Towards defining morphologic parameters of normal parous and nulliparous breast tissues by artificial intelligence

KTB was established in 2007 to acquire questionnaires, blood, and normal breast tissues from consented volunteers [16]. Donors provide multiple breast tissue cores that are snap-frozen or fixed and paraffin-embedded. The KTB uses standardized tissue sampling and our previous analysis of samples (n = 50) demonstrated that adjusting for tissue area in images of hematoxylin and eosin (H&E)-stained sections does not impact conclusions (unpublished). The KTB protocol was approved by the Institutional Review Board at Indiana University. The current project was also approved by the Mayo Clinic Institutional Review Board.

To compare histologic features of breast tissues of younger parous and nulliparous women, we obtained data for all 1,082 KTB donors enrolled from 2009 to 2018 and aged ≤ 45 years at enrollment, including 801 nulliparas and 281 parous women who had given birth within 10 years of tissue acquisition. We excluded women who did not self-identify as white or African American (29 parous, 96 nulliparous), women who were found to be above age 45 years at donation on further review (17 parous, 52 nulliparous), women whose samples did not contain evaluable TDLUs (36 parous, 63 nulliparous), women with a history of cancer (9 parous, 11 nulliparous), and women who had undergone an oophorectomy (7 parous, 5 nulliparous). For 32 repeat donors (3 parous and 29 nulliparous), only the first tissue donation was analyzed. The final sample set analyzed by AI and visual morphometry included 725 women, of whom 180 were parous (149 white, 31 African American; mean age = 35 years) and 545 were nulliparous (479 Caucasian, 66 African American; mean age = 25 years.

Women completed self-administered BC risk factor questionnaires assessing age at donation, race, ethnicity, age at first period, menopausal status, age at first birth, number of live births, any relatives with a history of breast/ovarian cancer, medical conditions, weight and height (assessed as body mass index (BMI) (Kg/m²)), smoking and ethanol consumption. A supplementary early life questionnaire assessing onset of puberty, dates of live births, and breastfeeding history (total months across all pregnancies) was collected https://​komentissuebank.​iu.​edu/​donate-tissue/​docs/​early-life-questionnaire-letter-a071.​pdf.

We applied previously validated automated AI methods to quantitate features related to scanned whole slide images (Aperio ScanScope XT Slide Scanner, Leica Biosystems) at 20X with resolution of 0.495 × 0.495  μm² per pixel or specific to TDLUs (Additional file 1: Fig. 1). Whole slide metrics assessed, included: TDLU count and adipose tissue fraction; TDLU-specific features, included: mean acini count/TDLU; mean dilated acini; mean average acini area; mean “capillary” area; mean epithelial area; mean ratio of epithelial area versus intralobular stroma; mean mononuclear cell count; mean fat area proximate to TDLUs and TDLU area. Mononuclear cell counts refer to all cells with small round nuclei, which may represent lymphocytes, plasma cells, macrophages, and less common immune cell types (e.g., mast cells). Previously, we found that our AI measurements agreed well with results of a four-member panel using a 6-tier classification of TDLU involution (unweighted kappa = 0.747 ± 0.01) (Additional file 1: Fig. 1) [17].

A trained reviewer (JO) independently assessed morphometric features of TDLUs using a validated method [8, 15, 17]. We limited morphometric analysis to normal appearing TDLUs and excluded TDLUs showing subtle features of benign breast disease, represented mainly by dilatation of acini two- to three-fold or focal metaplastic or proliferative changes. Total normal TDLUs per section were counted and ≤ 10 TDLUs first encountered per section in scanning were evaluated for acini number/TDLU scored categorically (1 = 1–9, 2 = 10–19, 3 = 20–29, 4 = 30–39, 5 = 40–49, 6 =  > 50), and TDLU span in um [8, 15, 17]. Data were analyzed per participant as the mean acini counts/TDLU and median TDLU span. A pathologist (MES) masked to other data, assessed images that contained ≥ 5 TDLUs (147 parous and 381 nulliparous donors) for “TDLU inflammation”, subjectively assessed as hypercellularity versus surrounding TDLUs, plasma cells (present if ≥ 3 per section), and dilated acini. AI and morphometric features were highly correlated: TDLU counts (r = 0.84; p < 0.0001); acini counts (p = 0.80; p < 0.0001). AI mononuclear cell counts were associated with visual assessment of inflammation (mean/maximum counts per TDLU for inflamed:118.2/444.0 versus not inflamed: 82.0/233.5, respectively; both comparisons p < 0.001). Our AI method was not trained to identify plasma cells.

We compared epidemiologic characteristics between parous and nulliparous women, and between white and African American women, using a Wilcoxon rank sum test (continuous and ordinal variables) or Fisher’s exact test (categorical variables). Histologic features (assessed by AI or visually) were compared between nulliparous and parous women (overall and by time between last birth and donation [ recent birth: ≤ 5 years versus remote birth: > 5 years]) using multivariable regression models (linear regression, negative binomial regression, proportional odds logistic regression, and binary logistic regression) appropriate to the nature of the specific histologic feature (continuous, count, ordinal or dichotomous, respectively; see table footnotes). Dichotomization at 5 years was chosen to reflect peak postpartum BC risk [13]. We performed exploratory analysis to assess variability among women who donated ≤ 5 years of last birth. For linear regression models, data transformations were applied to account for distributional skewness as needed (see table footnotes), and additive effects on the mean value of the histologic feature (i.e., parous minus nulliparous) were estimated with 95% confidence intervals (CIs). For negative binomial regression models, multiplicative effects on mean values of histologic features were estimated with 95% CIs. For proportional odds logistic regression models, odds ratios (ORs) and 95% CIs were estimated and are interpreted as the multiplicative increase in the odds of a higher category for parous versus nulliparous women. For logistic regression models, ORs and 95% CIs were estimated and are interpreted as the multiplicative increase in the odds of histologic features among parous versus nulliparous women. Multivariable models were initially adjusted for age at donation, and then further adjusted for any characteristic that differed between groups (p value < 0.15). Comparisons of AI features between white and African American women were made following this same multivariable regression analysis strategy. A Bonferroni correction for multiple testing was applied separately to AI analyses (11 tests, p values < 0.0045 considered as statistically significant) and for visual/morphometric assessments (6 tests, p values < 0.0083 considered as statistically significant).

We also performed exploratory analyses of determinants of TDLU number, dilated acini and mononuclear cells count (AI) or inflammation (morphometry) separately among parous and nulliparous women using regression models and compared AI histologic features by race. For reference, analyses among parous women after Bonferroni correction would require p ≤ 0.0036 and among nulliparas, p ≤ 0.0056 to achieve statistical significance. All statistical tests were two-sided and were performed using SAS (version 9.4; SAS Institute, Inc., Cary, North Carolina).

Participants included 545 nulliparous and 180 parous women, of whom 103 had their most recent birth within 5 years and 77 more than 5 years previously (median time since birth of 4 years). Characteristics are presented in Table 1. Compared with nulliparas, parous women were older (median 35 years versus 25 years; p < 0.001), more frequently reported any family history of breast/ovarian cancer (61.9% vs 49.3%, p = 0.006) and were marginally heavier (median BMI of 27.0 versus 25.4; p = 0.062). Compared with white women, African American donors were older (median age of 33 versus 27 years (p < 0.001)), heavier (median BMI = 31.6 versus 25.1 kg/m²; p < 0.001), reported less frequent ethanol use (59.8% versus 73.9%; p = 0.005); earlier age at menarche (p < 0.001), and, among parous women, longer interval from birth to donation (median 6.0 versus 3.7 years; p = 0.012) (Additional file 1: Table 1).

AI analyses revealed that tissues of parous women contained significantly higher TDLU counts and acini counts, more frequent dilated acini, higher mononuclear cell counts in TDLUs and smaller acini area per TDLU than nulliparas (all multivariable analyses p < 0.001; significant after multiple comparison adjustment) (Table 2). Using visual morphometric assessment, parous women had significantly higher TDLU counts (p < 0.001), and suggestively, but non-significantly increased plasma cells (p = 0.20) and inflammation (p = 0.14) (Additional file 1: Table 2). Associations remained after adjusting for age and other potential confounding variables.

We performed additional analysis stratifying data for parous women as ≤ 5 versus > 5 years postpartum. Irrespective of postpartum interval (i.e., time from most recent birth to donation), AI analysis showed that TDLU counts were higher and acinar area smaller among parous women compared to nulliparous women. In contrast, only specimens obtained within 5 years of childbirth showed higher mean acini counts (multiplicative effect on mean: 1.91, 95% CI = 1.60–2.28; p < 0.001), more frequent dilated acini (multiplicative effect on mean: 2.07, 95% CI = 1.59–2.70; p < 0.001) and increased mononuclear cells in TDLUs (multiplicative effect on mean: 1.82, 95% CI = 1.54–2.14; p < 0.001) compared with those from nulliparas (Fig. 1; Table 3). Similarly, visual analysis confirmed the AI findings above regarding a statistically significant increase in TDLU counts among parous versus nulliparous women, irrespective of the interval from birth to donation, and increased mononuclear cells in TDLUS only in samples obtained within 5 years of a birth (surrogate for all types of immune cells) (Additional file 1: Fig. 2). To assess AI changes evolving over the first 5 years postpartum, we performed an exploratory analysis of AI data for 101 parous women who donated tissue within 5 years of a last birth (per one-year intervals), which revealed suggestive declines in mean acini count (p = 0.015), number of dilated acini (p = 0.70) and mononuclear cells over the course of the early postpartum period (Additional file 1: table 3: changes per year for years 1–5).

We performed separate multivariable analyses of AI-derived histologic features in nulliparous and in parous tissues. In both groups, higher BMI (and percentage fat in tissue) was associated with statistically significantly lower TDLU counts and older age was related to increased detection of dilated acini (Additional file 1: Tables 4–5). Among parous women, higher numbers of births were related to reduced detection of dilated acini (OR = 0.19; 95%CI: 0.08–0.47; p < 0.001) and breastfeeding was related to increased detection of dilated acini (OR = 2.22; 95% CI: 1.30–3.80; p < 0.004; Additional file 1: Table 5). Also, among nulliparas, later age at menarche was associated with lower mononuclear cell counts (OR = 0.73, 95% CI: 0.58–0.92; p = 0.008; Additional file 1: Table 6). Additional associations that were statistically significant in multivariate analyses but did not reach significance when considering multiple comparisons, included: 1) current ethanol use and increased TDLU numbers in both parous and nulliparous women; increased detection of dilated acini among nulliparas of Hispanic ethnicity and higher mononuclear cell counts among women with a family history of breast/ovarian cancer. To assess effects of breastfeeding after last birth prior to donation (data were collected as total months across all pregnancies), we restricted analyses to uniparous women, which revealed that breastfeeding was associated with reduced TDLU counts (multiplicative effect 0.72, 95%CI = 0.53–0.98; p = 0.037); increased frequency of dilated acini (multiplicative effect 2.01, 95%CI = 1.37–2.93; p < 0.001), higher mean acini area (additive effect 0.16, 95%CI = 0.01–0.31; p = 0.032); increased mean capillary counts (multiplicative effect 1.21, 95%CI = 1.03–1.42; p = 0.023) and increased mononuclear cells (multiplicative effect 1.30, 95%CI = 1.08–1.57; p = 0.006) (data not shown).

In multivariable analyses, African American women’s tissues showed lower TDLU counts and increased acini counts, capillary area and mononuclear cell counts and TDLU area; however, none of these results were significant after considering multiple comparisons (Additional file 1: Table 7).

Given that obesity and percentage of fat on slides are correlated, and that poor correlation may sometimes reflect chance non-representative sampling [8], we re-analyzed data stratified by obesity status, percentage of fat on the slide (≤ 50%; 60–80% and ≥ 90%) and cross-classified by these factors. Results were generally consistent with the analysis presented above (data not shown). Results were substantially unchanged after excluding postmenopausal women.