Machine Learning Prediction of Lymph Node Metastasis in Breast Cancer: Performance of a Multi-institutional MRI-based 4D Convolutional Neural Network.

Purpose To develop a custom deep convolutional neural network (CNN) for noninvasive prediction of breast cancer nodal metastasis. Materials and Methods This retrospective study included patients with newly diagnosed primary invasive breast cancer with known pathologic (pN) and clinical nodal (cN) status who underwent dynamic contrast-enhanced (DCE) breast MRI at the authors' institution between July 2013 and July 2016. Clinicopathologic data (age, estrogen receptor and human epidermal growth factor 2 status, Ki-67 index, and tumor grade) and cN and pN status were collected. A four-dimensional (4D) CNN model integrating temporal information from dynamic image sets was developed. The convolutional layers learned prognostic image features, which were combined with clinicopathologic measures to predict cN0 versus cN+ and pN0 versus pN+ disease. Performance was assessed with the area under the receiver operating characteristic curve (AUC), with fivefold nested cross-validation. Results Data from 350 female patients (mean age, 51.7 years ± 11.9 [SD]) were analyzed. AUC, sensitivity, and specificity values of the 4D hybrid model were 0.87 (95% CI: 0.83, 0.91), 89% (95% CI: 79%, 93%), and 76% (95% CI: 68%, 88%) for differentiating pN0 versus pN+ and 0.79 (95% CI: 0.76, 0.82), 80% (95% CI: 77%, 84%), and 62% (95% CI: 58%, 67%), respectively, for differentiating cN0 versus cN+. Conclusion The proposed deep learning model using tumor DCE MR images demonstrated high sensitivity in identifying breast cancer lymph node metastasis and shows promise for potential use as a clinical decision support tool. Keywords: MR Imaging, Breast, Breast Cancer, Breast MRI, Machine Learning, Metastasis, Prognostic Prediction Supplemental material is available for this article. Published under a CC BY 4.0 license.

Radiomics Analysis of Contrast-Enhanced Breast MRI for Optimized Modelling of Virtual Prognostic Biomarkers in Breast Cancer.

Objective: Breast cancer clinical stage and nodal status are the most clinically significant drivers of patient management, in combination with other pathological biomarkers, such as estrogen receptor (ER), progesterone receptor or human epidermal growth factor receptor 2 (HER2) receptor status and tumor grade. Accurate prediction of such parameters can help avoid unnecessary intervention, including unnecessary surgery. The objective was to investigate the role of magnetic resonance imaging (MRI) radiomics for yielding virtual prognostic biomarkers (ER, HER2 expression, tumor grade, molecular subtype, and T-stage). Materials and Methods: Patients with primary invasive breast cancer who underwent dynamic contrast-enhanced (DCE) breast MRI between July 2013 and July 2016 in a single center were retrospectively reviewed. Age, N-stage, grade, ER and HER2 status, and Ki-67 (%) were recorded. DCE images were segmented and Haralick texture features were extracted. The Bootstrap Lasso feature selection method was used to select a small subset of optimal texture features. Classification of the performance of the final model was assessed with the area under the receiver operating characteristic curve (AUC). Results: Median age of patients (n = 209) was 49 (21-79) years. Sensitivity, specificity, positive predictive value, negative predictive value and accuracy of the model for differentiating N0 vs N1-N3 was: 71%, 79%, 76%, 74%, 75% [AUC = 0.78 (95% confidence interval (CI) 0.72-0.85)], N0-N1 vs N2-N3 was 81%, 59%, 24%, 95%, 62% [AUC = 0.74 (95% CI 0.63-0.85)], distinguishing HER2(+) from HER2(-) was 79%, 48%, 34%, 87%, 56% [AUC = 0.64 (95% CI 0.54-0.73)], high nuclear grade (grade 2-3) vs low grade (grades 1) was 56%, 88%, 96%, 29%, 61% [AUC = 0.71 (95% CI 0.63-0.80)]; and for ER (+) vs ER(-) status the [AUC=0.67 (95% CI 0.59-0.76)]. Radiomics performance in distinguishing triple-negative vs other molecular subtypes was [0.60 (95% CI 0.49-0.71)], and Luminal A [0.66 (95% CI 0.56-0.76)]. Conclusion: Quantitative radiomics using MRI contrast texture shows promise in identifying aggressive high grade, node positive triple negative breast cancer, and correlated well with higher nuclear grades, higher T-stages, and N-positive stages.

A Tale of Two Hospitals: Effect of Access to Care Through a Safety Net Hospital on Adjuvant Therapy, Imaging Compliance and 5-Year Survival Rates Compared to the University Hospital Served by the Same Breast Cancer Clinical Teams.

RATIONALE AND OBJECTIVES: To compare rates of guideline-concordant care, imaging surveillance, recurrence and survival outcomes between a safety-net (SNH) and tertiary-care University Hospital (UH) served by the same breast cancer clinical teams. MATERIALS AND METHODS: 647 women with newly diagnosed breast cancer treated in affiliated SNH and UH between 11.1.2014 and 3.31.2017 were reviewed. Patient demographics, completion of guideline-concordant adjuvant chemotherapy, radiation and hormonal therapy were recorded. Two multivariable logistic regression models were performed to investigate the effect of hospital and race on cancer stage. Kaplan-Meier log-rank and Cox-regression were used to analyze five-year recurrence-free (RFS) and overall survival (OS) between hospitals and races, (p < 0.05 significant). RESULTS: Patients in SNH were younger (mean SNH 53.2 vs UH 57.9, p < 0.001) and had higher rates of cT3/T4 disease (SNH 19% vs UH 5.5%, p < 0.001). Patients in the UH had higher rates of bilateral mastectomy (SNH 17.6% vs UH 40.1% p < 0.001) while there was no difference in the positive surgical margin rate (SNH 5.0% vs UH 7.6%, p = 0.20), completion of adjuvant radiation (SNH 96.9% vs UH 98.7%, p = 0.2) and endocrine therapy (SNH 60.8% vs UH 66.2%, p = 0.20). SNH patients were less compliant with mammography surveillance (SNH 64.1% vs UH 75.1%, p = 0.02) and adjuvant chemotherapy (SNH 79.1% vs UH 96.3%, p < 0.01). RFS was lower in the SNH (SNH 54 months vs UH 57 months, HR 1.90, 95% CI: 1.18-3.94, p = 0.01) while OS was not significantly different (SNH 90.5% vs UH 94.2%, HR 1.78, 95% CI: 0.97-3.26, p = 0.06). CONCLUSION: In patients experiencing health care disparities, having access to guideline-concordant care through SNH resulted in non-inferior OS to those in tertiary-care UH.

Imaging Challenges in Diagnosing Triple-Negative Breast Cancer.

Triple-negative breast cancer (TNBC) refers to a heterogeneous group of carcinomas that have more aggressive biologic features, faster growth, and a propensity for early distant metastasis and recurrence compared with other breast cancer subtypes. Due to the aggressiveness and rapid growth of TNBCs, there are specific imaging challenges associated with their timely and accurate diagnosis. TNBCs commonly manifest initially as circumscribed masses and therefore lack the typical features of a primary breast malignancy, such as irregular shape, spiculated margins, and desmoplastic reaction. Given the potential for misinterpretation, review of the multimodality imaging appearances of TNBCs is important for guiding the radiologist in distinguishing TNBCs from benign conditions. Rather than manifesting as a screening-detected cancer, TNBC typically appears clinically as a palpable area of concern that most commonly corresponds to a discrete mass at mammography, US, and MRI. The combination of circumscribed margins and hypoechoic to anechoic echogenicity may lead to TNBC being misinterpreted as a benign fibroadenoma or cyst. Therefore, careful mammographic and sonographic evaluation with US image optimization can help avoid misinterpretation. Radiologists should recognize the characteristics of TNBCs that can mimic benign entities, as well as the subtle features of TNBCs that should raise concern for malignancy and aid in timely and accurate diagnosis. ©RSNA, 2023 Quiz questions for this article are available in the supplemental material.

Radial Scars Without Atypia Diagnosed at Percutaneous Core Needle Breast Biopsy: Support for Imaging Surveillance.

Objective: Radial scar (RS) is a low-risk breast lesion that can be associated with or mimic malignancy. Management guidelines remain controversial for patients with RS without atypia on core needle biopsy (CNB). The aim was to evaluate the upgrade rate of these lesions and factors associated with malignancy risk and excision rate to more definitively guide management. Materials and Methods: In this retrospective study, 123 patients with RS without atypia, diagnosed with CNB between January 2008 to December 2014 who were either referred for surgical excision or followed-up with imaging, were reviewed. The differences in clinical presentation, imaging features, and biopsy technique among the benign RS patients and those upgraded, as well as the excised versus the observed patients were compared. Results: Of 123 RS reviewed, 93 cases of RS without atypia as the highest-grade lesion in the ipsilateral breast and with either 24-month imaging follow-up or surgical correlation were included. Seventy-four (79.6%) lesions were surgically excised and 19 (20.4%) were followed-up for at least 24 months. A single upgrade to malignancy (1%) and 15 upgrades to high-risk lesions (16%) were found. There was no association of any upgraded lesion with presenting symptoms or imaging features. The use of vacuum-assistance and larger biopsy needles, along with obtaining a higher number of samples, was associated with fewer upgrades and lower surgical excision rates. Conclusion: The upgrade rate of RS without atypia in our population was low, regardless of the imaging features and biopsy technique utilized. Close imaging surveillance is an acceptable alternative to surgical excision in these patients.

Outcome of Imaging and Biopsy of BI-RADS Category 3 Lesions: Follow-Up Compliance, Biopsy, and Malignancy Rates in a Large Patient Cohort.

OBJECTIVES: To identify biopsy rates and indications for BI-RADS 3 lesions in a large cohort of patients and compare with follow-up compliance and malignancy outcomes. METHODS: We retrospectively reviewed all BI-RADS category-3 lesions seen on mammography and/or ultrasound between 2013 and 2015. Patient age, lesion size, follow-up rates at 6-, 12-, and 24-months were collected. Biopsy timing, indication, and outcomes (malignant vs benign) were recorded using at least 2-year follow-up or biopsy pathology as endpoint. RESULTS: Of 2319 BI-RADS 3 lesions in 2075 women analyzed, biopsy was performed in 173 (7.5%). Most biopsies were performed upfront (99, 57.2%), followed by at 6 (44, 25.4%), 12 (21, 12.1%), and 24-month follow-up (9, 5.2%; P < .001). Palpable (P < .001) and larger (median 1.4 vs 1.0 cm, P < .001) lesions in women <40 years (15.2% vs 4.8%, P < .001) were more likely to undergo biopsy. Most biopsies were prompted by patient/physician desire (64.5%, P < .001). Of 783 lesions with available endpoint, 5 (0.6%) were cancer. All cancers were identified either at presentation (in 0-5 months, n = 1) or 6-month follow-up (in 5-9 months, n = 4) with biopsy prompted by either morphology change (n = 3) or lesion growth (n = 2). Of the 1855 lesions which were expected for follow up, only 310 (16.7%) underwent all follow-ups, while 482 (26.1%) had two, 489 (26.5%) one, and 565 (30.6%) had no follow-up. CONCLUSIONS: In our cohort, BI-RADS category 3 lesions had significantly higher biopsy rates compared with the small malignancy rate, all of which were identified at baseline or first follow-up. Overall patient follow-up compliance low. Imaging follow-up, especially at first 6-month time point, should be encouraged in BI-RADS 3 lesions, instead of upfront biopsies.

ACR BI-RADS Category 3 Lesions in Women Younger Than 30: Follow-up Outcomes and Factors Associated With Biopsy.

OBJECTIVE: Investigate imaging follow-up patterns and assessment of malignancy rate of BI-RADS 3 lesions in women younger than 30 years. METHODS: We retrospectively reviewed consecutive studies between January 1, 2013 and January 1, 2015 with BI-RADS 3 assessment in women <30 years. Lesion size, follow-up rate, and biopsy rate were recorded. Completion of 24-month imaging follow-up or biopsy determined the endpoint. Statistical analysis of follow-up rates and biopsy timing was performed. RESULTS: Of 2525 BI-RADS 3 lesions, 278 were identified in 215 women <30 years. Fifty-two (24%) women underwent a biopsy which was more frequently done at patient request than for lesion growth [33 (63.4%) versus 19 (36.5%), P <.01]. The odds of having biopsy upfront was significantly higher in lesions >2 cm in diameter (OR: 4.4 [95% CI 2.1-9.4], P <.01). The malignancy rate in our cohort was 0% (95% CI 0-1.7%). Of the 188 women expected for follow-up imaging, 58 (30%) were lost to follow-up, while 103 (55%) had 6-month follow-up, 74 (39%) 12-month follow-up, and 56 (30%) 24-month follow-up. CONCLUSIONS: BI-RADS 3 lesions identified in our cohort had high biopsy rates and low compliance with no cancers. Our findings suggest that probable fibroadenomas in young women may only warrant abbreviated short-term follow-up at 6-months.

Deep learning in breast radiology: current progress and future directions.

This review provides an overview of current applications of deep learning methods within breast radiology. The diagnostic capabilities of deep learning in breast radiology continue to improve, giving rise to the prospect that these methods may be integrated not only into detection and classification of breast lesions, but also into areas such as risk estimation and prediction of tumor responses to therapy. Remaining challenges include limited availability of high-quality data with expert annotations and ground truth determinations, the need for further validation of initial results, and unresolved medicolegal considerations. KEY POINTS: • Deep learning (DL) continues to push the boundaries of what can be accomplished by artificial intelligence (AI) in breast imaging with distinct advantages over conventional computer-aided detection. • DL-based AI has the potential to augment the capabilities of breast radiologists by improving diagnostic accuracy, increasing efficiency, and supporting clinical decision-making through prediction of prognosis and therapeutic response. • Remaining challenges to DL implementation include a paucity of prospective data on DL utilization and yet unresolved medicolegal questions regarding increasing AI utilization.

Stereotactic Vacuum-Assisted Needle Biopsy Outcomes of Non-calcified Mammographic Lesions.

RATIONALE AND OBJECTIVES: To identify the outcomes of stereotactic vacuum-assisted large bore biopsies performed on sonographically-occult non-calcified mammographic lesions (NCL). MATERIALS AND METHODS: In an IRB-approved retrospective study, we reviewed all NCL that underwent stereotactic biopsy from January 1, 2014 to December 31, 2017 at our institution, comparing patient age, lesion type, size and location with pathology outcome (benign, high-risk or malignant) using Wilcoxon-Mann-Whitney or Fisher's exact tests as appropriate. Multivariable logistic regression models were developed to decrease benign biopsies in our cohort with diagnostic performance assessed using receiver operating characteristic curve and area under the curve (AUC). RESULTS: Of 222 biopsied lesions in 213 patients, 79.3% (176/222) were benign, 5.9% (13/222) malignant, and 14.9% (33/222) high-risk. NCL were less likely to be malignant compared to calcifications biopsied in the same period [5.9% vs 19.0% (243/1279), p < 0.001]. All 42 asymmetries and 33 architectural distortions were benign, while 8.7% (4/46) of masses and 8.9% (9/101) of focal asymmetries were malignant. Cancers were associated with older age (mean 65.2 vs 52.7 years, p < 0.001), smaller size (mean 9.5 mm vs 15.5 mm, p < 0.01), and concurrent breast cancer (p < 0.01) compared to benign/high-risk lesions. Multivariable logistic regression model using patient age >50 years, lesion type, and size <15 mm had a high diagnostic performance [AUC=0.89, 95%CI (0.83, 0.94)], and yielded the highest PPV [0.24; 95%CI (0.13, 0.38)], and highest number of avoided, unnecessary biopsies (172/209, 82%). CONCLUSION: NCL biopsied under stereotactic guidance have low cancer yield (5.9%). A multivariate model integrating age, lesion size and type could potentially help avoid unwarranted biopsies in our cohort.

Performance of a clinical and imaging-based multivariate model as decision support tool to help save unnecessary surgeries for high-risk breast lesions.

PURPOSE: To investigate the performance of an imaging and biopsy parameters-based multivariate model in decreasing unnecessary surgeries for high-risk breast lesions. METHODS: In an IRB-approved study, we retrospectively reviewed all high-risk lesions (HRL) identified at imaging-guided biopsy in our institution between July 1, 2014-July 1, 2017. Lesions were categorized high-risk-I (HR-I = atypical ductal hyperplasia, atypical lobular hyperplasia, lobular carcinoma in situ and atypical papillary lesion) and II (HR-II = Flat epithelial atypia, radial scar, benign papilloma). Patient risk factors, lesion features, detection and biopsy modality, excision and cancer upgrade rates were collected. Reference standard for upgrade was either excision or at least 2-year imaging follow-up. Multiple logistic regression analysis was performed to develop a multivariate model using HRL type, lesion and biopsy needle size for surgical cancer upgrade with performance assessed using ROC analysis. RESULTS: Of 699 HRL in 652 patients, 525(75%) had reference standard available, and 48/525(9.1%) showed cancer at surgical excision. Excision (84.5% vs 51.1%) and upgrade (17.6%vs1.8%) rates were higher in HR-I compared to HR-II (p < 0.01). In HR-I, small needle size < 12G vs ≥ 12G [32.1% vs 13.2%, p < 0.01] and less cores [< 6 vs ≥ 6, 28.6%vs13.7%, p = 0.01] were significantly associated with higher cancer upgrades. Our multivariate model had an AUC = 0.87, saving 28.1% of benign surgeries with 100% sensitivity, based on HRL subtype, lesion size(mm, continuous), needle size (< 12G vs ≥ 12G) and biopsy modality (US vs MRI vs stereotactic) CONCLUSION: Our multivariate model using lesion size, needle size and patient age had a high diagnostic performance in decreasing unnecessary surgeries and shows promise as a decision support tool.

Is There a Difference in the Diagnostic Outcomes of Calcifications Initially Identified on Synthetic Tomosynthesis Versus Full-Field Digital Mammography Screening?

PURPOSE: To compare the outcomes of microcalcifications recalled on full-field digital (FFDM) and FFDM and combined tomosynthesis (Combo) to synthetic (SM) screening mammograms. METHOD: We reviewed medical records, radiology, and pathology reports of all patients found to have abnormal calcifications requiring further evaluation on mammography screening at our institution between 11/1/2016-11/1/2018 and collected patient demographics, calcification morphology and distribution, and mammography technique (SM, FFDM, or Combo). We used biopsy pathology or at least 1-year imaging follow-up to establish overall diagnostic outcome (benign or malignant). Fisher's exact test was used to compare validation rates at diagnostic work-up, BI-RADS category, and final outcome of calcifications identified on each screening technique. T-test was used for continuous variables. RESULTS: Of 699 calcifications in 596 women recalled, 176 (30%) of 596 were from SM and 420 (70%) FFDM/Combo. There was a significantly higher rate of calcifications unvalidated at diagnostic work-up for SM compared to FFDM/Combo (0.8% vs. 10%, p < 0.0001). SM calcifications were more likely to receive BI-RADS 2/3 at diagnostic work-up compared to FFDM/Combo ones (55% vs. 42%, p = 0.003). Of 346 (49%) calcifications that underwent biopsy, 88 (25%) were malignant (36% of SM vs. 22% of FFDM/Combo, OR:0.5 [95% CI: 0.3, 0.8] p = 0.01). Of 622 lesions with established diagnostic outcome, there was no difference between having an overall benign or malignant outcome between SM and FFDM/Combo (17% vs. 13%, OR: 0.8 [95% Cl: 0.5, 1.2] p = 0.27). CONCLUSIONS: Synthetic tomosynthesis screening results in a higher rate of false positive and unvalidated calcification recalls compared to FFDM/Combo.

Contrast-Enhanced Digital Mammography: Technique, Clinical Applications, and Pitfalls.

OBJECTIVE. Contrast-enhanced digital mammography (CEDM) combines the high spatial resolution of mammography with the improved enhancement provided by contrast medium. In this article, CEDM technique-the current and potential clinical applications and current challenges-will be reviewed. CONCLUSION. CEDM is a promising technique in the supplemental evaluation of patients with mammographically inconclusive findings and potentially in the screening of women with mammographically dense breasts. CEDM is emerging as a cost-effective alternative to dynamic contrast-enhanced MRI to stage newly diagnosed breast cancer and evaluate response to neoadjuvant chemotherapy.

Benign breast papillomas without atypia diagnosed with core needle biopsy: Outcome of surgical excision and imaging follow-up.

PURPOSE: To evaluate the surgical upgrade rate to malignancy and high-risk lesions in cases of papilloma without atypia diagnosed with imaging-guided percutaneous core needle biopsy (CNB) and to determine whether any lesion imaging features, biopsy techniques, and pathological factors can predict lesion upgrade to help guide clinical management. MATERIALS AND METHODS: Benign papillomas without atypia (n = 399) diagnosed with CNB were retrospectively reviewed. The surgical upgrade rate to malignancy or high-risk lesion (atypical ductal hyperplasia, atypical lobular hyperplasia, lobular carcinoma in-situ, flat epithelial atypia and atypical papilloma) was determined. Detection modality (i.e. mammography, ultrasonography (US), magnetic resonance imaging (MRI)), lesion type and size, biopsy-guidance modality (US, stereotactic, MRI), biopsy needle size (<14 G vs ≥14 G), use of vacuum assistance, and presenting symptoms were statistically analyzed. The reference standard for evaluation of upgrade was either excision or at least 24 months of imaging follow-up. Chi Square test and Fisher exact tests were performed for categorical variables, and the Mann-Whitney-U test was used for continuous variables. RESULTS: Ultrasound was the predominant biopsy modality (78.4 %, p < 0.001). Of the 399 benign papilloma lesions in 329 women, 239 (59.9 %) were excised and 93 others were followed for at least 24 months (total of 332). Of these 332 lesions, 7 (2.1 %) were upgraded to ductal carcinoma in-situ and 41 (12.3 %) to high-risk lesions at excision. Larger lesion size (≥15 mm, p = 0.009), smaller biopsy needle size (≥14 G, p = 0.027), and use of spring-loaded biopsy device (p = 0.012) were significantly associated with upgrade to atypia. Only lesion size (≥15 mm, p = 0.02) was associated with upgrade to cancer. CONCLUSION: Upgrade to malignancy of biopsy-proven benign papillomas without atypia at the time of surgery was sufficiently low (2.1 %) to support non-operative management. Surgery may be performed for selected cases- those with larger lesion size and those whose biopsies were performed with smaller spring-loaded biopsy needles.

Preoperative Prediction of Lymph Node Metastasis from Clinical DCE MRI of the Primary Breast Tumor Using a 4D CNN.

In breast cancer, undetected lymph node metastases can spread to distal parts of the body for which the 5-year survival rate is only 27%, making accurate nodal metastases diagnosis fundamental to reducing the burden of breast cancer, when it is still early enough to intervene with surgery and adjuvant therapies. Currently, breast cancer management entails a time consuming and costly sequence of steps to clinically diagnose axillary nodal metastases status. The purpose of this study is to determine whether preoperative, clinical DCE MRI of the primary tumor alone may be used to predict clinical node status with a deep learning model. If possible then many costly steps could be eliminated or reserved for only those with uncertain or probable nodal metastases. This research develops a data-driven approach that predicts lymph node metastasis through the judicious integration of clinical and imaging features from preoperative 4D dynamic contrast enhanced (DCE) MRI of 357 patients from 2 hospitals. Innovative deep learning classifiers are trained from scratch, including 2D, 3D, 4D and 4D deep convolutional neural networks (CNNs) that integrate multiple data types and predict the nodal metastasis differentiating nodal stage N0 (non metastatic) against stages N1, N2 and N3. Appropriate methodologies for data preprocessing and network interpretation are presented, the later of which bolster radiologist confidence that the model has learned relevant features from the primary tumor. Rigorous nested 10-fold cross-validation provides an unbiased estimate of model performance. The best model achieves a high sensitivity of 72% and an AUROC of 71% on held out test data. Results are strongly supportive of the potential of the combination of DCE MRI and machine learning to inform diagnostics that could substantially reduce breast cancer burden.