Moran’s index habitat and peritumoral radiomics based on digital breast tomosynthesis for predicting Ki-67 status in breast cancer Authors: Jiaai Gong, Zijian Zhuang, Shutong Liu, Lulu Xu, Huihui Liu, Haitao Zhu, Dongqing Wang, Lirong Zhang Published: 2026 ABSTRACT AbstractBackgroundThe heterogeneity of breast cancer is a key factor influencing treatment decisions and patient prognosis. As a crucial cell proliferation marker, the expression level of the Ki-67 protein is closely associated with tumor aggressiveness, recurrence risk, and prognosis. The goal of this study was to develop and validate an integrated radiomics model based on digital breast tomosynthesis (DBT) that incorporates both habitat subregions and peritumoral features to preoperatively predict Ki-67 expression status in breast cancer.MethodsA total of 146 patients with invasive breast cancer who underwent preoperative DBT were retrospectively enrolled in the study and divided into training (n=102) and testing (n=44) cohorts. Intratumoral regions were manually segmented, whereas peritumoral regions (2, 4, and 6 mm) were automatically delineated using mask padding. Moran’s index clustering was used to divide the tumors into four habitat subregions. Radiomics features were extracted from the intratumoral regions, habitat subregions, and peritumoral regions. Following feature selection, radiomics models were established using machine learning algorithms. A combined model was subsequently developed by integrating habitat subregions and peritumoral radiomics features with clinical factors. The area under the curve (AUC) was used to assess the performance of radiomics models.ResultsIn predicting Ki-67 expression status, the rbf_support vector machine (SVM) model, constructed from habitat radiomics features, outperformed the other models, with AUC values of 0.841 and 0.820 in the training and testing cohorts, respectively. The combined model slightly increased the AUC values to 0.849 and 0.852, respectively.ConclusionsA combined model integrating habitat, peritumoral radiomics features, and clinical factors demonstrated optimal accuracy in predicting Ki-67 expression levels in breast cancer; this model may help clinicians develop personalized treatment plans for breast cancer patients.Keywords: Radiomics, habitat imaging, peritumoral, digital breast tomosynthesis (DBT), breast cancer Introduction Breast cancer is one of the malignant tumors with the highest incidence rate among women worldwide (1). Importantly, its biological behaviors exhibit considerable heterogeneity, which directly affects treatment decision-making and assessments of patient outcomes. Ki-67 is an important marker of cell proliferation activity (2), and its expression level is closely related to tumor aggressiveness, treatment response and patient outcomes (3-5); patients with high levels of Ki-67 have worse outcomes but respond better to chemotherapy (6). In addition, in patients receiving neoadjuvant endocrine therapy, early changes in the level of Ki-67 can be used to determine whether the patient will achieve therapeutic benefits, and its residual value can be used to determine whether additional therapy will subsequently be needed (7). Currently, the preoperative detection of Ki-67 relies mainly on immunohistochemistry (IHC) analysis (8), which is usually performed by obtaining tumor tissue through core-needle aspiration biopsy (CNB). However, this approach has several limitations. For example, biopsy sampling is not representative of the entire tumor and thus may lead to missing spatial heterogeneity (9). In addition, the level of Ki-67 can change dynamically during treatment (10), and the results from a single assay cannot be used for dynamic monitoring of the Ki-67 status, limiting the precise formulation of neoadjuvant chemotherapy regimens. Therefore, the development of a noninvasive and comprehensive preoperative prediction technique is highly clinically important. Radiomics, a technique that can transform medical imaging data into high-dimensional features (11), has provided new ideas for the noninvasive assessment of tumor phenotypes (12). Unlike traditional radiomics, habitat analysis employs various algorithms to partition tumors into multiple voxel clusters with similar characteristics, thereby more effectively capturing intratumoral heterogeneity (ITH) (13,14). Relatedly, Moran’s index is a statistical measure of spatial autocorrelation in spatial data (15), reflecting tumor spatial heterogeneity. Previous studies have explored the predictive value of magnetic resonance imaging (MRI)-based habitat radiomics for the human epidermal growth factor receptor-2 (HER2) status in breast cancer (16), and similar approaches have been applied to esophageal cancer, colorectal cancer, and other tumor types (17,18). In addition to the tumor itself, the peritumoral microenvironment often influences tumor behavior and proliferation. Studies have demonstrated that angiogenesis, immune infiltration, and stromal remodeling in the peritumoral region may be closely associated with high Ki-67 expression (19-21). Therefore, a comprehensive evaluation of both intra- and peritumoral characteristics could facilitate a more complete characterization of the biological behavior of the tumor. In this study, digital breast tomosynthesis (DBT) images were used for radiomics analysis. Compared with traditional mammography, DBT can provide high-resolution, three-dimensional images that can better depict the tumor morphology and surrounding tissue structure more clearly, providing a data basis for imaging histology analysis (22,23). However, the potential of DBT-based habitat radiomics for assessing Ki-67 expression remains understudied. Based on of the above background, this study aimed to provide a novel approach for the preoperative noninvasive assessment of the Ki-67 expression status in patients with breast cancer, thereby optimizing personalized treatment decision-making. In this study, DBT radiomics features of the intratumoral, peritumoral, and habitat subregions were used to construct a machine learning model for predicting the Ki-67 status in breast cancer. By comparing the predictive performance across different peritumoral expansion ranges (2, 4, and 6 mm), the optimal size of the peritumoral region was determined. We present this article in accordance with the TRIPOD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0047/rc). Methods Patients The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Institutional Review Board of The Affiliated Hospital of Jiangsu University (No. KY2025K0801), and individual consent for this retrospective analysis was waived. A total of 146 patients who were diagnosed with invasive breast cancer at The Affiliated Hospital of Jiangsu University between July 2018 and September 2023 were retrospectively enrolled. The inclusion criteria were as follows: (I) postoperative pathological confirmation of invasive breast cancer; (II) DBT performed within two weeks prior to surgery; and (III) complete postoperative pathological data, including the Ki-67 status. Of note, the Breast Imaging Reporting and Data System (BI-RADS) category on DBT was not used as a selection criterion. Two patients with an initial DBT assessment of BI-RADS 3 were included because they had pathologically proven malignancy after supplementary imaging evaluation. The exclusion criteria were as follows: (I) lesions on DBT images that did not manifest as masses but were present in other forms; (II) lesions that were not entirely visible because of the depth of their location; (III) poor image quality; (IV) history of breast surgery; and (V) history of other malignant tumors (Figure 1). Flow chart for the selection of participants. DBT, digital breast tomosynthesis. DBT examination In this study, DBT was performed using a mammography machine (MAMMOMAT Inspiration, Siemens Healthcare, Germany). Automatic exposure mode was adopted for 3D data acquisition to obtain images of the patient’s bilateral breasts in mediolateral oblique (MLO) and craniocaudal (CC) positions. For DBT imaging, X-ray tube was rotated in a 15° arc, with projection angles ranging from 40° to 65° based on patient anatomy. Each field of view captured 15 to 30 projections. For all images acquired, the layer thickness was reconstructed using the filtered back projection (FBP) algorithm; finally, the layer thickness of the DBT images was controlled to 1 mm. In our study, the median average glandular dose (AGD) at the CC position was 1.79 mGy [interquartile range (IQR), 1.43–2.10 mGy], and it was 1.78 mGy (IQR, 1.55–2.26 mGy) at the MLO position. The DBT scanning parameters are detailed in Appendix 1. Data collection and analysis Clinical and pathological data: the clinical and pathological data of the patients were obtained from the electronic medical records system of the institution. Pathological data were obtained by two pathologists who diagnosed postoperative tissue samples. If the diagnoses were inconsistent, the final decision was made through consultation. The pathologists examined the stained sections under a microscope and counted the number of tumor cells with positive staining. Ki-67 proliferation levels were expressed as the percentage of malignant cells exhibiting positive staining. In accordance with the St. Gallen international expert consensus guidelines, a threshold of 14% was used to classify the Ki-67 level; specifically, a Ki-67 proliferation ≥14% indicated high expression, while a proliferation <14% indicated low expression (7). Radiographic features: in accordance with the BI-RADS Version 5 classification standards established by the American College of Radiology (ACR) (24), the radiographic features of the patients were evaluated and analyzed by two radiologists with at least 5 years of experience in diagnosing breast disease; if the radiologists disagreed on the diagnosis, a decision was made by consensus. The data collected in this study included the tumor density, maximum tumor diameter (MTD), microlobulated margins, spiculate margins, indistinct margins and calcification. With regard to the margins of the lesions, as DBT rather than conventional two-dimensional mammography was used for margin assessment, no lesions were classified as having obscured margins. All the observed calcifications were classified as suspicious calcifications according to BI-RADS. The presence of suspicious calcifications was recorded as a binary feature (present/absent). Cases with amorphous, coarse heterogeneous, fine polymorphic, or fine linear branching morphologies, as well as those with regional, clustered, linear, or segmental distributions, were classified as suspicious calcifications. Volume of interest (VOI) segmentation Breast lesions were manually segmented on the CC view using 3D-Slicer software (version 5.6.2) by a radiologist who was blinded to the initial radiological reports and pathological findings of the patients. Intratumoral regions were obtained by outlining the tumor along its margins layer-by-layer construct a VOI for the entire tumor. The MLO view was used as a reference to confirm the boundaries. The segmentations were subsequently reviewed by a senior radiologist. To verify the reproducibility of feature extraction, 30 patients were randomly selected for resegmentation by another senior radiologist and the same radiologist one month later. A sample size of 30 is considered adequate to obtain a robust estimate of the intraclass correlation coefficient (ICC) with a reasonable confidence interval (CI), in accordance with common practice in radiomics methodology studies (25). Peritumoral and habitat subregions We used the mask padding toolkit to automatically expand the border of the manually delineated VOI outward 2, 4, and 6 mm, generating the corresponding peritumoral regions. The generated peritumoral mask was subsequently restricted to the boundaries of the mammary tissue to avoid including nonbreast tissue (such as air or the pectoral muscle) when the peritumoral region near the skin or chest wall was expanded. Any voxels extending beyond these boundaries were excluded from the analysis. The local Moran’s index method was employed to generate subregions within these regions of interest. This index, which originates from spatial statistics and geographic information science, serves as a statistic for assessing the autocorrelations in spatial data (26). In a medical image analysis, this method effectively quantifies the spatial distribution patterns of pathological tissue features. We constructed a weight matrix W based on neighborhood relationships to capture the local spatial relationships between voxels within a region. This weight matrix was calculated based on the spatial distance between voxels, and the Euclidean distance was used to determine whether the voxels were neighbors. For example, if the distance between two voxels was less than or equal to the diagonal distance 3 of their octagonal neighborhood, they were defined as neighbors, and the weight wij was set to 1; otherwise, wij was set to 0. Based on these definitions, the weight matrix W is a symmetric sparse matrix that represents the spatial relationships among local voxels within the region. On the basis of the local Moran’s index and voxel attribute values, the tumor was divided into four habitat subregions: A positive local Moran’s index (I >0) indicates a positive spatial correlation between the voxel and its neighboring voxels. By also incorporating the relationship between the voxel value and the mean, the voxel can be further said to be within a high-high (HH) or a low-low (LL) region. If I <0, indicating a negative spatial correlation between the voxel and its neighbors, the voxel is said to be within a high-low (HL) or a low-high (LH) region on the basis of the relationship between the voxel value and the mean (Figure 2). We did not set a minimum region size threshold; therefore, subregions with fewer voxels were also retained for subsequent radiomic feature extraction, and each patient had four habitat subregions. Specifically, the local Moran’s index statistic for spatial autocorrelation in our study was calculated as follows: Example of ROI segmentation on DBT images. The patient is a 57-year-old woman with a 2.5 cm mass in the left breast and high Ki-67 expression. (A) The original DBT image. (B-D) The green masks are the intratumoral ROIs, and the yellow area in (C-E) represent expansions of 2 mm, 4 mm, and 6 mm peritumoral ROIs. (F) The tumor was subdivided into habitat subregions 1 (purple), 2 (orange), 3 (blue), and 4 (red). DBT, digital breast tomosynthesis; ROI, region of interest. where xi is the attribute value of voxel i, xj is the attribute value of voxel j, x^ is the mean value of the attributes within the region, wij is the spatial weight between voxels i and j, and M2 is defined as follows: where n is the total number of voxels within this region. Radiomic feature extraction and selection Prior to the radiomics feature extraction process, the images were subjected to resampling and grayscale quantization for normalization. We employed the open-source software PyRadiomics (version 3.0.1) to extract radiomics features in accordance with the image biomarker standardization initiative (27), as PyRadiomics is a well-validated and standardized tool that ensures the reproducibility and compliance of our feature extraction process. Features were extracted from the intratumoral region, the three peritumoral regions, and the four habitat subregions; these features included first order (n=18), shape (n=14), gray-level dependence matrix (GLDM) (n=14), gray-level co-occurrence matrix (GLCM) (n=24), gray-level run-length matrix (GLRLM) (n=16), gray-level size zone matrix (GLSZM) (n=16), and neighboring gray-tone difference matrix (NGTDM) features (n=5). The ICC1 and ICC2 were calculated to assess the reproducibility of the manually extracted radiological features; only those with both an ICC1 and an ICC2 ≥0.75 were retained. We used the PixelMedAI platform (written in Python 3.8) for feature selection. All the extracted features were normalized via z-score transformation. We implemented a three-step process exclusively within the training cohort to reduce dimensionality and identify robust features. First, the Student’s t-test was used to identify features potentially associated with the Ki-67 expression level; features with a P value <0.05 were retained for the next step. To mitigate collinearity, we subsequently assessed the correlations between features using Pearson’s correlation coefficient. If the correlation coefficient between any two features exceeded 0.9, only one of the features was retained. Finally, feature selection was performed using a nested 10-fold cross-validation strictly within the training cohort. For each fold, least absolute shrinkage and selection operator (LASSO) with inner cross-validation was applied exclusively to the training subfolds to select features and optimize λ. A flowchart of the study is shown in Figure 3. Overview of the workflow. First, collect data and manually delineate regions of interest. Based on this, define the peritumoral regions and habitat subregions. Subsequently, extract radiomics features from each region. After feature selection and dimensionality reduction, construct predictive models. Finally, evaluate the models’ performance. DBT, digital breast tomosynthesis; DCA, decision curve analysis; LR, logistic regression; ROC, receiver operating characteristic; ROI, region of interest; SVM, support vector machine. Prediction model construction and evaluation The univariate and multivariate logistic regression (LR) analyses were employed to identify independent predictors associated with the Ki-67 expression levels. Subsequently, a clinical model was constructed using LR. By employing LASSO for stringent feature selection, we constructed predictive models using four machine learning algorithms: LR, support vector machine (SVM) with linear (linear_SVM) and radial basis function (rbf_SVM) kernels, and light gradient boosting machine (LightGBM). The optimal parameters for the classification model were selected in the training cohort using fivefold cross-validation. In this study, multiple prediction models for different regions were established: traditional radiomics model (Intra), habitat models (Habitat), and peritumoral models (Peri2mm, Peri4mm, and Peri6mm). Moreover, a combined model was constructed that integrated habitat features, peritumoral features, and clinical factors. The diagnostic efficacy of the machine learning models was evaluated in the testing cohort using the area under the operator characteristic curve (AUC) to assess the ability of the models to predict the Ki-67 expression status. The calibration of the models was analyzed using calibration curves supplemented by Hosmer-Lemeshow goodness-of-fit tests to verify the reliability of the models. In addition, decision curve analysis (DCA) was used to evaluate the clinical utility of our predictive models, specifically to assess their potential benefits in clinical contexts. Interpretability of the model We used Shapley additive explanation (SHAP) analysis to explain the machine learning model. Specifically, the SHAP method quantifies the contributions of each radiomics feature to the model predictions, allowing a clear understanding of their individual impacts (28). SHAP values help explain why a model made a particular prediction by showing the extent to which each feature drives that prediction relative to a baseline (typically the average prediction across the dataset). Statistical analysis Continuous variables were compared with the t-test or the Mann-Whitney U test based on their distribution characteristics. Categorical variables were compared using chi-square (χ2) tests. The area under the curve (AUC) was calculated to assess model performance. The clinical value of each predictive model was assessed by DCA and calibration curve analysis. The DeLong test was used to compare differences in the receiver operating characteristic (ROC) curves. All the statistical analyses were performed using SPSS software (version 26.0). For all the statistical analyses, P<0.05 indicated statistical significance. Results Patient characteristics This study ultimately included 146 patients with invasive breast cancer who were randomly assigned to a training cohort or a testing cohort at a 7:3 ratio. The patient characteristics are reported in Table 1. A total of 102 patients (53.88±9.69 years) were included in the training cohort, among whom 69 (67.65%) had high Ki-67 expression and 33 (32.35%) had low Ki-67 expression, and 44 patients (52.84±10.87 years) were included in the testing cohort, among whom 32 (72.73%) had high Ki-67 expression and 12 (27.27%) had low Ki-67 expression. Differences in the demographic and clinicoradiological characteristics were not statistically significant between the training and testing cohorts (all P>0.05), indicating that the dataset partitioning was reasonable and that the two cohorts were comparable, allowing the subsequent construction and evaluation of the models. Table 1Demographic and clinicoradiological characteristics of the training and testing cohortsCharacteristicTraining cohort (n=102)Testing cohort (n=44)P valueAge (years)53.88±9.6952.84±10.870.567MTD (cm)1.90±0.081.95±0.120.788Ki-67 status0.542 High expression69 (67.65)32 (72.73) Low expression33 (32.35)12 (27.27)BI-RADS0.543 32 (1.96)0 (0.00) 4a9 (8.82)6 (13.63) 4b18 (17.65)10 (22.73) 4c48 (47.06)14 (31.82) 525 (24.51)14 (31.82)Calcification0.642 Absent80 (78.43)36 (81.82) Present22 (21.57)8 (18.18)Spiculate margin0.718 Absent52 (50.98)21 (47.73) Present50 (49.02)23 (52.27)Microlobulated margin0.071 Absent56 (54.90)17 (38.64) Present46 (45.10)27 (61.36)Mass density0.728 High density69 (67.65)31 (70.45) Isodensity33 (32.35)13 (29.55)Data are presented as mean ± standard deviation or n (%). BI-RADS, Breast Imaging Reporting and Data System; MTD, maximum tumor diameter. Demographic and clinicoradiological characteristics of the training and testing cohorts Data are presented as mean ± standard deviation or n (%). BI-RADS, Breast Imaging Reporting and Data System; MTD, maximum tumor diameter. Clinical model To identify independent predictors associated with the Ki-67 expression status and construct a clinical prediction model, we first performed univariable LR analysis. The results indicated that age, BI-RADS, MTD, and microlobulated margin (all P<0.05) were associated with the Ki-67 expression status (Table S1). These factors were subsequently incorporated into the multivariable LR analysis. The results revealed that only age and MTD (all P<0.05) were independent predictors of the level of Ki-67 expression. The clinical model constructed with these two factors yielded AUC values of 0.679 (95% CI: 0.562–0.784) and 0.729 (95% CI: 0.505–0.892) in the training and testing cohorts, respectively, indicating moderate predictive performance for the Ki-67 expression status (Table 2). Table 2Performance of the prediction models across the training and testing cohortsModelCohortsAUC (95% CI)AccuracySensitivitySpecificityClinicTraining0.679 (0.562–0.784)0.5880.5070.758IntraTraining0.776 (0.671–0.871)0.7160.7100.727Peri4mmTraining0.824 (0.723–0.899)0.7750.8260.667HabitatTraining0.841 (0.749–0.909)0.7940.7970.788CombineTraining0.849 (0.752–0.929)0.7840.7390.879ClinicTesting0.729 (0.505–0.892)0.7500.7500.750IntraTesting0.737 (0.559–0.898)0.6820.6560.750Peri4mmTesting0.745 (0.512–0.884)0.7500.8130.583HabitatTesting0.820 (0.678–0.936)0.6590.5940.833CombineTesting0.852 (0.674–0.975)0.8410.8130.917AUC, area under the curve; CI, confidence interval. Performance of the prediction models across the training and testing cohorts AUC, area under the curve; CI, confidence interval. Radiomics models construction and evaluation To investigate tumor spatial heterogeneity and construct effective predictive models, we first used local Moran’s index clustering to segment the tumor into four habitat subregions (HH, LL, HL, and LH regions), which corresponded to Habitat 1, Habitat 2, Habitat 3, and Habitat 4, respectively. Radiomics features were extracted from the intratumoral region, peritumoral regions, and habitat subregions. The intraobserver and interobserver reproducibility of the radiomics features were excellent. The number of features retained for each region was as follows: 98 for the intratumoral region, 97 for the 2 mm peritumoral region, 96 for the 4 mm peritumoral region, 94 for the 6 mm peritumoral region, and 380 for the habitat subregions (Habitat 1–4). Afterward, these features were subjected to a meticulous “three-step process” for feature selection. Finally, the intratumoral model retained 8 features, the habitat model retained 6 features, and the peritumoral models retained 5, 6, and 5 features at 2, 4, and 6 mm, respectively. Tables S2-S6 provide a detailed summary of the final features selected via LASSO and their corresponding LASSO coefficients. Predictive models were constructed using four machine learning algorithms: LR, linear_SVM, rbf_SVM, and LightGBM. In the traditional Intra model, the linear_SVM classifier had the best predictive performance, with AUC values of 0.776 (95% CI: 0.671–0.871) and 0.737 (95% CI: 0.559–0.898) in the training and testing cohorts, respectively. Among the peritumoral models, the performance of the LightGBM-based model constructed with the Peri4mm features was superior to that of the models constructed from the features in the other regions, with AUC values of 0.824 (95% CI: 0.723–0.899) and 0.745 (95% CI: 0.512–0.884) in the training and testing cohorts, respectively. Finally, the rbf_SVM classifier based on the habitat model had the best predictive performance, with AUC values of 0.841 (95% CI: 0.749–0.909) and 0.820 (95% CI: 0.678–0.936) in the training and testing cohorts, respectively. The detailed model specifications are summarized in Tables S7-S11. Thus, compared to other models, the habitat model displayed superior predictive performance. Compared with the habitat model alone, the combined model integrating the habitat and peritumoral features and clinical factors showed a slight improvement in the AUC, with values of 0.849 (95% CI: 0.752–0.929) and 0.852 (95% CI: 0.674–0.975) in the training and testing cohorts, respectively (Table 2). However, in the testing cohort, the DeLong test indicated that this difference was not statistically significant (P=0.62) (Figure S1). We plotted ROC curves for each predictive model across both cohorts (Figure 4). Calibration curve (Figure S2) analysis demonstrated good agreement between the combined model’s predicted and actual Ki-67 status, while DCA (Figure 5) indicated that the combined model provided a clinical net benefit. The ROC of prediction models (A) in the training cohort and (B) testing cohort. AUC, area under the curve; ROC, receiver operating characteristic. The DCA of prediction models (A) in the training cohort and (B) testing cohort. DCA, decision curve analysis. Subgroup analysis Given the differences in breast tissue structure among patients, we divided the testing cohort into a nondense group (ACR a–b, n=19) and a dense group (ACR c–d, n=25) based on the ACR BI-RADS breast density classification system, in order to assess the impact of breast tissue composition on the performance of the Combined model. The model achieved an AUC of 0.957 (95% CI: 0.866–1.000) in the nondense group and 0.778 (95% CI: 0.543–1.000) in the dense group (Figure S3). The discriminative ability of the model was better for nondense breasts than for dense breasts. Interpretation of the habitat model The SHAP analysis method was employed to interpret the rbf_SVM model in Habitat, which by quantifying the contribution of each feature to the model’s predictions. A beeswarm plot (Figure 6) was generated to illustrate the contribution of each radiomics feature to predicting the level of Ki-67 expression. As shown in Figure 6, the feature glcm_Idmn_h1 contributed most to the model’s prediction performance and is thus the most important feature for distinguishing the Ki-67 expression status; this feature indicates that the Habitat 1 subregion significantly influences the model’s predictive performance. Details of the SHAP analysis of the combined model are provided in Figure S4. SHAP analysis of the habitat model. Each point represents the SHAP value of a feature for a patient, with point color indicating feature value magnitude, ranging from low (blue) to high (red). These six radiomic features are ranked from highest to lowest absolute contribution, with the top feature having the most significant impact on prediction outcomes. SHAP, Shapley additive explanation. A SHAP waterfall plot (Figure 7) was generated to explain the prediction results for an individual patient. As shown in Figure 7A, the model output [f(x)] for this patient was 1, which is greater than the baseline value of 0.67, leading to a classification of high Ki-67 expression. Among the features, glcm_Idmn_h1 contributed substantially to this prediction. Conversely, for the patient shown in Figure 7B, f(x) was 0, which was below the baseline of 0.67, resulting in a classification of low Ki-67 expression. In this patient, the feature glcm_Imc1_h4 contributed negatively to the prediction of low Ki-67 expression. Waterfall plots illustrate individual patient assessments. (A) Patient A exhibits high Ki-67 expression, while (B) Patient B shows low Ki-67 expression. The baseline value represents the average of all sample predictions (the merged training and testing cohorts), while the model output value indicates the final prediction for the sample. Arrow colors denote whether a feature positively (red) or negatively (blue) influences the prediction outcome. The numerical values indicate each feature’s contribution to the final prediction, representing its SHAP value. SHAP, Shapley additive explanation. Additionally, we generated a heatmap (Figure 8) to illustrate the predictive performance of the habitat model for the Ki-67 expression status. By comparing the model-predicted probability of Ki-67 expression with the ground truth, the heatmap hows that our model effectively discriminates between patients with high and low Ki-67 expression, indicating favorable predictive accuracy and strong generalizability. Heatmap showing the rbf_SVM model (habitat model) predictions for Ki-67 expression status (high expression vs. low expression) across all datasets. Each row represents a variable, each column represents a sample, with variable values indicated by color. Variables include rbf_SVM predicted value, spiculated margin, MTD, mass density, calcification, microlobulated margin, BI-RADS, and age. Color gradients indicate the range of variable values, with categorical variables represented by distinct colors for each category. BI-RADS, Breast Imaging Reporting and Data System; MTD, maximum tumor diameter; SVM, support vector machine. Discussion In this study, we developed and validated a machine learning model based on DBT images combined with radiomics features from habitat subregions and peritumoral regions to predict the Ki-67 expression status in patients with breast cancer. Our findings indicate that the habitat analysis effectively captured spatial heterogeneity within the tumor, whereas peritumoral regions (particularly within a 4 mm range) contain biologically relevant information associated with the tumor microenvironment. The rbf_SVM classifier constructed from habitat features achieved an AUC of 0.820 in the testing cohort; further integration of peritumoral features and clinical factors increased the AUC to 0.852, underscoring the value of multiparametric feature integration for improving the predictive accuracy. Additionally, the results of this study revealed that the features of the Habitat 1 subregion (the HH region) may be closely associated with the Ki-67 expression status. As a key biomarker of proliferative activity in breast cancer (2), Ki-67 is important for predicting patient outcomes and tumor recurrence in clinical practice (29). High Ki-67 expression typically indicates active tumor cell division and high invasiveness and is therefore closely associated with poor patient outcomes; however, these tumors are often more sensitive to chemotherapy (6). Furthermore, the Ki-67 index has important discriminatory value in breast cancer molecular subtyping, particularly in distinguishing luminal A from luminal B breast cancer. Treatment approaches differ between these two groups: luminal A tumors require only endocrine therapy following surgery, whereas luminal B tumors necessitate the addition of chemotherapy drugs to the treatment regimen (30). Therefore, a preoperative noninvasive prediction of the Ki-67 expression level is essential for individualized treatment decision-making. Radiomics has potential as a tool for predicting the level of Ki-67 expression. Liang et al. (31) developed a radiomics classifier based on MRI images for the preoperative prediction of the Ki-67 expression status; similarly, Liu et al. (32) constructed and validated a nomogram model using ultrasound radiomics features to noninvasively predict the level of Ki-67 expression. Tagliafico et al. (33) similarly employed DBT to investigate whether radiomics features were correlated with Ki-67 expression in breast cancer. These studies analyzed tumors as a whole but did not thoroughly assess subregions within the tumor, which may contain information regarding ITH. Consequently, conventional radiomics may not be the most sensitive method for assessing ITH (13). Advances in habitat imaging technology have enabled us to divide tumors into multiple subregions with distinct metabolic and structural characteristics, thereby revealing ITH with greater precision (34). Furthermore, imaging features extracted from these habitat subregions typically exhibit greater stability than those derived from the entire tumor region (35). Consequently, previous studies have explored the role of habitat radiomics in breast cancer. Wang et al. (36) explored the value of whole-tumor and subregional contrast-enhanced mammography (CEM) radiomics in predicting the HER2 expression status in breast cancer. Although many radiomics-based habitat analyses have focused on tumor spatial heterogeneity, they predominantly employed unsupervised clustering algorithms (such as K-means) to partition the tumor regions. For instance, Jiang et al. (37) used the K-means algorithm to segment DBT images based on the DBT intensity and local entropy. Wu et al. (38) used the simple linear iterative clustering (SLIC) algorithm for the superpixel segmentation of computed tomography (CT) scans and 18F-fluorodeoxyglucose (FDG) positron emission tomography (PET) imaging data. In contrast, this study innovatively used the local Moran’s index to analyze the grayscale autocorrelations of image pixels. We used this approach because DBT provides high-spatial-resolution volumetric data. When conventional habitat segmentation methods such as K-means or SLIC superpixel clustering are applied to such high-dimensional images, incur substantial computational cost and time are incurred. The use of the local Moran’s index method dramatically reduced the computational burden while preserving fine spatial details. This approach divides tumors into four distinct subtypes: HH, LL, HL, and LH (39). With its basis in spatial statistical inference, this method provides information on the structure of the tissue within a lesion more effectively than traditional approaches do. The HH region is defined by high-intensity voxels and surrounding high-intensity voxels, which may correspond to relatively dense tissue areas. This property could be due to a high cellular density or other factors leading to increased X-ray attenuation. The LL region consists of low-intensity voxels and their adjacent low-intensity voxels, which may represent low-density components such as adipose tissue or necrotic tissue. The HL region consists of isolated high-intensity voxels within a low-density background, which may indicate microcalcifications. LH regions consist of isolated low-intensity voxels within a high-density background, which may reflect cystic changes within the dense tumor parenchyma. The Moran’s index approach improves the interpretability of subregion classification and provides deeper insights into the relationships between imaging features and pathological foundations. However, these interpretations still require direct histopathological validation. Future studies combining image registration with pathological specimens are needed to confirm the biological relevance of these habitat subregions. In a previous study, Moran’s index was used to identify spatial microbiome modules from whole-slide pathological images of patients with breast cancer, demonstrating that these modules exhibit highly structured, spatially distinct patterns within tissues (40). This finding further elucidates the spatial diversity within the breast cancer ecosystem, confirming that subregional segmentation based on spatial autocorrelation effectively captures biologically meaningful structural heterogeneity within tumors. It holds significant research value for understanding the clinical implications of tumor heterogeneity. In summary, these findings underscore that the Moran’s index serves as an effective tool for analyzing breast medical imaging subregions, highlighting the need for further exploration of its application potential. During the subregion segmentation process, we did not set a minimum region size threshold for the habitat subregions, and thus some subregions with a small number of voxels being included in the analysis. While this approach may introduce noise and affect the reliability of the radiomic features extracted from these regions, we consistently divided each voxel within a tumor into four subregions, ensuring the consistency of subregion types across different patients. More importantly, the results of the SHAP analysis reveal that the top three features contributing most to the model’s predictions all originate from the larger habitat regions 1 and 2 (i.e., the HH and LL regions), indicating that the features we extracted exhibit good overall stability. Nevertheless, future studies may consider excluding subregions smaller than the preset size to further investigate whether features from these regions influence the results, thereby improving the robustness of the findings. In recent years, increasing attention has been given to the peritumoral region in radiomics studies of breast cancer (41-43), probably because it contains biological features that are closely related to the tumor. Breast cancer includes not only tumor cells but also the altered mesenchyme surrounding them; thus, the peritumoral region of breast cancer is considered to play a key role in tumorigenesis and progression and is thus a potential therapeutic target (44). In this study, a comparison of peritumoral regions of different widths (2, 4, and 6 mm) revealed that the features in the 4 mm region displayed the highest predictive efficacy for the Ki-67 status, suggesting that the peritumoral region in this range may best represent the biological behavioral characteristics at the tumor–stroma interface. This result may have been obtained because the 4 mm zone better captures microenvironmental information, thereby providing a more representative imaging correlation for the Ki-67 proliferation index. We further evaluated whether the breast composition affected model performance by stratifying patients according to ACR BI-RADS density categories. The Combined model achieved an AUC of 0.957 in nondense breasts (ACR a–b) and 0.778 in dense breasts (ACR c–d). The higher predictive accuracy observed in nondense breasts may be attributed to clearer tumor-background interfaces and more stable radiomic features in fatty-dominant tissue, where the contrast between the tumor and surrounding parenchyma is greater. Conversely, dense fibroglandular tissue can introduce greater image heterogeneity, partial volume effects, and overlapping structures, which may reduce the reproducibility and discriminative power of radiomic features. However, the sample sizes for these two subgroups were relatively small, and thus these results should be interpreted with caution. Importantly, the model’s performance in dense breast tissue remained at a clinically acceptable level (AUC =0.778), suggesting that radiomics techniques based on DBT are applicable to patients with dense breast tissue. Nevertheless, larger and more balanced cohort studies are needed to validate these findings. Furthermore, optimization strategies specifically tailored for dense breasts, such as customized image preprocessing, density-specific feature normalization, or the development of independent radiomic models for different density categories may further improve the predictive accuracy in this challenging patient subgroup. This study has several limitations. First, this study was a small-sample, single-center retrospective study with a relatively small testing cohort (n=44). This results in wide CIs for the AUC of some models and low statistical stability. Therefore, the generalizability and statistical robustness of the proposed models must be validated using larger, independent, and preferably multicenter external datasets in future studies. Second, our DBT images were acquired from a single vendor (Siemens), which may introduce model dependence. Future multicenter studies with data from different DBT vendors are necessary to validate the generalizability of our proposed framework. Third, subgroup analyses based on breast density were limited by the small sample sizes in each subgroup; therefore, these results should be considered exploratory. Future studies should focus on optimizing the model for women with dense breasts. Finally, suspicious calcifications were present in only 30 patients. Consequently, we were unable to perform meaningful subgroup analyses based on the calcification morphology or distribution, nor could we assess whether the model’s performance differed between patients with and without suspicious calcifications. Future studies with larger numbers of calcification-positive cases are needed to address this question. Conclusions In conclusion, multiple DBT-based radiomics models were constructed and evaluated in this study. The combined model integrating habitat, peritumoral, and clinical features achieved optimal performance in the preoperative noninvasive prediction of the Ki-67 expression status in patients with invasive breast cancer presenting as a mass on DBT images. These findings demonstrate the potential clinical translational value of the model for this specific population. Consequently, this approach holds potential as a precise, noninvasive assessment tool to effectively guide individualized clinical treatment decision making. Future studies involving larger cohorts are necessary to validate the findings of this study. Supplementary The article’s supplementary files as Acknowledgments None. Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Institutional Review Board of The Affiliated Hospital of Jiangsu University (No. KY2025K0801), and individual consent for this retrospective analysis was waived. Footnotes Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0047/rc Funding: This work was supported by the National Natural Science Foundation of China (grant No. 82272066) and the Zhenjiang Science and Technology Program (grant No. JC2025014). Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0047/coif). The authors have no conflicts of interest to declare. Data Sharing Statement Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0047/dss References Associated Data Supplementary Materials The article’s supplementary files as Data Availability Statement Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0047/dss