A simple morphometric algorithm based on nuclear atypia features of invasive breast carcinoma: relationship with nuclear grade, Ki-67 level, and hormone receptor status Authors: Tomohiro Tsuru, Takanori Watanabe, Keita Kouzu, Youji Yamagishi, Takahiro Einama, Tamio Yamasaki, Hironori Tsujimoto, Yoji Kishi, Hideki Ueno, Hitoshi Tsuda Published: 2026 ABSTRACT AbstractBackground With the widespread availability of whole-slide imaging, many studies have utilized digital images of hematoxylin and eosin (H&E)-stained breast cancer tissues and applied convolutional neural networks (CNNs) for pathological diagnosis. However, CNN-based diagnosis is largely a black box and may be limited in quantitative morphological research. In this study, we developed a simple algorithm for morphometric analysis of three nuclear atypia features on H&E-stained whole-slide images to predict nuclear grade, hormone receptor status, and Ki-67 levels in breast cancer.Materials and Methods Using 43,183 H&E-stained nuclear images larger than 20 µm2 from 131 invasive ductal breast carcinomas, we calculated the following features of nuclear atypia using a computer vision algorithm: anisonucleosis (variation in nuclear size), inhomogeneous chromatin density, and the average size of prominent nucleoli. Anisonucleosis was quantified as the percentage of nuclei larger than 47 μm². Inhomogeneous chromatin was defined as the percentage of blue-saturated structures with 0.92-fold luminance or less than the average nuclear luminance. Prominent nucleoli were identified based on blue-saturated structures with 0.87-fold luminance or less, circularity greater than 0.65, and size greater than 1.15 μm². Using these values of nuclear atypia features, the thresholds that were the most associated with grade and biomarkers were calculated using receiver operating characteristic curves by Youden index.Results The morphometric algorithm using these thresholds predicted nuclear grade (grade 1 and 3), Ki-67 index ≧ 20%, and hormone receptor negative status with sensitivities of 52.1 to 100% and specificities from 34.3 to 85.3%. Two multivariable logistic regression models combining these three thresholds predicted nuclear grade, Ki-67, and hormone receptor negative with much better accuracy, sensitivities ranging from 57.1 to 91.3%, specificities from 50.9 to 82.2%, and area under curve of 0.70–0.82. The algorithm was applied to an independent set of 42 tumors.Conclusion The present morphological algorithm of nuclear atypia might provide new insights into the computational grading of invasive breast cancer.Supplementary InformationThe online version contains supplementary material available at 10.1007/s12282-026-01870-3.Keywords: Invasive breast carcinoma, Morphometry, Nuclear grade, Nuclear atypia, Hormone receptor, Digital pathology, Breast cancer, Whole-slide imaging, Nucleoli Introduction The pathological grading of malignancy, such as histological and nuclear grades, is utilized in daily diagnostic practice as a prognostic indicator and for treatment decisions in the management of patients with primary breast cancer [1–3]. Nuclear atypia or nuclear pleomorphism is one of the major elements of histological and nuclear grades and has been shown to be correlated with patient prognosis. However, because the criteria for nuclear atypia are semiquantitative and somewhat subjective, numerous studies have suggested that pathologists sometimes assign different scores for nuclear atypia to the same area of a tumor [4]. Therefore, the necessity and utility of automatic imaging diagnosis for the quantitative and objective pathological grading of breast cancer have been recognized in many studies. Whole-slide digital imaging has become widely available since 2010. Veta et al. analyzed nuclear morphometric features on whole-slide breast cancer images of 101 breast cancers and developed a method to quantify nuclear size deviations of automatically segmented hematoxylin and eosin (H&E)-stained images of breast cancer nuclei to predict patient prognosis [5]. In the late 2010 s, digital pathology gained momentum with the advent of deep learning as a strong tool for histopathological image analysis. Deep learning automatically detects a hierarchy of features, where higher-level concepts are built on lower-level concepts. Couture et al. trained the neural network classifier by learning 571 H&E-stained breast cancer pathology images and distinguished tumor grade, estrogen receptor (ER) status, histological type, and intrinsic subtype with 85 ± 10% accuracy [6]. More recently, several studies have explored deep learning-based tumor grades, focusing on nuclear morphological features. Mahmood et al. succeeded in accurately counting mitotic cells on histological breast cancer images based on faster region convolutional neural network (CNN) and deep CNN methods [7]. Jaroensri et al. developed deep learning models for each component of the Nottingham Grading System, that is, mitotic count, nuclear pleomorphism, and tubule formation, and showed their ability to predict these histopathological features relatively accurately at both the patch and slide levels [8]. However, these CNN-based methods have a black box in some parts and sometimes have difficulties in quantitative morphological studies for the diagnosis of the indication of systemic therapies and the prediction of response to therapies. According to the World Health Organization (WHO) Classification of Breast Tumours, 5th edition, the grade of malignancy for invasive carcinoma of no special type (also known as invasive ductal carcinoma) is determined by histological grading. This grading is based on three criteria: tubule and gland formation, nuclear pleomorphism, and mitotic count [3]. According to the General Rules for Clinical and Pathological Recording of Breast Cancer, 18th edition (Japan), the grade of malignancy of invasive ductal carcinoma is also classified by nuclear grade, which is based on nuclear atypia and mitotic counts [9, 10]. Nuclear atypia—equivalent to nuclear pleomorphism—is semi-quantitatively scored based on the degree of nuclear size variation (anisonucleosis), uneven chromatin distribution, and prominent nucleoli [3, 9]. However, few studies have used computer morphometric approaches to evaluate nuclear atypia, presumably because these three parameters are difficult to quantify or stratify. Here, we developed simple computer morphometric approaches to analyze anisonucleosis, chromatin density inhomogeneity, and the average size of prominent nucleoli. We attempted to devise algorithms that predict grade of invasive breast carcinoma by the combination of morphometric data related to nuclear atypia score. In addition, we examined whether these algorithms were also applicable to predict biomarker status, including Ki-67 labeling index, ER, progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2), and triple-negative (TN) status. Materials and methods Specimens We analyzed H&E-stained pathology slides from representative cut surfaces of primary breast cancer tissues obtained from 173 patients who underwent total or partial mastectomy at the National Defense Medical College Hospital (Tokorozawa, Japan) between 2010 and 2013. All specimens were invasive ductal carcinomas (also referred to as invasive carcinomas of no special type). During the same period, 337 patients received surgical treatment for primary breast cancer at the hospital; however, 164 cases were excluded due to special histological subtypes, ductal carcinoma in situ, preoperative chemotherapy or hormone therapy, or the absence of suitable slides. All 173 selected slides were reviewed by three authors (TT, TW, and HTd) to confirm sufficient invasive carcinoma content and acceptable morphological preservation for histological evaluation. The slides were converted into whole-slide digital images at 200x mode using Nanozoomer Digital Pathology version 1.0.9 (Hamamatsu Photonics, Hamamatsu, Japan). We used 131 specimens (operated on in 2011–2013) as the test cohort (cohort 1) and 42 (operated on in 2010) as the validation cohort (cohort 2). Patient data We referred to pathology reports to collect clinicopathological and immunohistochemical information for 173 cases. These data included age, pT, pN, nuclear grade, nuclear atypia score, ER, PR, HER2, subtype, Ki-67 labeling index, and clinical outcomes. We evaluated ER and PR expression according to proportional score of the Allred score [11]. ER and PR were regarded as positive and negative, when the proportional scores were 2–5 (i.e., ≥ 1% positive) and 0–1 (< 1% positive), respectively, according to the 2020 American Society of Clinical Oncology/College of American Pathologists [ASCO/CAP] guideline [12]. In all cases, HER2 status was evaluated or re-evaluated according to the ASCO/CAP guidelines at that time [13]. All patients with HER2 IHC 2 + were determined to be HER2-positive or HER2-negative by fluorescence in situ hybridization. Based on immunohistochemical data, ER-negative, PR-negative, and HER2-negative tumors were defined as TN breast cancer. Ki-67 labeling index was classified as low, intermediate, and high when < 10%, 10–<20%, and ≥ 20%, respectively [14–16]. Data processing flow An overview of the image analysis pipeline in the present study is shown in Fig. 1. Briefly, from the whole slide images of 131 cohort 1 cases, regions of invasive carcinoma were selected by the pathologists, nuclear segmentation was performed using “nuclear phenotyper” of DenseNet V2 architecture on HALO AI (Indica Labs, Albuquerque, NM, U.S.A.), and then quality was manually confirmed. A histogram of the nuclear sizes of the three nuclear atypia scores is shown in Supplementary Fig. 1. With reference to this histogram, the nuclei 20 µm2 or smaller were excluded from segmentation because they might represent only a part of nuclear features or be of non-cancer cells. Annotation of 43,183 cancer cell nuclei larger than 20 µm2 was conducted. These imaging data were subjected for extraction of morphometric features of nuclear atypia using Python 3.7.4 and OpenCV 4.5.1-based algorithms. Overview of image analysis pipeline for morphometric assessment of nuclear atypia features Whole slide images of HE-stained breast carcinoma specimens were acquired at x200 magnification and analyzed at x400 magnification. Regions of invasive carcinoma were manually selected by pathologists. Nuclear segmentation was performed using HALO AI with manual inspection. Morphometric features were extracted using OpenCV-based algorithms, followed by case-level aggregation. Reference values for calculating Ax, Bx, and Cx systems are determined by area under curve (AUC) in the receiver operating characteristic (ROC) curves for Ax, Bx, and Cx by correlating with nuclear atypia score 3 and manual checking of actual image Then, the thresholds of Ax, Bx, and Cx that were the most associated with Grade 1/3, Ki-67, hormone receptor status were computed using the ROC analysis by the Youden index These thresholds optimization were done using Cohort 1, and validation were done in an independent cohort (Cohort 2) As three elemental findings of nuclear atypia, variations in nuclear size, or anisonucleosis (Ax), inhomogeneity in chromatin density (Bx), and prominent nucleoli (Cx) were selected based on the criteria of nuclear atypia score described in the General Rules for Clinical and Pathological Recording of Breast Cancer, 18th edition [9]. Key parameters for Ax, Bx, and Cx were content percentage of abnormally large nuclei, area ratio of highly blue-saturated structure in nucleus, and average size of large, circular and highly blue-saturated structures in nucleus, respectively. At first, basic reference values for Ax, Bx, and Cx were determined with reference to the area under the curve (AUC) associated with nuclear atypia score 3 and manual inspection of digital H&E images. Ax (%) was defined as the ratio of the number of nuclei larger than the reference size to the total number of nuclei larger than 20 µm2. The reference nuclear size for anisocytosis was calculated to maximize the AUC in the receiver operating characteristic curve (ROC) associated with nuclear atypia 3 by comparing the AUC among 28 different values of candidate reference nuclear sizes. Before analyzing the inhomogeneous chromatin density, we standardized the luminance of the images of annotated nuclei on HALO AI by color normalization with noise reduction and emphasized the contours of the structures using adaptive histogram equalization to improve visual accuracy (Supplementary Fig. 2). Image preprocessing parameters were fixed across all cases. Brightness normalization was performed to standardize global image intensity by setting brightness with a target mean intensity of 160 and by setting contrast with a standard deviation of 32. Subsequently, contrast-limited adaptive histogram equalization (CLAHE) was applied using OpenCV (version 4.5.1) with a clipLimit of 2.0 and a tileGridSize of 3 × 3 to enhance local contrast. Bx (%) was defined as the area ratio of highly blue-saturated structure to the total area of nuclei, and was calculated by dividing the total area of blue-saturated structure by the total area of nuclei larger than 20 μm². The blue-saturated structure in Bx was defined as the hematoxylin-stained basophilic structure with X-fold luminance or less (0 < X < 1.0) compared with the average luminance of all nuclei. Basic reference value of X for inhomogeneous chromatin density was determined by considering the balance between the AUC value for predicting nuclear atypia score 3 at 33 points of candidate X values in the ROC curves and manual inspection of H&E-stained nuclear images. Cx (µm2) was defined as the average size of large, circular and highly blue-saturated structure in nuclei. For the evaluation of Cx, the following three criteria were calculated: (1) blue-saturated structure: area of blue structure with X’-fold luminant or less (0 < X’ < 1.0) compared with the average luminance in the total nuclei; (2) circularity of each structure more than the reference value of circularity (circularity = 4π* area/[contour length]²); and (3) structure larger than the reference value of nucleolar size. The values of X’, circularity, and nucleolar size were calculated by considering the balance of AUC value in the ROC for predicting nuclear atypia score 3 and manual inspection of digital H&E images of nucleoli. The reference value of X’ was determined by considering the balance between the AUC value at 10 points of candidate X’ values in the ROC curves. The reference values of circularity and nucleolar size were determined by both considering the AUC values at several points of their values and manual checking whether they fit the extracted structures. Cx (µm2) was calculated as the average size of the structures that satisfied conditions (1), (2), and (3) with their optimal reference values. These morphometric features were then aggregated at the case level, followed by threshold optimization of Ax, Bx, and Cx for correlating with grade and biomarker statuses in Cohort 1, and validated in Cohort 2. Statistics In Cohort 1, after determination of basic reference values, we measured the mean values of three parameters (Ax, Bx, and Cx) for each of the 131 cases. Mean Ax, Bx, and Cx values were compared between the following subgroups: Grade 1 vs. Grade 2 vs. Grade 3, Ki-67 low vs. intermediate vs. high, lymph node metastasis negative vs. positive, ER negative vs. positive, PR negative vs. positive, HER2 negative vs. positive, TN vs. non-TN, recurrence positive (R) vs. negative (non-R), and age under 70 years (younger) vs. else (elder). Statistical significance was analyzed using an unrelated t-test (Welch’s t-test) with Benjamini-Hochberg false discovery rate method. In the 131 cases of Cohort 1, Ax, Bx, and Cx thresholds were calculated to achieve the optimal sensitivity and specificity for each clinicopathological and biomarker subgroup, i.e., Grade 3, Grade 1, Ki67-high, Ki67-low, ER-negative, PR-negative, HER2 positive, and TN tumors, using the Youden method in ROC curves. To predict each grade and biomarker status more effectively, the optimal coefficients of the simultaneous linear equations for Ax, Bx, and Cx were calculated using a multivariable logistic regression model analysis. Owing to multicollinearity issues, several candidate values were sampled using a computer, and two types of combination of coefficients with the maximum AUC of the ROC curve were adopted as formulae Dx and Ex. Then, we tested 42 specimens for validation (Cohort 2) for prediction of grades and biomarker statuses using Dx, and Ex cutoff values that were used for Cohort 1 without modification. Results Determination of reference values for elements of nuclear atypia: nuclear size, inhomogeneous chromatin density, and prominent nucleoli According to the ROC curve analyses, for the evaluation of Ax, the basic reference value of nuclear size associated with Grade 3 was set at 47 µm2 (Supplementary Fig. 3). For the evaluation of Bx, the reference value of the luminance of the blue-saturated structure was set at 0.92 times or less of the average luminance in the nuclei (Supplementary Fig. 4). For the evaluation of Cx, the reference values of three parameters were set: (1) blue-saturated structures with 0.87 times or less average luminance in the nuclei (Supplementary Fig. 5), (2) circularity of each structure greater than 0.65, and (3) blue structures larger than 1.15 µm2. Figure 2 shows 14 nuclei and nucleoli with blue-saturated structures, colored orange, that were assumed to be inhomogeneous chromatin compaction, and blue round structures, edged green, that were assumed to be prominent nucleoli (Fig. 2). Example images of carcinoma cell nuclei and nucleoli. (upper) H&E-stained images. Blue-stained nuclear structures, other than the nuclear membrane, are regarded as representing inhomogeneous chromatin compaction or prominent nucleoli. (lower) Extracted inhomogeneous chromatin was colored orange, while extracted prominent nucleoli are edged green Clinicopathological implication of anisonucleosis, inhomogeneous chromatin density, and prominent nucleoli The clinicopathological information and immunohistochemical profiles of the tumors in cohorts 1 and 2 are presented in Table 1.. Table 1Clinicopathological and biomarker profiles of two cohorts of breast cancer specimensLabeled groupNumber of specimens (%)Cohort 1, operated in 2011–2013 (n = 131)Cohort 2, operated in 2010 (n = 42)Age (years) < 7099No data ≧7032No dataLymph node metastasis pN090No data pN1-241No dataNuclear grade Grade 13519 Grade 23313 Grade 36310Nuclear atypia score Score 1107 Score 25925 Score 36210ER (Proportional score) Negative (0-1)300 Positive (2-5)10142PR (Proportional score) Negative (0-1)426 Positive (2-5)8936HER2 Negative12031 Positive113Subtype Triple negative (TN)24No data Others107No dataKi67 labeling index Low (< 10%)2316 Intermediate (10–< 20%)3714 High (≧20%)7111Recurrence (in 8 years) Yes8No data No123No dataNo patient received neo-adjuvant chemotherapy Clinicopathological and biomarker profiles of two cohorts of breast cancer specimens No patient received neo-adjuvant chemotherapy Figure 3 shows box plots comparing mean Ax, Bx, and Cx between the subgroups of each labeled group. Mean Ax was significantly associated with high nuclear grade (grades 3 vs.2 vs. 1: q = 0.0044), high Ki-67 levels (high vs. intermediate vs. low: q = 0.02), and ER negative (q = 0.014). Similarly, mean Bx was significantly associated with nuclear grade 3 (vs. 2 vs. 1, q = 0.0048), high Ki-67 levels (vs. intermediate vs. low, q = 0.0025), ER negative (q = 0.016), and PR negative (q = 0.034). Mean Cx was also positively associated with high nuclear grade 3 (vs. 2 vs. 1: q = 0.0002), high Ki-67 levels (vs. intermediate vs. low: q = 0.0033), ER negative (q = 0.00074), PR negative (q = 0.0057), and TN status (vs. non-TN, q = 0.0014). Mean values of Ax, Bx, and Cx tended to be higher in HER2-positive tumors than in HER2-negative tumors, but the differences were not statistically significant. The box plots of Ax, Bx, and Cx in clinicopathological and biomarker subgroups in Cohort 1. Differences in mean values of Ax, Bx, and Cx between the subgroups in each labeled group were tested with Welch’s t-test and were corrected using the Benjamini–Hochberg false discovery rate (FDR) method. Numbers indicated in the figure are adjusted p values (q values) Ki67 high (H): ≧20%, intermediate (M): 10–<20%, low (L): <10% ER estrogen receptor, PR progesterone receptor, TN triple negative, R recurrence Table 2. shows the cutoff values, sensitivity, and specificity of Ax, Bx, and Cx for stratifying each Grade, Ki-67, ER, PR, HER2, and TN subgroups. Figure 4 presents their respective ROC curves with AUC and 95% confidence intervals. When one of Ax, Bx, and Cx met the cutoffs, the sensitivities and specificities were 52.1–100% and 34.3–85.3%, respectively, to detect the subgroups of Grade 3, Grade 1, Ki-67 high, Ki-67 low, ER negative, PR negative, HER2 positive, and TN (Table 2.B). Table 2Sensitivity and specificity of Ax (anisonucleosis), Bx (inhomogeneous chromatin density), and Cx (prominent nucleoli) and their combination that are the most associated with grade and biomarker subgroups in Cohort 1 (n = 131)A. Optimal cutoff of Ax, Bx, and Cx indices for prediction of subgroupSubgroupCutoff valueAx (%)Bx (%)Cx (µm2)Grade 3 > 12.8 > 20.7 > 2.14Grade 1 < 15.6 < 20.5 < 2.03Ki67 high (≧20%) > 7.55 > 20.7 > 2.03Ki67 low (< 10%) < 3.01 < 20.7 < 2.26ER negative (PS 0–1) > 11.5 > 20.8 > 2.30PR negative (PS 0–1) > 11.5 > 20.8 > 2.24HER2 positive > 6.86 > 20.8 > 2.20Triple-negative > 11.5 > 19.8 > 2.24B. Sensitivity and specificity of Ax, Bx, and Cx for subgroup (n = 131)SubgroupAxBxCxSe (%)Sp (%)Se (%)Sp (%)Se (%)Sp (%)Grade 358.775.054.085.352.476.5Grade 191.442.791.450.077.164.6Ki67 high (≧20%)73.261.752.185.067.666.7Ki67 low (< 10%)52.284.310043.595.734.3ER negative (PS 0–1)80.062.460.075.263.383.2PR negative (PS 0–1)64.360.752.476.452.478.7HER2 positive90.943.363.671.763.668.3Triple-negative79.259.870.858.970.877.6C. Cutoff values, Sensitivity and specificity of Dx and Ex for subgroup in Cohort 1SubgroupDxExCutoff valueSe (%)Sp (%)Cutoff valueSe (%)Sp (%)Grade 3 > 29.576.270.6 > 32.977.866.2Grade 1 < 29.988.660.4 < 34.888.658.3Ki67 high (≧20%) > 28.678.970.0 > 32.974.668.3Ki67 low (< 10%) < 30.491.350.9 < 34.887.052.8ER negative (PS 0–1) > 31.776.781.2 > 37.373.380.2PR negative (PS 0–1) > 31.759.580.9 > 37.357.179.8HER2 positive > 31.481.869.2 > 37.472.773.3Triple-negative > 32.075.080.4 > 38.166.782.2ER estrogen receptor, PR progesterone receptor, PS proportional score, Se sensitivity [%], Sp specificity [%] Sensitivity and specificity of Ax (anisonucleosis), Bx (inhomogeneous chromatin density), and Cx (prominent nucleoli) and their combination that are the most associated with grade and biomarker subgroups in Cohort 1 (n = 131) ER estrogen receptor, PR progesterone receptor, PS proportional score, Se sensitivity [%], Sp specificity [%] The ROC curves for Ax, Bx, and Cx of nuclear atypia features and Dx and Ex values to predict Grade 1/3, Ki-67 low/high, ER-negative, PR-negative, HER2-positive, and triple negative statuses The Dx and Ex values were computed by the sum of Ax, Bx, and Cx, that were calculated as thresholds for each subgroup (Table 2.)A and multiplied with their respective coefficients as follows: Dx = 0.10*Ax + 0.45*Bx + 9.52*Cx. Ex = 0.27*Ax + 0.06*Bx + 14.2*Cx, where the cutoff values of Dx and Ex for subgroups were estimated from the maximum AUC of ROC curve analyses. When the cutoff value for Dx was set at 29.5, the sensitivity and specificity for grade 3 were 76.2 and 70.6%, respectively. When the cutoff value for Ex was set at 32.9, these values were 77.8 and 66.2%, respectively. (Table 2.C and Fig. 4). In Cohort 1, Dx and Ex models combining Ax, Bx, and Cx values could achieve 0.7 or higher AUC, with high sensitivity and specificity, for the prediction of nuclear grades 3 and 1, Ki-67 high and low, ER negative, PR negative, and HER2 positive (Fig. 4, Table 2.C). In Cohort 2, no data were available for validation of the ER-negative subgroup because of the absence of these samples. The sensitivities and specificities detecting Grade 3, Grade 1, Ki-67 high, and Ki-67 low subgroups were 63.6–100% and 4.3–62.5%, respectively (Data not shown). As combinations of Ax, Bx, and Cx, Ex for Grade 3 and Dx for HER2 positive showed relatively high accuracy, but these models showed much lower sensitivity for detection of high-grade tumors, e.g., Grade 3 and Ki-67 high, and much lower specificity for detection of low-grade tumors, e.g., Grade 1 and Ki-67 low (Table 3). Table 3Sensitivity and specificity of Dx and Ex for subgroup in Cohort 2 (n = 42)SubgroupDxExSe (%)Sp (%)Se (%)Sp (%)Grade 350.087.570.068.8Grade 189.530.478.939.1Ki67 high (≧20%)45.574.245.561.3Ki67 low (< 10%)87.526.975.034.6ER negative (PS 0-1)No data90.5No data83.3PR negative (PS 0-1)16.786.116.783.3HER2 positive66.789.733.384.6 Sensitivity and specificity of Dx and Ex for subgroup in Cohort 2 (n = 42) Discussion In this study, we developed a simple morphometric algorithm to assess nuclear atypia in breast cancer using 131 H&E-stained slides. The algorithm was designed based on the criteria for histological grading outlined in the WHO Classification of Breast Tumors, 5th edition [3], and for nuclear grading in the General Rules for Clinical and Pathological Recording of Breast Cancer, 18th edition [9]. We focused on three key features of nuclear atypia: variation in nuclear size (Ax), inhomogeneous chromatin density (Bx), and the presence of prominent nucleoli (Cx), as summarized in Table 4.. By applying the morphometric algorithm combining the measurements of these three elements, we examined if morphometric features related to nuclear atypia alone could predict nuclear grade, which contains both findings of nuclear atypia and number of mitotic figures: Grade 3 was supported when Ax > 12.8, Bx > 20.7, and Cx > 2.14, whereas Grade 1 was supported when Ax < 15.6, Bx < 20.5, and Cx < 2.03. Table 4.Summary of our simple morphometrics for quantifying nuclear atypia features of invasive ductal carcinomas from computer vision of H&E images Summary of our simple morphometrics for quantifying nuclear atypia features of invasive ductal carcinomas from computer vision of H&E images Studies attempting to determine simple morphometric parameters in pathological images have been conducted for various cancers. In a study measuring nuclear size using electron microscopy on 56 breast fine-needle aspiration cytology specimens (40 cases of breast cancer and 16 cases of benign disease), all specimens with a mean nucleolar area of 2.5 µm2 or more, or a mean maximum nucleolar diameter of 2.5 μm or more, were malignant [17], which might be relevant with the present results of the reference value of nucleolar size (2.14 µm2) to detect Grade 3 (Table 2.). Andrea et al. found that preoperative specimens of osteosarcoma with both nuclei larger than 42.5 µm2 and the circularity larger than 0.7 predicted worse prognosis by analyzing 56 cases [18]. Although cancer type was different, their nuclear-size threshold was similar to the reference value (47 μm²) for Ax in the present study (Supplementary Fig. 2). In prostate cancer, the morphometric features of nuclei, including size, shape, and texture, are clinically relevant to histopathological features [19]. The major impact of this exploratory study is that objective quantification of nuclear atypia parameters was able to predict not only nuclear grade but also certain ranges of Ki-67 and hormone receptor status. To date, it is insufficient to evaluate the expression status of these molecules using only H&E-stained cancer specimens, and the information and costs of immunohistochemistry are needed [20]. In addition, inter-institutional variation in the analytical process, such as the concentration of reagents, staining methods, and reaction time, could give rise to differences in nuclear features on H&E staining. We attempted to construct an improved algorithm (Supplementary Fig. 2) that standardizes the luminance between slide images and emphasizes the contours of structures by adaptive histogram equalization [21]. Application of such an algorithm may be helpful in the future for daily practice and multi-institutional clinical trials that use computer vision of pathological slide images by equalizing H&E staining between institutes and samples [22]. This study had several implications. Quantifying the nuclear structures of cancer cells by analyzing H&E-stained images is expected to be of more statistical evidence. Although cancer cell nuclei show obvious morphological changes in H&E-stained images, few studies have investigated the link between these changes and cell function [23]. Previous studies have indicated that anisonucleosis and nuclear morphological changes are related to DNA polyploidy/aneuploidy or abnormalities in nuclear structural proteins, perinuclear structures, cytoplasmic factors, and intracellular transport systems [24]. Recently, some studies have revealed that a nuclear pore complex, including ELYS, which regulates nuclear size [25], seems to be involved in transcription and chromatin organization, DNA repair, aging, viral infection, cancer, and other diseases [26]. In another example, the nucleolar size of TN was demonstrated to be larger than that of non-TNs in the present study. These nucleolar findings may be morphometric characteristics of TN (Fig. 3). When conducting such basic research, it would be helpful to employ objective methods to evaluate the morphology of cancer cell nuclei. In our study, Ki67-high had a strong positive relationship with Bx and Cx and ER negative, PR negative, and TN with Cx. These findings may reflect the etiological molecular events in the nuclei of each type of breast cancer. This study had some limitations. First, reproducibility of the Dx and Ex models for prediction of grade and biomarkers were not very high for Cohort 2. We considered that the discrepancy in Dx and Ex results between these two cohorts resulted from a small number and skewed characteristics of Cohort 2. Cohort 2 represented a limited-spectrum validation cohort rather than a comprehensive external validation set. Accordingly, the generalizability of our findings acquired in Cohort 1 was limited. There are realistic challenges of generalizing morphometric algorithms across cohorts. Broader multi-institutional validation, ideally incorporating cross-validation and external datasets, including those acquired with multiple scanners and staining conditions will be required in future studies. Second, subgroups in the test and validation cohorts were classified based on surgical pathology reports, in which different attending pathologists had individually diagnosed the patients. Third, at present, the computed values indicated in our algorithms are no more than forecast figures based on only 173 breast cancer cases. Fourth, there might have been sampling bias when selecting one pathological image per specimen. With the optimal thresholds of Dx and Ex, the sensitivity and specificity for predicting Grade 3 were still 76.2–77.8% and 66.2–70.6%. respectively, with AUC of 0.74–0.75 (Table 2.C; Fig. 4). For other biomarkers, the sensitivity and specificity were 57.1–91.3% and 50.9–81.2%, respectively, with AUC of 0.70–0.82. Nonetheless, the present algorithm for nuclear atypia measurement is considered potentially useful as a simple and quantitative tool for predicting grade 3 or other biomarker statuses of invasive breast carcinoma. We believe that the present work on rewriting the pathological guideline criteria into certain digital values for computer vision could be a springboard for the impending epoch of AI-assisted clinical medicine. Conclusions The simple algorithm, which was composed of morphometrically quantified features of nuclear atypia elements, i.e., anisocytosis, inhomogeneous chromatin density, and prominent nucleoli, of invasive breast carcinoma, and may potentially be able to predict nuclear grade, Ki-67, and hormone receptor statuses. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements We thank Taro Kuroda, M.D.; Taiyo Hasegawa, M.D.; Mariko Saito, M.D; Sinya Yoshimatsu, M.D.; Wataru Yano, Ph.D.; and Susumu Matsukuma, M.D., Ph.D., National Defense Medical College, for their valuable technical advice. The authors sincerely appreciate language editing by Editage. Author contributions TT and HTd designed this study, TT, TW, YY, TE, and TY collected clinical and pathological data, TT, TW, and KK conducted data analyses. TT and HTd drafted original manuscript, and HTj, YK, and HU substantively revised the manuscript. All authors have read and approved the manuscript. Funding Open Access funding provided by National Defense Medical College. This study was supported in part by Defense Medicine Basic Research Program B. Data availability The data analyzed during the current study are available from the corresponding author on reasonable request. Declarations Conflict of interest Keita Kouzu received honorarium from Indica Labs, Hitoshi Tsuda received scholarship donation from Chugai, research grant from EPS, and consultant fee from PCL Japan. The remaining authors declare no conflict of interest. The disclosed commercial affiliations did not influence the study design, data analysis, or interpretation of results. Ethical approval All procedures involving human participants performed in this study were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. This study was performed in accordance with the Declaration of Helsinki and approved by the Institutional Review Board of the National Defense Medical College (registration No. 4558, 4689). Informed consent An opt-out approach was used with the disclosure of website. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The data analyzed during the current study are available from the corresponding author on reasonable request.