio [aOR], 1.52; 95% confidence interval [CI], 1.12-2.07; P = .007), Medicaid or other means-tested insurance (aOR, 2.75; 95% CI, 2.29-3.30; P < .001), and ZCTAs with ≥20% poverty (aOR, 1.76; 95% CI, 1.14-2.72; P = .011). At 1-year follow-up, 40.6% (302 of 743) of patients with no-shows had not completed screening mammogram. DISCUSSION: Screening mammogram no-shows is a health equity issue in which socio-economically disadvantaged and racially and ethnically minoritized patients are more likely to experience missed appointments and continued delays in screening mammogram completion.

Incidence, Timing, and Long-Term Outcomes of COVID-19 Vaccine-Related Lymphadenopathy on Screening Mammography.

OBJECTIVE: To determine the incidence, timing, and long-term outcomes of unilateral axillary lymphadenopathy ipsilateral to vaccine site (UIAL) on screening mammography after COVID-19 vaccination. METHODS: This retrospective, multisite study included consecutive patients undergoing screening mammography February 8, 2021, to January 31, 2022, with at least 1 year of follow-up. UIAL was typically considered benign (BI-RADS 1 or 2) in the setting of recent (≤6 weeks) vaccination or BI-RADS 0 (ultrasound recommended) when accompanied by a breast finding or identified >6 weeks postvaccination. Vaccination status and manufacturer were obtained from regional registries. Lymphadenopathy rates in vaccinated patients with and without UIAL were compared using Pearson's χ2 test. RESULTS: There were 44,473 female patients (mean age 60.4 ± 11.4 years) who underwent screening mammography at five sites, and 40,029 (90.0%) received at least one vaccine dose. Ninety-four (0.2%) presented with UIAL, 1 to 191 days postvaccination (median 13.5 [interquartile range: 5.0-31.0]). Incidence declined from 2.1% to 0.9% to ≤0.5% after 1, 2, and 3 weeks and persisted up to 36 weeks (P < .001). UIAL did not vary across manufacturer (P = .15). Of 94, 77 (81.9%) were BI-RADS 1 or 2 at screening. None were diagnosed with malignancy at 1-year follow-up. Seventeen (18.1%) were BI-RADS 0 at screening. At diagnostic workup, 13 (76.5%) were BI-RADS 1 or 2, 2 (11.8%) were BI-RADS 3, and 2 (11.8%) were BI-RADS 4. Both BI-RADS 4 patients had malignant status and ipsilateral breast malignancies. Of BI-RADS 3 patients, at follow-up, one was biopsied yielding benign etiology, and one was downgraded to BI-RADS 2. DISCUSSION: Isolated UIAL on screening mammography performed within 6 months of COVID-19 vaccination can be safely assessed as benign.

Comparison of the Diagnostic Accuracy of Mammogram-based Deep Learning and Traditional Breast Cancer Risk Models in Patients Who Underwent Supplemental Screening with MRI.

Background Access to supplemental screening breast MRI is determined using traditional risk models, which are limited by modest predictive accuracy. Purpose To compare the diagnostic accuracy of a mammogram-based deep learning (DL) risk assessment model to that of traditional breast cancer risk models in patients who underwent supplemental screening with MRI. Materials and Methods This retrospective study included consecutive patients undergoing breast cancer screening MRI from September 2017 to September 2020 at four facilities. Risk was assessed using the Tyrer-Cuzick (TC) and National Cancer Institute Breast Cancer Risk Assessment Tool (BCRAT) 5-year and lifetime models as well as a DL 5-year model that generated a risk score based on the most recent screening mammogram. A risk score of 1.67% or higher defined increased risk for traditional 5-year models, a risk score of 20% or higher defined high risk for traditional lifetime models, and absolute scores of 2.3 or higher and 6.6 or higher defined increased and high risk, respectively, for the DL model. Model accuracy metrics including cancer detection rate (CDR) and positive predictive values (PPVs) (PPV of abnormal findings at screening [PPV1], PPV of biopsies recommended [PPV2], and PPV of biopsies performed [PPV3]) were compared using logistic regression models. Results This study included 2168 women who underwent 4247 high-risk screening MRI examinations (median age, 54 years [IQR, 48-60 years]). CDR (per 1000 examinations) was higher in patients at high risk according to the DL model (20.6 [95% CI: 11.8, 35.6]) than according to the TC (6.0 [95% CI: 2.9, 12.3]; P < .01) and BCRAT (6.8 [95% CI: 2.9, 15.8]; P = .04) lifetime models. PPV1, PPV2, and PPV3 were higher in patients identified as high risk by the DL model (PPV1, 14.6%; PPV2, 32.4%; PPV3, 36.4%) than those identified as high risk with the TC (PPV1, 5.0%; PPV2, 12.7%; PPV3, 13.5%; P value range, .02-.03) and BCRAT (PPV1, 5.5%; PPV2, 11.1%; PPV3, 12.5%; P value range, .02-.05) lifetime models. Conclusion Patients identified as high risk by a mammogram-based DL risk assessment model showed higher CDR at breast screening MRI than patients identified as high risk with traditional risk models. © RSNA, 2023 Supplemental material is available for this article. See also the editorial by Bae in this issue.

Standalone AI for Breast Cancer Detection at Screening Digital Mammography and Digital Breast Tomosynthesis: A Systematic Review and Meta-Analysis.

Background There is considerable interest in the potential use of artificial intelligence (AI) systems in mammographic screening. However, it is essential to critically evaluate the performance of AI before it can become a modality used for independent mammographic interpretation. Purpose To evaluate the reported standalone performances of AI for interpretation of digital mammography and digital breast tomosynthesis (DBT). Materials and Methods A systematic search was conducted in PubMed, Google Scholar, Embase (Ovid), and Web of Science databases for studies published from January 2017 to June 2022. Sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) values were reviewed. Study quality was assessed using the Quality Assessment of Diagnostic Accuracy Studies 2 and Comparative (QUADAS-2 and QUADAS-C, respectively). A random effects meta-analysis and meta-regression analysis were performed for overall studies and for different study types (reader studies vs historic cohort studies) and imaging techniques (digital mammography vs DBT). Results In total, 16 studies that include 1 108 328 examinations in 497 091 women were analyzed (six reader studies, seven historic cohort studies on digital mammography, and four studies on DBT). Pooled AUCs were significantly higher for standalone AI than radiologists in the six reader studies on digital mammography (0.87 vs 0.81, P = .002), but not for historic cohort studies (0.89 vs 0.96, P = .152). Four studies on DBT showed significantly higher AUCs in AI compared with radiologists (0.90 vs 0.79, P < .001). Higher sensitivity and lower specificity were seen for standalone AI compared with radiologists. Conclusion Standalone AI for screening digital mammography performed as well as or better than radiologists. Compared with digital mammography, there is an insufficient number of studies to assess the performance of AI systems in the interpretation of DBT screening examinations. © RSNA, 2023 Supplemental material is available for this article. See also the editorial by Scaranelo in this issue.

Breast cancer polygenic risk scores are associated with short-term risk of poor prognosis breast cancer.

PURPOSE: Polygenic risk scores (PRS) for breast cancer may help guide screening decisions. However, few studies have examined whether PRS are associated with risk of short-term or poor prognosis breast cancers. The study purpose was to evaluate the association of the 313 SNP breast cancer PRS with 2-year risk of poor prognosis breast cancer. METHODS: We evaluated the association of breast cancer PRS with breast cancer overall, ER + and ER- breast cancer, and poor prognosis breast cancer diagnosed within 2 years of a negative mammogram among a cohort of 3657 women using logistic regression adjusted for age, breast density, race/ethnicity, year of screening, and genetic ancestry principal components. Breast cancers were considered poor prognosis if they were metastatic, positive lymph nodes, ER/PR + HER2- and > 2 cm, ER/PR/HER2-, or HER2 + and > 1 cm. RESULTS: Of the 308 breast cancers, 137 (44%) were poor prognosis. The overall breast cancer PRS was significantly associated with breast cancer diagnosis within 2 years (OR 1.39, 95% CI 1.23-1.57, p < 0.001). The breast cancer PRS was also associated specifically with diagnosis of poor prognosis disease (OR 1.24, 95% CI 1.03-1.49, p = 0.018), but was more strongly associated with good prognosis cancer (OR 1.52 95% CI 1.29-1.80 p = 3.60 × 10-7) The ER + PRS was significantly associated with ER/PR + breast cancer (OR 1.41, 95% CI 1.24-1.61, p < 0.001) and the ER- PRS was significantly associated with ER- breast cancer (OR 1.48, 95% CI 1.08-2.02, p = 0.015). CONCLUSION: Breast cancer PRS was independently and significantly associated with diagnosis of both breast cancer overall and poor prognosis breast cancer within 2 years of a negative mammogram, suggesting PRS may help guide decisions about screening intervals and supplemental screening.

Stacked dilated convolutions and asymmetric architecture for U-Net-based medical image segmentation.

Deep learning has been widely utilized for medical image segmentation. The most commonly used U-Net and its variants often share two common characteristics but lack solid evidence for the effectiveness. First, each block (i.e., consecutive convolutions of feature maps of the same resolution) outputs feature maps from the last convolution, limiting the variety of the receptive fields. Second, the network has a symmetric structure where the encoder and the decoder paths have similar numbers of channels. We explored two novel revisions: a stacked dilated operation that outputs feature maps from multi-scale receptive fields to replace the consecutive convolutions; an asymmetric architecture with fewer channels in the decoder path. Two novel models were developed: U-Net using the stacked dilated operation (SDU-Net) and asymmetric SDU-Net (ASDU-Net). We used both publicly available and private datasets to assess the efficacy of the proposed models. Extensive experiments confirmed SDU-Net outperformed or achieved performance similar to the state-of-the-art while using fewer parameters (40% of U-Net). ASDU-Net further reduced the model parameters to 20% of U-Net with performance comparable to SDU-Net. In conclusion, the stacked dilated operation and the asymmetric structure are promising for improving the performance of U-Net and its variants.

Deep Learning vs Traditional Breast Cancer Risk Models to Support Risk-Based Mammography Screening.

BACKGROUND: Deep learning breast cancer risk models demonstrate improved accuracy compared with traditional risk models but have not been prospectively tested. We compared the accuracy of a deep learning risk score derived from the patient's prior mammogram to traditional risk scores to prospectively identify patients with cancer in a cohort due for screening. METHODS: We collected data on 119 139 bilateral screening mammograms in 57 617 consecutive patients screened at 5 facilities between September 18, 2017, and February 1, 2021. Patient demographics were retrieved from electronic medical records, cancer outcomes determined through regional tumor registry linkage, and comparisons made across risk models using Wilcoxon and Pearson χ2 2-sided tests. Deep learning, Tyrer-Cuzick, and National Cancer Institute Breast Cancer Risk Assessment Tool (NCI BCRAT) risk models were compared with respect to performance metrics and area under the receiver operating characteristic curves. RESULTS: Cancers detected per thousand patients screened were higher in patients at increased risk by the deep learning model (8.6, 95% confidence interval [CI] = 7.9 to 9.4) compared with Tyrer-Cuzick (4.4, 95% CI = 3.9 to 4.9) and NCI BCRAT (3.8, 95% CI = 3.3 to 4.3) models (P < .001). Area under the receiver operating characteristic curves of the deep learning model (0.68, 95% CI = 0.66 to 0.70) was higher compared with Tyrer-Cuzick (0.57, 95% CI = 0.54 to 0.60) and NCI BCRAT (0.57, 95% CI = 0.54 to 0.60) models. Simulated screening of the top 50th percentile risk by the deep learning model captured statistically significantly more patients with cancer compared with Tyrer-Cuzick and NCI BCRAT models (P < .001). CONCLUSIONS: A deep learning model to assess breast cancer risk can support feasible and effective risk-based screening and is superior to traditional models to identify patients destined to develop cancer in large screening cohorts.

Impact of a Deep Learning Model for Predicting Mammographic Breast Density in Routine Clinical Practice.

OBJECTIVE: Legislation in 38 states requires patient notification of dense mammographic breast tissue because increased density is a marker of breast cancer risk and can limit mammographic sensitivity. Because radiologist density assessments vary widely, our objective was to implement and measure the impact of a deep learning (DL) model on mammographic breast density assessments in clinical practice. METHODS: This institutional review board-approved prospective study identified consecutive screening mammograms performed across three clinical sites over two periods: 2017 period (January 1, 2017, through September 30, 2017) and 2019 period (January 1, 2019, through September 30, 2019). The DL model was implemented at sites A (academic practice) and B (community practice) in 2018 for all screening mammograms. Site C (community practice) was never exposed to the DL model. Prospective densities were evaluated, and multivariable logistic regression models evaluated the odds of a dense mammogram classification as a function of time and site. RESULTS: We identified 85,124 consecutive screening mammograms across the three sites. Across time intervals, odds of a dense classification decreased at sites exposed to the DL model, site A (adjusted odds ratio [aOR], 0.93; 95% confidence interval [CI], 0.86-0.99; P = .024) and site B (aOR, 0.81 [95% CI, 0.70-0.93]; P = .003), and odds increased at the site unexposed to the model (site C) (aOR, 1.13 [95% CI, 1.01-1.27]; P = .033). DISCUSSION: A DL model reduces the odds of screening mammograms categorized as dense. Accurate density assessments could help health care systems more appropriately use limited supplemental screening resources and help better inform traditional clinical risk models.

Association Between Surgery Preference and Receipt in Ductal Carcinoma In Situ After Breast Magnetic Resonance Imaging: An Ancillary Study of the ECOG-ACRIN Cancer Research Group (E4112).

Importance: Guiding treatment decisions for women with ductal carcinoma in situ (DCIS) requires understanding patient preferences and the influence of preoperative magnetic resonance imaging (MRI) and surgeon recommendation. Objective: To identify factors associated with surgery preference and surgery receipt among a prospective cohort of women with newly diagnosed DCIS. Design, Setting, and Participants: A prospective cohort study was conducted at 75 participating institutions, including community practices and academic centers, across the US between March 25, 2015, and April 27, 2016. Data were analyzed from August 2 to September 24, 2021. This was an ancillary study of the ECOG-ACRIN Cancer Research Group (E4112). Women with recently diagnosed unilateral DCIS who were eligible for wide local excision and had a diagnostic mammogram within 3 months of study registration were included. Participants who had documented surgery and completed the baseline patient-reported outcome questionnaires were included in this substudy. Exposures: Women received preoperative MRI and surgeon consultation and then underwent wide local excision or mastectomy. Participants will be followed up for recurrence and overall survival for 10 years from the date of surgery. Main Outcomes and Measures: Patient-reported outcome questionnaires assessed treatment goals and concerns and surgery preference before MRI and after MRI and surgeon consultation. Results: Of the 368 participants enrolled 316 (86%) were included in this substudy (median [range] age, 59.5 [34-87] years; 45 women [14%] were Black; 245 [78%] were White; and 26 [8%] were of other race). Pre-MRI, age (odds ratio [OR] per 5-year increment, 0.45; 95% CI, 0.26-0.80; P = .007) and the importance of keeping one's breast (OR, 0.48; 95% CI, 0.31-0.72; P < .001) vs removal of the breast for peace of mind (OR, 1.35; 95% CI, 1.04-1.76; P = .03) were associated with surgery preference for mastectomy. After MRI and surgeon consultation, MRI upstaging (48 of 316 [15%]) was associated with patient preference for mastectomy (OR, 8.09; 95% CI, 2.51-26.06; P < .001). The 2 variables with the highest ORs for initial receipt of mastectomy were MRI upstaging (OR, 12.08; 95% CI, 4.34-33.61; P < .001) and surgeon recommendation (OR, 4.85; 95% CI, 1.99-11.83; P < .001). Conclusions and Relevance: In this cohort study, change in patient preference for DCIS surgery and surgery received were responsive to MRI results and surgeon recommendation. These data highlight the importance of ensuring adequate information and ongoing communication about the clinical significance of MRI findings and the benefits and risks of available treatment options.

Magnetic Seeds: An Alternative to Wire Localization for Nonpalpable Breast Lesions.

INTRODUCTION: Magnetic seeds have emerged as an alternative to wires for localization of nonpalpable breast lesions. The purpose of this study was to evaluate the utility of magnetic seeds compared to wires for preoperative localization. MATERIALS AND METHODS: A retrospective cohort analysis of magnetic seed localization (MSL) and wire localization (WL) excisional biopsies and lumpectomies performed at a single institution was conducted. Indication, age, BMI, number of markers, procedure type, operative time, and postoperative opioid administration were reviewed. Impact of localization method on operative time, specimen volume, postoperative opioid administration, and re-excision rate were assessed. RESULTS: A total of 608 MSL procedures in 601 patients were compared to 628 WL procedures in 620 patients. MSL excisional biopsies were significantly longer (37.0 minutes) than WL excisional biopsies (31.9 minutes, P< .001), but in lumpectomies without axillary surgery, MSL procedures (42.3 minutes) were significantly shorter than WL procedures (46.9 minutes, P = .017). Significantly less tissue was excised during MSL lumpectomies (68.5 cm3) and excisional biopsies (32.3 cm3) than WL lumpectomies (78.1 cm3, P = .039) and excisional biopsies (38.7 cm3, P = .018). Postoperative opioid administration was similar for MSL and WL procedures (P = .076). Re-excision rates for MSL lumpectomies were significantly higher for ductal carcinoma in situ (35.3% MSL vs. 18.5% WL, P = .013), but were similar for invasive carcinoma (14.4% MSL vs. 17.7% WL, P = .290). Logistic regression analysis showed no association between localization method and re-excision (OR 1.007, 95% CI 0.681-1.488; P = .973). CONCLUSION: MSL is a feasible alternative to WL for excision of nonpalpable breast lesions with regard to surgical outcomes.

Artificial Intelligence (AI) for Screening Mammography, From the AJR Special Series on AI Applications.

Artificial intelligence (AI) applications for screening mammography are being marketed for clinical use in the interpretative domains of lesion detection and diagnosis, triage, and breast density assessment and in the noninterpretive domains of breast cancer risk assessment, image quality control, image acquisition, and dose reduction. Evidence in support of these nascent applications, particularly for lesion detection and diagnosis, is largely based on multireader studies with cancer-enriched datasets rather than rigorous clinical evaluation aligned with the application's specific intended clinical use. This article reviews commercial AI algorithms for screening mammography that are currently available for clinical practice, their use, and evidence supporting their performance. Clinical implementation considerations, such as workflow integration, governance, and ethical issues, are also described. In addition, the future of AI for screening mammography is discussed, including the development of interpretive and noninterpretive AI applications and strategic priorities for research and development.

Impact of a Same-Day Breast Biopsy Program on Disparities in Time to Biopsy for Patients With Serious Mental Illness.

PURPOSE: The aim of this study was to investigate disparities in time between breast biopsy recommendation and completion and the impact of a same-day biopsy (SDB) program for patients with serious mental illness (SMI), with a focus on more vulnerable individuals with public payer insurance. METHODS: In August 2017, the authors' academic breast imaging center started routinely offering needle biopsies on the day of recommendation. Primary outcomes were the proportion of biopsies performed as SDBs and days from biopsy recommendation to completion over a 2.5-year pre- versus postintervention period, comparing all patients with SMI versus those without, and public payer-insured patients <65 years of age with SMI (SMI-PP) versus without SMI (non-SMI-PP). Multivariable proportional odds and logistic regression models were fit to assess association of SMI status, age, race/ethnicity, language, and insurance with days to biopsy and SDB within each period. RESULTS: There were 2,026 biopsies preintervention and 2,361 biopsies postintervention. Preintervention, 8.43% of patients with SMI (7 of 83) underwent SDB compared with 15.59% of those without SMI (303 of 1,943) (P = .076), and 2.7% of the SMI-PP subgroup (1 of 37) underwent SDB compared with 15.88% of the non-SMI-PP subgroup (47 of 296) (P = .031). Adjusted for age, race/ethnicity, and language, disparities persisted in odds for undergoing SDB (adjusted odds ratio, 0.13; 95% confidence interval, 0.02-0.92; P = .04) and having longer days to biopsy (adjusted odds ratio, 2.35; 95% confidence interval, 1.26-4.37; P = .01) for the SMI-PP subgroup compared with the non-SMI-PP subgroup in the preintervention period. There was no evidence of these disparities postintervention for the SMI-PP subgroup. SDB proportion increased from 15.3% (310 of 2,026) to 36.09% (852 of 2,361) (P < .001) across all patients. CONCLUSIONS: A same-day breast biopsy program mitigates disparities in time to biopsy for patients with SMI and helps improve breast cancer care equity for this vulnerable population.

Optimizing risk-based breast cancer screening policies with reinforcement learning.

Screening programs must balance the benefit of early detection with the cost of overscreening. Here, we introduce a novel reinforcement learning-based framework for personalized screening, Tempo, and demonstrate its efficacy in the context of breast cancer. We trained our risk-based screening policies on a large screening mammography dataset from Massachusetts General Hospital (MGH; USA) and validated this dataset in held-out patients from MGH and external datasets from Emory University (Emory; USA), Karolinska Institute (Karolinska; Sweden) and Chang Gung Memorial Hospital (CGMH; Taiwan). Across all test sets, we find that the Tempo policy combined with an image-based artificial intelligence (AI) risk model is significantly more efficient than current regimens used in clinical practice in terms of simulated early detection per screen frequency. Moreover, we show that the same Tempo policy can be easily adapted to a wide range of possible screening preferences, allowing clinicians to select their desired trade-off between early detection and screening costs without training new policies. Finally, we demonstrate that Tempo policies based on AI-based risk models outperform Tempo policies based on less accurate clinical risk models. Altogether, our results show that pairing AI-based risk models with agile AI-designed screening policies has the potential to improve screening programs by advancing early detection while reducing overscreening.

AUR-RRA Review: Logistics of Academic-Industry Partnerships in Artificial Intelligence.

The Radiology Research Alliance (RRA) of the Association of University Radiologists (AUR) convenes Task Forces to address current topics in radiology. In this article, the AUR-RRA Task Force on Academic-Industry Partnerships for Artificial Intelligence, considered issues of importance to academic radiology departments contemplating industry partnerships in artificial intelligence (AI) development, testing and evaluation. Our goal was to create a framework encompassing the domains of clinical, technical, regulatory, legal and financial considerations that impact the arrangement and success of such partnerships.

Power Spectrum Analysis of Breast Parenchyma with Digital Breast Tomosynthesis Images in a Longitudinal Screening Cohort from Two Vendors.

RATIONALE AND OBJECTIVES: To quantitatively compare breast parenchymal texture between two Digital Breast Tomosynthesis (DBT) vendors using images from the same patients. MATERIALS AND METHODS: This retrospective study included consecutive patients who had normal screening DBT exams performed in January 2018 from GE and normal screening DBT exams in adjacent years from Hologic. Power spectrum analysis was performed within the breast tissue region. The slope of a linear function between log-frequency and log-power, ß, was derived as a quantitative measure of breast texture and compared within and across vendors along with secondary parameters (laterality, view, year, image format, and breast density) with correlation tests and t-tests. RESULTS: A total of 24,339 DBT slices or synthetic 2D images from 85 exams in 25 women were analyzed. Strong power-law behavior was verified from all images. Values of ß d did not differ significantly for laterality, view, or year. Significant differences of ß were observed across vendors for DBT images (Hologic: 3.4±0.2 vs GE: 3.1±0.2, 95% CI on difference: 0.27 to 0.30) and synthetic 2D images (Hologic: 2.7±0.3 vs GE: 3.0±0.2, 95% CI on difference: -0.36 to -0.27), and density groups with each vendor: scattered (GE: 3.0±0.3, Hologic: 3.3±0.3) vs. heterogeneous (GE: 3.2±0.2, Hologic: 3.4±0.1), 95% CI (-0.27, -0.08) and (-0.21, -0.05), respectively. CONCLUSION: There are quantitative differences in the presentation of breast imaging texture between DBT vendors and across breast density categories. Our findings have relevance and importance for development and optimization of AI algorithms related to breast density assessment and cancer detection.

Disparities in Same-Day Diagnostic Imaging in Breast Cancer Screening: Impact of an Immediate-Read Screening Mammography Program Implemented During the COVID-19 Pandemic.

BACKGROUND. The need for second visits between screening mammography and diagnostic imaging contributes to disparities in the time to breast cancer diagnosis. During the COVID-19 pandemic, an immediate-read screening mammography program was implemented to reduce patient visits and decrease time to diagnostic imaging. OBJECTIVE. The purpose of this study was to measure the impact of an immediate-read screening program with focus on disparities in same-day diagnostic imaging after abnormal findings are made at screening mammography. METHODS. In May 2020, an immediate-read screening program was implemented whereby a dedicated breast imaging radiologist interpreted all screening mammograms in real time; patients received results before discharge; and efforts were made to perform any recommended diagnostic imaging during the visit (performed by different radiologists). Screening mammographic examinations performed from June 1, 2019, through October 31, 2019 (preimplementation period), and from June 1, 2020, through October 31, 2020 (postimplementation period), were retrospectively identified. Patient characteristics were recorded from the electronic medical record. Multivariable logistic regression models incorporating patient age, race and ethnicity, language, and insurance type were estimated to identify factors associated with same-day diagnostic imaging. Screening metrics were compared between periods. RESULTS. A total of 8222 preimplementation and 7235 postimplementation screening examinations were included; 521 patients had abnormal screening findings before implementation, and 359 after implementation. Before implementation, 14.8% of patients underwent same-day diagnostic imaging after abnormal screening mammograms. This percentage increased to 60.7% after implementation. Before implementation, patients who identified their race as other than White had significantly lower odds than patients who identified their race as White of undergoing same-day diagnostic imaging after receiving abnormal screening results (adjusted odds ratio, 0.30; 95% CI, 0.10-0.86; p = .03). After implementation, the odds of same-day diagnostic imaging were not significantly different between patients of other races and White patients (adjusted odds ratio, 0.92; 95% CI, 0.50-1.71; p = .80). After implementation, there was no significant difference in race and ethnicity between patients who underwent and those who did not undergo same-day diagnostic imaging after receiving abnormal results of screening mammography (p > .05). The rate of abnormal interpretation was significantly lower after than it was before implementation (5.0% vs 6.3%; p < .001). Cancer detection rate and PPV1 (PPV based on positive findings at screening examination) were not significantly different before and after implementation (p > .05). CONCLUSION. Implementation of the immediate-read screening mammography program reduced prior racial and ethnic disparities in same-day diagnostic imaging after abnormal screening mammograms. CLINICAL IMPACT. An immediate-read screening program provides a new paradigm for improved screening mammography workflow that allows more rapid diagnostic workup with reduced disparities in care.

Multi-Institutional Validation of a Mammography-Based Breast Cancer Risk Model.

PURPOSE: Accurate risk assessment is essential for the success of population screening programs in breast cancer. Models with high sensitivity and specificity would enable programs to target more elaborate screening efforts to high-risk populations, while minimizing overtreatment for the rest. Artificial intelligence (AI)-based risk models have demonstrated a significant advance over risk models used today in clinical practice. However, the responsible deployment of novel AI requires careful validation across diverse populations. To this end, we validate our AI-based model, Mirai, across globally diverse screening populations. METHODS: We collected screening mammograms and pathology-confirmed breast cancer outcomes from Massachusetts General Hospital, USA; Novant, USA; Emory, USA; Maccabi-Assuta, Israel; Karolinska, Sweden; Chang Gung Memorial Hospital, Taiwan; and Barretos, Brazil. We evaluated Uno's concordance index for Mirai in predicting risk of breast cancer at one to five years from the mammogram. RESULTS: A total of 128,793 mammograms from 62,185 patients were collected across the seven sites, of which 3,815 were followed by a cancer diagnosis within 5 years. Mirai obtained concordance indices of 0.75 (95% CI, 0.72 to 0.78), 0.75 (95% CI, 0.70 to 0.80), 0.77 (95% CI, 0.75 to 0.79), 0.77 (95% CI, 0.73 to 0.81), 0.81 (95% CI, 0.79 to 0.82), 0.79 (95% CI, 0.76 to 0.83), and 0.84 (95% CI, 0.81 to 0.88) at Massachusetts General Hospital, Novant, Emory, Maccabi-Assuta, Karolinska, Chang Gung Memorial Hospital, and Barretos, respectively. CONCLUSION: Mirai, a mammography-based risk model, maintained its accuracy across globally diverse test sets from seven hospitals across five countries. This is the broadest validation to date of an AI-based breast cancer model and suggests that the technology can offer broad and equitable improvements in care.