Deep Learning Image Analysis of Benign Breast Disease to Identify Subsequent Risk of Breast Cancer

Benign Breast Disease and Breast Cancer Risk in the Percutaneous Biopsy Era

Importance

Benign breast disease (BBD) comprises approximately 75% of breast biopsy diagnoses. Surgical biopsy specimens diagnosed as nonproliferative (NP), proliferative disease without atypia (PDWA), or atypical hyperplasia (AH) are associated with increasing breast cancer (BC) risk; however, knowledge is limited on risk associated with percutaneously diagnosed BBD.

Objectives

To estimate BC risk associated with BBD in the percutaneous biopsy era irrespective of surgical biopsy.

Design, Setting, and Participants

In this retrospective cohort study, BBD biopsy specimens collected from January 1, 2002, to December 31, 2013, from patients with BBD at Mayo Clinic in Rochester, Minnesota, were reviewed by 2 pathologists masked to outcomes. Women were followed up from 6 months after biopsy until censoring, BC diagnosis, or December 31, 2021.

Exposure

Benign breast disease classification and multiplicity by pathology panel review.

Main Outcomes

The main outcome was diagnosis of BC overall and stratified as ductal carcinoma in situ (DCIS) or invasive BC. Risk for presence vs absence of BBD lesions was assessed by Cox proportional hazards regression. Risk in patients with BBD compared with female breast cancer incidence rates from the Iowa Surveillance, Epidemiology, and End Results (SEER) program were estimated.

Results

Among 4819 female participants, median age was 51 years (IQR, 43-62 years). Median follow-up was 10.9 years (IQR, 7.7-14.2 years) for control individuals without BC vs 6.6 years (IQR, 3.7-10.1 years) for patients with BC. Risk was higher in the cohort with BBD than in SEER data: BC overall (standard incidence ratio [SIR], 1.95; 95% CI, 1.76-2.17), invasive BC (SIR, 1.56; 95% CI, 1.37-1.78), and DCIS (SIR, 3.10; 95% CI, 2.54-3.77). The SIRs increased with increasing BBD severity (1.42 [95% CI, 1.19-1.71] for NP, 2.19 [95% CI, 1.88-2.54] for PDWA, and 3.91 [95% CI, 2.97-5.14] for AH), comparable to surgical cohorts with BBD. Risk also increased with increasing lesion multiplicity (SIR: 2.40 [95% CI, 2.06-2.79] for ≥3 foci of NP, 3.72 [95% CI, 2.31-5.99] for ≥3 foci of PDWA, and 5.29 [95% CI, 3.37-8.29] for ≥3 foci of AH). Ten-year BC cumulative incidence was 4.3% for NP, 6.6% for PDWA, and 14.6% for AH vs an expected population cumulative incidence of 2.9%.

Conclusions and Relevance

In this contemporary cohort study of women diagnosed with BBD in the percutaneous biopsy era, overall risk of BC was increased vs the general population (DCIS and invasive cancer combined), similar to that in historical BBD cohorts. Development and validation of pathologic classifications including both BBD severity and multiplicity may enable improved BC risk stratification.

Built to Last? Reproducibility and Reusability of Deep Learning Algorithms in Computational Pathology

Recent progress in computational pathology has been driven by deep learning. While code and data availability are essential to reproduce findings from preceding publications, ensuring a deep learning model's reusability is more challenging. For that, the codebase should be well-documented and easy to integrate into existing workflows and models should be robust toward noise and generalizable toward data from different sources. Strikingly, only a few computational pathology algorithms have been reused by other researchers so far, let alone employed in a clinical setting. To assess the current state of reproducibility and reusability of computational pathology algorithms, we evaluated peer-reviewed articles available in PubMed, published between January 2019 and March 2021, in 5 use cases: stain normalization; tissue type segmentation; evaluation of cell-level features; genetic alteration prediction; and inference of grading, staging, and prognostic information. We compiled criteria for data and code availability and statistical result analysis and assessed them in 160 publications. We found that only one-quarter (41 of 160 publications) made code publicly available. Among these 41 studies, three-quarters (30 of 41) analyzed their results statistically, half of them (20 of 41) released their trained model weights, and approximately a third (16 of 41) used an independent cohort for evaluation. Our review is intended for both pathologists interested in deep learning and researchers applying algorithms to computational pathology challenges. We provide a detailed overview of publications with published code in the field, list reusable data handling tools, and provide criteria for reproducibility and reusability.

Host, reproductive, and lifestyle factors in relation to quantitative histologic metrics of the normal breast

BACKGROUND

Emerging data indicate that variations in quantitative epithelial and stromal tissue composition and their relative abundance in benign breast biopsies independently impact risk of future invasive breast cancer. To gain further insights into breast cancer etiopathogenesis, we investigated associations between epidemiological factors and quantitative tissue composition metrics of the normal breast.

METHODS

The study participants were 4108 healthy women ages 18-75 years who voluntarily donated breast tissue to the US-based Susan G. Komen Tissue Bank (KTB; 2008-2019). Using high-accuracy machine learning algorithms, we quantified the percentage of epithelial, stromal, adipose, and fibroglandular tissue, as well as the proportion of fibroglandular tissue that is epithelium relative to stroma (i.e., epithelium-to-stroma proportion, ESP) on digitized hematoxylin and eosin (H&E)-stained normal breast biopsy specimens. Data on epidemiological factors were obtained from participants using a detailed questionnaire administered at the time of tissue donation. Associations between epidemiological factors and square root transformed tissue metrics were investigated using multivariable linear regression models.

RESULTS

With increasing age, the amount of stromal, epithelial, and fibroglandular tissue declined and adipose tissue increased, while that of ESP demonstrated a bimodal pattern. Several epidemiological factors were associated with individual tissue composition metrics, impacting ESP as a result. Compared with premenopausal women, postmenopausal women had lower ESP [β (95% Confidence Interval (CI)) = -0.28 (- 0.43, - 0.13); P < 0.001] with ESP peaks at 30-40 years and 60-70 years among pre- and postmenopausal women, respectively. Pregnancy [β (95%CI) vs nulligravid = 0.19 (0.08, 0.30); P < 0.001] and increasing number of live births (P-trend < 0.001) were positively associated with ESP, while breastfeeding was inversely associated with ESP [β (95%CI) vs no breastfeeding = -0.15 (- 0.29, - 0.01); P = 0.036]. A positive family history of breast cancer (FHBC) [β (95%CI) vs no FHBC = 0.14 (0.02-0.26); P = 0.02], being overweight or obese [β (95%CI) vs normal weight = 0.18 (0.06-0.30); P = 0.004 and 0.32 (0.21-0.44); P < 0.001, respectively], and Black race [β (95%CI) vs White = 0.12 (- 0.005, 0.25); P = 0.06] were positively associated with ESP.

CONCLUSION

Our findings revealed that cumulative exposure to etiological factors over the lifespan impacts normal breast tissue composition metrics, individually or jointly, to alter their dynamic equilibrium, with potential implications for breast cancer susceptibility and tumor etiologic heterogeneity.

Associations of alcohol consumption with breast tissue composition

BACKGROUND

We investigated the associations of alcohol with percentage of epithelium, stroma, fibroglandular tissue (epithelium + stroma), and fat in benign breast biopsy samples.

METHODS

We included 857 cancer-free women with biopsy-confirmed benign breast disease within the Nurses' Health Study (NHS) and NHSII cohorts. Percentage of each tissue was measured on whole slide images using a deep-learning algorithm and then log-transformed. Alcohol consumption (recent and cumulative average) was assessed with semi-quantitative food frequency questionnaires. Regression estimates were adjusted for known breast cancer risk factors. All tests were 2-sided.

RESULTS

Alcohol was inversely associated with % of stroma and fibroglandular tissue (recent ≥ 22 g/day vs. none: stroma: β = - 0.08, 95% Confidence Interval [CI] - 0.13; - 0.03; fibroglandular: β = - 0.08, 95% CI - 0.13; - 0.04; cumulative ≥ 22 g/day vs. none: stroma: β = - 0.08, 95% CI - 0.13; - 0.02; fibroglandular: β = - 0.09, 95% CI - 0.14; - 0.04) and positively associated with fat % (recent ≥ 22 g/day vs. none: β = 0.30, 95% CI 0.03; 0.57; cumulative ≥ 22 g/day vs. none: β = 0.32, 95% CI 0.04; 0.61). In stratified analysis, alcohol consumption was not associated with tissue measures in premenopausal women. In postmenopausal women, cumulative alcohol use was inversely associated with % of stroma and fibroglandular tissue and positively associated with fat % (≥ 22 g/day vs. none: stroma: β = - 0.16, 95% CI - 0.28; - 0.07; fibroglandular: β = - 0.18, 95% CI - 0.28; - 0.07; fat: β = 0.61, 95% CI 0.01; 1.22), with similar results for recent alcohol use.

CONCLUSION

Our findings suggest that alcohol consumption is associated with smaller % of stroma and fibroglandular tissue and a greater % of fat in postmenopausal women. Future studies are warranted to confirm our findings and to elucidate the underlying biological mechanisms.

Histological Findings of Mammary Gland Development and Risk of Breast Cancer in BRCA1 Mutant Mouse Models

Purpose

The breast cancer susceptibility gene, BRCA1, is involved in normal development and carcinogenesis of mammary glands. Here, we aimed to evaluate the relationship between histological findings of mammary gland development and breast cancer risk in BRCA1 mutant mice.

Methods

Five BRCA1 mutant mice and five non-mutant FVB/NJ mice were used for each group of 1-month-old (pubertal), 3-month-old (fertile), and 8-month-old (menopausal) mice. In another experiment, 15 BRCA1 mutant mice were followed up to 8 months after birth and classified into tumor-bearing (11 mice) and tumor-free (4 mice) groups. Excised mammary gland tissues were stained with Carmine Alum, and the number of terminal end buds (or alveolar buds), branching density, and duct elongation were measured using image analysis programs. Differences between the two groups were assessed using paired t-test.

Results

One-month-old BRCA1 mutant mice showed a higher number of terminal end buds (23.8 ± 1.0 vs. 15.6 ± 0.8, p = 0.0002), branching density (11.7 ± 0.4 vs. 9.6 ± 0.5%, p = 0.0082), and duct elongation (9.7 ± 0.7 vs. 7.3 ± 0.4 mm, p = 0.0186) than controls. However, there was no difference between the 3- and 8-month-old groups. In BRCA1 mutant mice, the tumor-bearing group showed a significantly higher number of alveolar buds (142.7 ± 5.5 vs. 105.5 ± 5.4, p = 0.0008) and branching density (30.0 ± 1.0 vs. 24.1 ± 1.1%, p = 0.008) than the tumor-free group; however, duct elongation was not different (23.9 ± 0.6 vs. 23.6 ± 0.6 mm, p = 0.8099) between the groups.

Conclusion

BRCA1 mutant mice exhibited early pubertal mammary gland development and delayed age-related mammary gland involution was associated with breast cancer. Our results may have clinical implications for predicting breast cancer risk and developing prevention strategies for BRCA1 mutation carriers.

Associations of reproductive breast cancer risk factors with breast tissue composition

BACKGROUND

We investigated the associations of reproductive factors with the percentage of epithelium, stroma, and fat tissue in benign breast biopsy samples.

METHODS

This study included 983 cancer-free women with biopsy-confirmed benign breast disease (BBD) within the Nurses' Health Study and Nurses' Health Study II cohorts. The percentage of each tissue type (epithelium, stroma, and fat) was measured on whole-section images with a deep-learning technique. All tissue measures were log-transformed in all the analyses to improve normality. The data on reproductive variables and other breast cancer risk factors were obtained from biennial questionnaires. Generalized linear regression was used to examine the associations of reproductive factors with the percentage of tissue types, while adjusting for known breast cancer risk factors.

RESULTS

As compared to parous women, nulliparous women had a smaller percentage of epithelium (β = - 0.26, 95% confidence interval [CI] - 0.41, - 0.11) and fat (β = - 0.34, 95% CI - 0.54, - 0.13) and a greater percentage of stroma (β = 0.04, 95% CI 0.01, 0.08). Among parous women, the number of children was inversely associated with the percentage of stroma (β per child = - 0.01, 95% CI - 0.02, - 0.00). The duration of breastfeeding of ≥ 24 months was associated with a reduced proportion of fat (β = - 0.30, 95% CI - 0.54, - 0.06; p-trend = 0.04). In a separate analysis restricted to premenopausal women, older age at first birth was associated with a greater proportion of epithelium and a smaller proportion of stroma.

CONCLUSIONS

Our findings suggest that being nulliparous as well as having a fewer number of children (both positively associated with breast cancer risk) is associated with a smaller proportion of epithelium and a greater proportion of stroma, potentially suggesting the importance of epithelial-stromal interactions. Future studies are warranted to confirm our findings and to elucidate the underlying biological mechanisms.

Early-Life and Adult Adiposity, Adult Height, and Benign Breast Tissue Composition

BACKGROUND

Early-life and adult anthropometrics are associated with breast density and breast cancer risk. However, little is known about whether these factors also influence breast tissue composition beyond what is captured by breast density among women with benign breast disease (BBD).

METHODS

This analysis included 788 controls from a nested case-control study of breast cancer within the Nurses' Health Study BBD subcohorts. Body fatness at ages 5 and 10 years was recalled using a 9-level pictogram. Weight at age 18, current weight, and height were reported via questionnaires. A deep-learning image analysis was used to quantify the percentages of epithelial, fibrous stromal, and adipose tissue areas within BBD slides. We performed linear mixed models to estimate beta coefficients (β) and 95% confidence intervals (CI) for the relationships between anthropometrics and the log-transformed percentages of individual tissue type, adjusting for confounders.

RESULTS

Childhood body fatness (level ≥ 4.5 vs. 1), BMI at age 18 (≥23 vs. <19 kg/m²), and current adult BMI (≥30 vs. <21 kg/m²) were associated with higher proportions of adipose tissue [β (95% CI) = 0.34 (0.03, 0.65), 0.19 (-0.04-0.42), 0.40 (0.12, 0.68), respectively] and lower proportions of fibrous stromal tissue [-0.05 (-0.10, 0.002), -0.03 (-0.07, 0.003), -0.12 (-0.16, -0.07), respectively] during adulthood (all P _trend < 0.04). BMI at age 18 was also inversely associated with epithelial tissue (P _trend = 0.03). Adult height was not associated with any of the individual tissue types.

CONCLUSIONS

Our data suggest that body fatness has long-term impacts on breast tissue composition.

IMPACT

This study contributes to our understanding of the link between body fatness and breast cancer risk.See related commentary by Oskar et al., p. 590.