Diagnosing breast masses in BI-RADS 4: construction and interpretation of an intralesional and perilesional radiomics nomogram based on digital breast tomosynthesis

Abstract

Background

To evaluate the diagnostic performance of a digital breast tomosynthesis (DBT)-based, interpretable radiomics nomogram for differentiating between benign and malignant Breast Imaging Reporting and Data System (BI-RADS) category 4 masses.

Methods

This retrospective study enrolled 312 patients with BI-RADS 4 lesions assessed via DBT. Radiomics features were extracted from intralesional and perilesional regions of interest (ROIs), and used to construct three models: Radscore_IL, Radscore_PL, and Radscore_Combine. Univariate and multivariate logistic regression analyses identified independent predictors, which were integrated into clinical and nomogram models. Model performance and interpretability were assessed using receiver operating characteristic (ROC) curves, calibration curves, decision curve analysis (DCA), and Shapley Additive Explanations (SHAP) values.

Results

Among the three radiomics models, Radscore_Combine demonstrated superior performance, with area under the curve (AUC) values of 0.935 (training) and 0.888 (validation). Age and BI-RADS 4 subclassification emerged as independent clinical predictors and were included in the clinical model. The final nomogram, incorporating Radscore_Combine, age, and BI-RADS 4 subclassification, achieved AUCs of 0.960 and 0.942 in the training and validation datases, respectively. DCA confirmed significant clinical utility of the nomogram. SHAP analysis enhanced interpretability by illustrating individual feature contributions to model predictions.

Conclusions

The DBT-based radiomics nomogram offers high diagnostic accuracy and good interpretability, making it a valuable tool for differentiating benign from malignant BI-RADS 4 masses.

Relevance statement

This DBT-based radiomics nomogram accurately differentiates benign and malignant BI-RADS 4 breast lesions, supporting clinical decision-making to optimize management of these ambiguous findings.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12880-026-02459-0.

Key points

Developed an interpretable nomogram using DBT-based Radscore and clinical features.

Combined intralesional and perilesional features improved diagnostic accuracy.

Nomogram accurately distinguished benign from malignant BI-RADS 4 breast masses.

SHAP analysis enhanced model transparency and feature interpretability.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12880-026-02459-0.

Background

Breast cancer has one of the highest incidence rates among all malignant tumors and is a leading cause of cancer-related mortality among women worldwide [1–5]. Early detection and prompt treatment are essential for improving patient survival rates [6].

Mammography is widely recognized as a main screening method for early detection of breast cancer [2, 7–9]. The breast imaging reporting and data system (BI-RADS) has established a standardized framework for diagnosing breast lesions, proving to be highly effective in the characterization of breast lesions [2]. BI-RADS 5 lesions demonstrate a malignancy probability exceeding 95%, representing reliable indicator for breast cancer diagnosis with minimal diagnostic uncertainty [2, 10]. In contrast, BI-RADS 4 encompasses a spectrum of indeterminate breast lesions that display ambiguous imaging features, neither fulfilling definitive benign criteria nor demonstrating conclusive malignant characteristics, with a variable malignancy risk ranging from 2% to 95% [2, 11, 12]. While BI-RADS 4 lesions can be further classified into 4 A, 4B, and 4 C subcategories, it is still recommended to proceed with an invasive biopsy for confirmation of their pathological characteristics [2]. Accurately identifying the BI-RADS 4 lesions can reduce unnecessary biopsies [2]. Digital breast tomosynthesis (DBT) is an advanced mammography technique that captures multiple angled images of a breast lesion to reconstruct detailed tomographic images. This innovative method improves lesion detection, particularly in dense breast tissue [1, 7]. DBT also enhances diagnostic accuracy in identifying and characterizing BI-RADS 4 lesions with indeterminate findings on digital mammography [13]. However, DBT interpretation for the image features has certain limitations. It requires experienced radiologists to perform a detailed analysis of lesion imaging features, and the BI-RADS evaluation process could be influenced by subjective factors [9, 14]. These limitations may affect the efficacy of DBT in distinguishing between benign and malignant lesions.

Radiomics, a burgeoning machine learning technique, offers a non-invasive approach supporting preoperative clinical decision-making. By extracting features from medical images that may elude human perception, radiomics has the potential to provide valuable insights for clinicians [1, 2, 15–20]. Several earlier studies have employed intralesional radiomics analysis in the diagnosis of breast cancer and have demonstrated encouraging outcomes [16–18]. The perilesional region merits particular attention, as it harbors blood/lymphatic plexuses, immune cell infiltrates, and stromal reactions—factors potentially constituting determinants of tumorigenic risk [21]. Incorporating extralesional radiomics features significantly enhances diagnostic accuracy in breast cancer assessment [2]. Recently, there has been growing recognition and utilization of radiomics analysis focusing on the perilesional region in the field of breast cancer diagnosis [2, 9, 21]. However, there has been a limited amount of research dedicated to perilesional radiomics analysis in distinguishing between benign and malignant BI-RADS 4 lesions [2], especially with regard to DBT.

The limited interpretability and unclear internal mechanisms of some models can pose challenges to their widespread clinical application [20]. Shapley Additive Explanations (SHAP), a game theory-based method, addresses this issue by quantifying the individual contribution of each feature to a model’s prediction for a given case [22, 23]. SHAP has been increasingly utilized in radiomics research to enhance model transparency and interpretability [20, 22–25].

Multi-region radiomics studies using DBT specifically for the risk stratification of BI-RADS category 4 lesions remain relatively scarce. This study integrated both intralesional and perilesional features to capture the subtle microenvironmental information essential for distinguishing malignancy. To enhance the clinical utility of our findings, we employed the SHAP algorithm to provide a transparent interpretation of the final model. By quantifying the contributions of each radiomic and clinical factor, we constructed a visualized nomogram. This approach offers clinicians a highly interpretable and non-invasive diagnostic aid, potentially facilitating more informed decision-making and helping to reduce unnecessary biopsies for BI-RADS 4 lesions.

Methods

Patients selection

This retrospective study was approved by the hospital’s ethics committee, with informed consent waived due to its retrospective design. Patients with breast lesions were included in the study, covering the period from December 2019 to June 2023.

All patients included in the study met the specified criteria for inclusion, which are detailed below: (1) Patients underwent DBT examination and were assessed with a BI-RADS category 4; (2) Pathology confirmed benign or malignant breast lesions via surgery or biopsy within one month post-DBT examination; (3) Prior to conducting DBT examination, no breast biopsies, treatments, or prior diagnoses of breast cancer had been recorded. Patients who met any of the exclusion criteria were deemed ineligible for participation in the study: (1) Patients lacking sufficient clinicopathological data - Missing any of the following: Family history of breast cancer, Body mass index (BMI), or Histopathological results; (2) Imaging that is partial or of substandard quality - Absence of standard DBT views; Poor image resolution or blurring affecting lesion assessment or segmentation; Incomplete visualization of the target lesion; (3) Breast lesions present as non-mass findings, including pure calcifications, asymmetry, and architectural distortion. Finally, 312 female patients with BI-RADS category 4 lesions were eligible in this study. All cases were randomly divided into training and validation datasets at a ratio of 7:3. In cases involving multifocal or bilateral lesions, the analysis focused on the largest BI-RADS 4 lesion [2]. Figure 1a provides a visual depiction of the patient selection process.

Fig. 1.

Study flowchart. (a) Patient selection process. (b) Workflow for constructing and validating the radiomics nomogram

DBT protocol

The Selenia Dimensions imaging system (Hologic, Marlborough, MA) was used for DBT examinations during this study. Details of the imaging acquisition parameters are provided in Appendix E1 of the Supplementary Materials.

Clinical information

Patient clinical information, including age, BMI, family history of breast cancer, and pathology results of breast lesions, was collected and recorded through a review of medical records. Pathology findings were used as the gold standard for categorizing lesions into benign and malignant groups. The DBT evaluation—based on prior studies [13, 26–28]—is presented in Appendix E2 of the Supplementary Materials.

Segmentation and concordance evaluation

The intralesional region of interest (IL-ROI) was carefully delineated through manual tracing along the tumor boundary at its maximum diameter, as observed in both the craniocaudal (CC) and mediolateral oblique (MLO) views of the DBT images. The segmentation of lesions was performed by two radiologists, referred to as Radiologist 1 and Radiologist 2, both with over five years of experience in breast lesion diagnosis. Using 3D Slicer (version 4.11; available at http://www.slicer.org), Radiologist 1 outlined all IL-ROIs. Once the IL-ROI was established, a perilesional region of interest (PL-ROI) was generated by expanding the IL-ROI outward using Python (version 3.7) with the SimpleITK library (version 2.1.1). Specifically, morphological dilation was performed on the IL-ROI mask to achieve a uniform 4‑mm outward expansion. The 4 mm extension for the perilesional region was determined based on prior DBT radiomics studies confirming this as the optimal distance for information capture, balancing anatomical relevance with maximal predictive performance [8, 21, 29, 30]. Any sections of the PL-ROI that extended beyond the breast parenchyma were manually removed. Preceding the radiomics feature extraction process, image resampling and grayscale discretization were carried out to achieve normalization, in alignment with previous research practices [9, 15, 31]. A comprehensive description of the feature extraction process and intraclass correlation coefficient (ICC) evaluation is provided in Appendix E3 of the Supplementary Materials. Radiomics features with intra- and inter- ICC values exceeding 0.75 were deemed to have good repeatability and were included in the subsequent feature selection process.

Development of Radscore models

Prior to the selection of radiomics features, all features were standardized using Z-score normalization. In the training dataset, a three-step procedure was conducted for the radiomics features [15]: (a) Statistical Significance: An independent samples t-test or Mann-Whitney U test was conducted to identify statistically significant features; (b) Redundancy Removal: Spearman rank correlation analysis was performed to identify and eliminate redundant features (correlation coefficient exceeding 0.90); and (c) Feature Optimization: The Least Absolute Shrinkage and Selection Operator (LASSO) regression, in combination with a 5-fold cross-validation approach, was used to identify the most predictive features. The optimal regularization parameter (λ) was determined based on the criterion of minimum binomial deviance. Features with non-zero coefficients at the optimal λ were selected for final model construction. The radiomics score (Radscore) for each patient was then calculated by linearly combining the selected features, with each feature weighted according to its corresponding coefficient [2, 32]. Using this methodology, three Radscore models were developed: Radscore_IL: Derived from intralesional radiomics features; Radscore_PL: Derived from perilesional radiomics features; Radscore_Combine: Derived from both intralesional and perilesional radiomics features. To ensure full reproducibility, a fixed random seed was applied to all stochastic operations, including dataset partitioning and cross-validation.

Nomogram construction and model validation

In the univariate analysis, clinical factors such as age, lesion size, BI-RADS category, and other related variables were investigated. Variables that exhibited statistical significance at a p-value less than 0.05 were subsequently incorporated into a multivariable logistic regression to formulate a clinical model. In order to improve the model’s fit and reduce redundancy, a backward stepwise selection method utilizing the Akaike information criterion (AIC) was implemented [15]. Moreover, a radiomics nomogram model was established by integrating independent clinical predictors and the optimal Radscore, which exhibited the highest area under the curve (AUC) value [33, 34]. To assess collinearity, we utilized the variance inflation factor (VIF). We used the Hosmer-Lemeshow test to gauge the radiomics nomogram’s goodness-of-fit, and the calibration curve was employed to assess the agreement between the predicted probability and the actual probability. The decision curve analysis (DCA) was applied to assess the clinical utility of different models. To assess the improvement brought about by the incorporation of Radscore, the net reclassification improvement (NRI) and integrated discrimination improvement (IDI) analyses were carried out. The NRI quantifies the net proportion of cases correctly reclassified into risk categories, while the IDI evaluates the overall improvement in the model’s discriminative power by comparing the mean change in predicted probabilities. Figure 1b provides a visual representation of the construction and validation of the radiomics nomogram model.

Model interpretability exploration

The SHAP technique enhances model transparency by offering both global and local interpretability [24]. Rooted in cooperative game theory, SHAP values represent the average marginal contribution of each feature across all possible feature combinations [22]. In this study, SHAP summary plots were used to identify and rank the most influential features within the model. SHAP dependence plots were employed to illustrate how specific features affect model predictions and to visualize potential interaction effects between feature pairs. Additionally, SHAP waterfall plots were generated to explain individual predictions, showcasing how each feature contributed to the final model output for a specific case [35].

Statistical analysis

The statistical analyses for this study were carried out using SPSS software (version 26.0, Chicago, IL) and R software (version 4.1.3, accessible at http://www.R-project.org). A detailed overview of the statistical methods and the primary R software packages used in this study can be found in Appendix E4 and Table S1 of the Supplementary Materials.

Results

Comparison of clinical information and construction of clinical model

This study included a total of 312 BI-RADS 4 patients, comprising 180 with malignant lesions and 132 with benign lesions. Among the malignant cases, 150 were invasive ductal carcinomas, 5 were ductal carcinomas in situ, and 25 were other malignant types. The benign lesions included 90 cases of fibroadenomas, 7 phyllodes tumors, 3 papillomas, and 32 cases of other benign types.

The training dataset included 97 benign and 121 malignant lesions, while the validation dataset consisted of 35 benign and 59 malignant lesions. As shown in Table 1, no significant differences were observed in clinically relevant variables between the training and validation datasets (p = 0.234–0.759). The comparison of clinical variables between benign and malignant groups in the training and validation datasets is detailed in Table S2. Univariate and multivariate logistic regression analyses were performed on all clinical variables. Patient age and BI-RADS 4 subclassification were identified as independent predictive factors for BI-RADS 4 malignant lesions (all p < 0.001), as demonstrated in Tables 2 and 3. These two factors were subsequently used to construct clinical models. The VIF for both factors was approximately 1.000, indicating no significant collinearity.

Table 1.

Comparisons of patient characteristics in the training and validation datasets

Characteristics	Training dataset (n = 218)	Validation dataset (n = 94)	P-value
Groups, n (%)			0.234
Benign	97 (44.50)	35 (37.23)
Malignant	121 (55.50)	59 (62.77)
Age, years	47.71 ± 12.88	48.60 ± 12.31	0.571
Family history of breast cancer, n (%)			0.709
No	204 (93.58)	89 (94.68)
Yes	14 (6.42)	5 (5.32)
BMI, n (%)			0.475
≤ 24 kg/m2	118 (54.13)	55 (58.51)
> 24 kg/m2	100 (45.87)	39 (41.49)
Breast density, n (%)			0.131
Nondense	98 (44.95)	51 (54.26)
Dense	120 (55.05)	43 (45.74)
Tumor location, n (%)			0.679
Upper outer quadrant	117 (53.67)	54 (57.45)
Lower outer quadrant	17 (7.80)	5 (5.32)
Upper inner quadrant	31 (14.22)	9 (9.57)
Lower inner quadrant	26 (11.93)	14 (14.89)
Others	27 (12.39)	12 (12.77)
Lesion size, mm	23.31 (18.06, 32.09)	22.20 (16.58, 29.67)	0.257
BI-RADS category, n (%)			0.552
4a	58 (26.61)	22 (23.40)
4b or 4c	160 (73.39)	72 (76.60)
Radscore_Combine	0.25 (-0.93, 1.75)	0.40 (-0.67, 1.54)	0.746
Radscore_IL	0.30 ± 1.16	0.38 ± 1.13	0.595
Radscore_PL	0.25 ± 0.98	0.29 ± 0.84	0.759

Nondense ACR a + b, Dense ACR c + d

Table 2.

Univariate logistic analysis for the risk factors

Characteristics	OR	95%CI	P-value
Age	1.137	1.095–1.180	< 0.001
Family history of breast cancer
No	Reference
Yes	0.789	0.267–2.334	0.669
BMI
≤ 24 kg/m2	Reference
> 24 kg/m2	1.512	0.880–2.597	0.134
Breast density
Nondense	Reference
Dense	0.956	0.558–1.635	0.868
Tumor location
Upper outer quadrant	Reference
Lower outer quadrant	0.522	0.186–1.468	0.218
Upper inner quadrant	1.182	0.526–2.655	0.686
Lower inner quadrant	0.640	0.272–1.503	0.305
Others	1.085	0.464–2.541	0.850
Lesion size	0.999	0.980–1.019	0.936
BI-RADS category
4a	Reference
4b or 4c	10.873	5.083–23.262	< 0.001
Radscore_Combine	5.897	3.721–9.347	< 0.001

Nondense ACR a + b, Dense ACR c + d, CI confidence interval, OR Odds ratio

Table 3.

Multivariate logistic analysis for the risk factors

Characteristics	Clinical model			Radiomics nomogram model
	OR	95%CI	P-value	OR	95%CI	P-value
Age	1.121	1.078–1.166	< 0.001	1.104	1.047–1.164	< 0.001
BI-RADS category
4a	Reference			Reference
4b or 4c	5.646	2.379–13.401	< 0.001	5.660	1.507–21.250	0.010
Radscore_Combine	NA	NA	NA	4.849	3-7.837	< 0.001

CI confidence interval, OR Odds ratio

Feature selection and radscore model establishment

Using the "three-step process" for feature selection, 18, 9, and 17 optimal radiomics feature subsets were ultimately identified through LASSO regression. These subsets were used to construct three Radscore models: Radscore_Combine, Radscore_IL, and Radscore_PL, to differentiate between benign and malignant BI-RADS 4 lesions (Figures S1-3). A detailed list of the radiomics features and their respective coefficients used in constructing these three Radscore models is provided in Table S3.

There were no statistically significant differences in the Radscore values between the training and validation datasets for all three Radscore models (p = 0.595–0.759; details in Table 1). In both the training and validation datasets, Radscore values for malignant lesions were significantly higher than those for benign lesions across all Radscore models (all p < 0.001, as shown in Table S2 and Figures S4-6).

As illustrated in Table 4; Fig. 2a-b, the Radscore_Combine model demonstrated superior performance compared with both the Radscore_IL and Radscore_PL models, achieving AUCs of 0.935 (95% CI: 0.900–0.970) in the training dataset and 0.888 (95% CI: 0.822–0.954) in the validation dataset.

Table 4.

Performance of established three radscore models in both the training and validation datasets

Prediction models	Datasets	Cutoff value	AUC(95% CI)	ACC	SEN	SPE	PPV	NPV	P-value
RS1	Training	-0.061	0.935(0.900–0.970)	0.904	0.909	0.897	0.917	0.888
RS1	Validation	-0.061	0.888(0.822–0.954)	0.798	0.797	0.800	0.870	0.700
RS2	Training	0.072	0.906(0.865–0.948)	0.862	0.860	0.866	0.889	0.832
RS2	Validation	0.072	0.847(0.771–0.924)	0.755	0.780	0.714	0.821	0.658
RS3	Training	0.456	0.834(0.781–0.887)	0.743	0.653	0.856	0.849	0.664
RS3	Validation	0.456	0.737(0.630–0.844)	0.702	0.627	0.829	0.861	0.569
DeLong test for the comparison of AUC values between different models
RS1 vs. RS2	Training								< 0.001
RS1 vs. RS2	Validation								0.047
RS1 vs. RS3	Training								< 0.001
RS1 vs. RS3	Validation								0.003
RS2 vs. RS3	Training								0.011
RS2 vs. RS3	Validation								0.072

ACC accuracy, CI confidence interval, NPV negative predictive value, PPV positive predictive value, RS1 Radscore_Combine, RS2 Radscore_IL, RS3 Radscore_PL, SEN sensitivity, SPE specificity

Fig. 2.

Radscore_combine distribution. Bar charts showing Radscore_Combine in the training (a) and validation (b) datasets. The y-axis represents the adjusted Radscore_Combine. Specifically, the cutoff value of − 0.061 was subtracted from the raw Radscore_Combine to center the distribution, thereby yielding a clearer visualization

Development and validation of radiomics nomogram

Patient age, BI-RADS 4 subclassification, and Radscore_Combine were identified as independent predictive factors for malignant BI-RADS 4 lesions (Table 3). The VIF values for these factors ranged from 1.007 to 1.030, indicating no significant collinearity issues. Thus, the radiomics nomogram model, constructed using patient age, BI-RADS 4 subclassification, and Radscore_Combine, is stable and reliable (Fig. 3a).

Fig. 3.

Radiomics nomogram construction and performance evaluation. (a) Visualization of the constructed radiomics nomogram. (b) Receiver operating characteristic curve (ROC) curves for the Radscore_Combine model, clinical model, and radiomics nomogram in the training dataset. (c) ROC curves for the three models in the validation dataset.

In both the training and validation datasets, the nomogram model outperformed the Radscore_Combine and clinical models in diagnostic efficacy, achieving AUCs of 0.960 (95% CI: 0.934–0.985) and 0.942 (95% CI: 0.899–0.985), respectively (details in Table 5). Figure 3b and c illustrate the respective ROC curves for the three models in both datasets, respectively. Figures S7a and S7b present the calibration curves for the nomogram model, indicating excellent agreement between the predicted and actual probabilities. Specifically, the calibration slope and intercept were 1.000 and 0.000 for the training dataset, and 0.850 and 0.503 for the validation dataset, respectively. The Hosmer-Lemeshow test yielded P-values of 0.272 and 0.160 for the training and validation datasets, respectively, both exceeding 0.05, demonstrating a good model fit.

Table 5.

Diagnostic performance of the Radscore_Combine, clinical model, and Radiomics nomogram in both the training and validation datasets

Prediction models	Datasets	Cutoff value	AUC(95% CI)	ACC	SEN	SPE	PPV	NPV	P-value
M1	Training	-0.061	0.935(0.900–0.970)	0.904	0.909	0.897	0.917	0.888
M1	Validation	-0.061	0.888(0.822–0.954)	0.798	0.797	0.800	0.870	0.700
M2	Training	0.600	0.851(0.801–0.901)	0.766	0.760	0.773	0.807	0.721
M2	Validation	0.600	0.858(0.780–0.937)	0.787	0.780	0.800	0.868	0.683
M3	Training	0.498	0.960(0.934–0.985)	0.917	0.926	0.907	0.926	0.907
M3	Validation	0.498	0.942(0.899–0.985)	0.883	0.881	0.886	0.929	0.816
DeLong test for the comparison of AUC values between different models
M1 vs. M2	Training								0.005
M1 vs. M2	Validation								0.573
M1 vs. M3	Training								0.002
M1 vs. M3	Validation								0.033
M2 vs. M3	Training								< 0.001
M2 vs. M3	Validation								0.009

ACC accuracy, CI confidence interval, NPV negative predictive value, PPV positive predictive value, M1 Radscore_Combine, M2 Clinical model, M3 Radiomics nomogram, SEN sensitivity, SPE specificity

Across most threshold probability ranges, the nomogram model demonstrated higher net benefits compared with the Radscore_Combine and clinical models (Figures S8a and S8b). Additionally, NRI and IDI analyses revealed that incorporating Radscore_Combine into the nomogram model significantly enhanced its discriminatory power compared with the model based solely on clinical factors (details in Table 6).

Table 6.

Assessment of radiomics nomogram and clinical models through the use of NRI and IDI

Characteristics	Training dataset			Validation dataset
	Categorical NRI [95% CI]	Continuous NRI [95% CI]	IDI [95% CI]	Categorical NRI [95% CI]	Continuous NRI [95% CI]	IDI [95% CI]
Radiomics nomogram vs. clinical models	0.353 [0.236–0.469]	1.505 [1.333–1.676]	0.322 [0.261–0.384]	0.205 [0.019–0.392]	1.149 [0.814–1.483]	0.201 [0.114–0.287]
P-value	< 0.001	< 0.001	< 0.001	0.031	< 0.001	< 0.001

CI confidence interval, NRI net reclassification improvement, IDI integrated discrimination improvement

Exploration of model interpretability

SHAP values were computed for all variables included in the nomogram model. Figure 4a presents a summary plot, where positive SHAP values indicate an increased predicted risk of malignancy, while negative values suggest a decreased risk. Variables positioned higher in the plot exhibit greater predictive power and are considered more influential on the model’s output than those lower down.

Fig. 4.

Model interpretability using SHAP. (a) SHAP summary plot showing the contribution of each feature to model predictions. (b-d) SHAP dependence plots for Radscore_Combine, age, and BI-RADS subclassification, illustrating feature impact and interactions. (e) SHAP waterfall plot for a representative benign case. (f) SHAP waterfall plot for a representative malignant case

SHAP dependence plots for Radscore_Combine, age, and BI-RADS 4 subclassification (Figs. 4b–d) illustrate the relationships between individual feature values and their corresponding SHAP values, as well as their interactions with other features. Figure 4e and f present SHAP waterfall plots for two representative test cases. Figure 4e depicts a case confirmed pathologically as benign. Here, the negative contributions (purple bars) lead to a predicted value ƒ(x) = -10.2, which is below the base value E[ƒ(x)] = 0, indicating a benign outcome. In contrast, Fig. 4f shows a case confirmed as malignant, where positive contributions (yellow bars) result in a predicted value of ƒ(x) = 1.23, surpassing the base value and suggesting malignancy.

Discussion

This study successfully developed three Radscore models—Radscore_IL, Radscore_PL, and Radscore_Combine—by extracting and analyzing radiomics features from the intralesional and perilesional regions based on DBT images. Among these models, the Radscore_Combine model demonstrated superior diagnostic performance, achieving significantly higher AUC values compared with the other models in distinguishing benign from malignant BI-RADS 4 lesions. To further enhance diagnostic efficacy, a radiomics nomogram model was developed by integrating patient age, BI-RADS 4 subclassification, and Radscore_Combine. The nomogram model exhibited exceptional efficacy in distinguishing between benign and malignant BI-RADS 4 lesions.

In clinical practice, biopsy is commonly regarded as the preferred method for determining whether BI-RADS 4 lesions are benign or malignant. However, this approach is invasive and may cause physical discomfort and potential risks to patients [36]. Therefore, actively exploring non-invasive auxiliary diagnostic methods is of significant importance. Through a detailed analysis of clinical risk factors associated with breast cancer, this study identified patient age and the BI-RADS 4 subclassification based on DBT as independent predictors of malignancy in BI-RADS 4 lesions. Specifically, older age or lesions classified as BI-RADS 4b/4c were significantly associated with a higher risk of malignancy. These findings align with the results of previous studies [2, 37, 38]. Consequently, a clinical model was constructed by integrating patient age and BI-RADS 4 subclassification. In both the training and validation datasets, this model achieved AUC values of 0.851 and 0.858, respectively, indicating strong diagnostic performance. These results are comparable to those of the clinical model developed by Zhang et al. [2], further reinforcing the critical role of age and BI-RADS 4 subclassification in distinguishing benign from malignant breast lesions.

Previous studies have successfully developed radiomics models based on the intralesional region for differentiating benign from malignant breast lesions, achieving impressive model performance [39–41]. However, these studies frequently overlook the potential of radiomics features in the perilesional region. Recent research has highlighted that perilesional radiomics analysis plays a significant role in distinguishing between benign and malignant breast lesions. In recent years, with the deepening of research, accumulating evidence has confirmed that radiomics research based on the surrounding area of lesions plays an indispensable role in the differentiation of benign and malignant breast lesions, and its value has been widely acknowledged and validated [2, 9, 21].

Currently, the field of multi-regional radiomics remains divided on the identification of benign and malignant breast lesions, with ongoing debate about the optimal extent of the perilesional region for analysis. Different research teams have reported varying conclusions based on diverse imaging modalities and methodologies. Some studies, however, have raised concerns that increasing the size of the perilesional region may reduce diagnostic accuracy, as larger regions may incorporate more normal breast tissue, diluting the predictive value of radiomics models [42]. Conversely, too narrow a perilesional region may fail to capture critical tumor infiltration information, potentially diminishing diagnostic efficacy [21]. Recent DBT-based radiomics research has provided evidence that a 4 mm perilesional region contains valuable information relevant for predicting Ki-67 expression levels [8], molecular subtype [29], and lymphovascular invasion [30] in breast cancer. This 4 mm expansion may serve as an empirical threshold that maintains an optimal balance between capturing sufficient perilesional stromal reactions and minimizing background noise from surrounding normal glandular tissue. Therefore, this study defined the perilesional region as extending 4 mm beyond the lesion boundary and conducted radiomics analysis on both this region and the intralesional region.

In this study, radiomics features from the intralesional, perilesional, and combined regions of breast lesions were analyzed using the LASSO algorithm, which enabled the successful development of three radiomics models: Radscore_IL, Radscore_PL, and Radscore_Combine. The Radscore_Combine model integrates features from multiple regions, incorporating first-order features (such as the 10th percentile of grayscale values) from the intralesional region, two-dimensional (2D) shape features (such as sphericity), and wavelet filtering features from both the intralesional and perilesional regions. The 10th percentile, a first-order feature, reflects the grayscale value at the 10th percentile of the lesion’s intensity distribution. Some studies suggest that this metric plays a key role in distinguishing splenic lymphoma from benign splenomegaly [43]. Sphericity, a dimensionless measure, quantifies how closely the shape of the lesion approximates a perfect sphere [44]. In 2D shape features, the sphericity value ranges from 0 to 1, with values approaching 1 indicating a shape that is more regular and closer to a circle. This study found a negative correlation between the sphericity of the intralesional region and the risk of malignancy in BI-RADS 4 lesions, suggesting that malignant lesions are more likely to have irregular shapes compared with benign ones. Additionally, wavelet filtering features, which capture multi-scale characteristics of lesions, are known for their ability to quantify tumor heterogeneity accurately [7, 45]. This study found that wavelet features from both the intralesional and perilesional regions played a crucial role in differentiating between benign and malignant BI-RADS 4 lesions. These findings are consistent with previous radiomics research by Zhang et al. [2], further highlighting the significant potential of wavelet filtering features in distinguishing benign from malignant BI-RADS 4 lesions.

This study thoroughly analyzed radiomics features from both intralesional and perilesional regions, resulting in the development of the Radscore_Combine model. This model demonstrated significantly higher diagnostic efficacy compared with the Radscore_IL and Radscore_PL models, which are limited to features derived from a single region (either the lesion’s internal or surrounding area). The rationale behind this finding may be that the perilesional region contains critical information that aids in distinguishing between benign and malignant BI-RADS 4 lesions. Radiomics analysis enables the efficient integration and quantification of features from multiple regions, providing a more comprehensive and in-depth view of the lesion’s biological characteristics. In some challenging BI-RADS 4 cases, where the lesion’s nature could not be reliably determined using radiomics model based solely on intralesional features, the Radscore_Combine model may offer significant support.

Previous studies have shown that combining radiomics features with clinical factors to construct a nomogram model can improve diagnostic accuracy and provide personalized risk assessments, thereby guiding clinical decision-making [1]. Building on this approach, our study successfully developed a radiomics nomogram model by integrating patient age, BI-RADS 4 subclassification, and Radscore_Combine. This model demonstrated superior discriminatory ability in distinguishing between benign and malignant BI-RADS 4 lesions, with significantly better diagnostic performance compared with the clinical model and the Radscore_Combine model alone. The analysis results of NRI and IDI highlight the critical role of Radscore_Combine in enhancing model performance. Additionally, the SHAP algorithm was employed to visualize the individual contribution of each feature for every case. This not only clarifies the influence of each variable on the model’s predictions but also improves interpretability from a clinical perspective. The inclusion of SHAP greatly supports the practical application of the model by enhancing its transparency and clinical utility.

This study has several limitations. First, this was a single-center retrospective study with a limited sample size and a lack of external validation. There was a class imbalance between benign and malignant lesions in our study cohort, which might lead to a slight overestimation of the model performance. Furthermore, the independent cumulative risk of multiple lesions was not evaluated in this study. Future studies should adopt strategies for imbalanced data, conduct large-scale multicenter validation, develop multi-lesion combined prediction models, and perform adequate sensitivity analyses. Second, lesion ROI segmentation in this study was performed manually using a 2D method. While this approach improves segmentation efficiency, it may not fully capture information about the entire lesion. Third, this study only extracted and analyzed radiomics features from a 4 mm area surrounding the lesion, without exploring the potential value of features from other regions. Future research will compare the predictive performance of different expansion distances (e.g., 2 mm, 4 mm, 6 mm) for differentiating benign and malignant BI-RADS 4 lesions, to determine the optimal parameter. Fourth, BI-RADS assessment may be influenced by the subjective judgement of radiologists, and this study exclusively focused on mass lesions without exploring other types such as pure calcifications, asymmetry, and architectural distortion. Fifth, the inclusion of all BI-RADS 4 subclasses with wide malignancy probabilities may limit model generalizability in specific subgroups, necessitating larger subcategory-focused validation; the absence of a direct comparison between our nomogram and the current BI-RADS 4 subclassification may to some extent limit the clinical translatability of the findings. Finally, although DM is an important tool for breast cancer screening, this study did not analyze DM images of breast lesions. Systematic DM-versus-DBT comparisons warrant future dedicated validation studies.

In summary, the Radscore_Combine, constructed by combing radiomics features extracted from the intralesional and perilesional regions on DBT images, effectively differentiates between benign and malignant BI-RADS 4 lesions. Compared with both the Radscore_Combine and clinical models, the radiomics nomogram model—integrating Radscore_Combine, patient age, and BI-RADS 4 subclassification—demonstrated superior efficacy. However, while these results are promising, the nomogram should currently serve as an auxiliary decision-support tool; broader validation across diverse cohorts and integration of automated segmentation for standardization are needed to solidify its clinical utility as a reliable auxiliary tool for biopsy deferral.

Electronic Supplementary Material

Below is the link to the electronic supplementary material.

Abbreviations

DBT

Digital breast tomosynthesis

BI-RADS

Breast imaging reporting and data system

ROI

Region of interest

Radscore

Radiomics score

AUC

Area under the curve

SHAP

Shapley Additive Explanations

Author contributions

Maolin Xu: conceptualization (equal), investigation (equal), methodology (equal), visualization (equal), writing original draft (equal). Xiaofang Guo: funding acquisition (equal), software (equal), visualization (equal). Zhenzhen Hu: data curation (equal), formal analysis (equal), software (equal). Guifeng Liu: funding acquisition (equal), conceptualization (equal), investigation (equal), methodology (equal), project administration (equal), supervision (equal), writing – review and editing (equal).Yulin Liu: methodology (equal), resources (equal), supervision (equal), writing– review and editing (equal).

Funding

This study was financially supported by Jilin Province Science and Technology Development Plan (Project No. 20220203113SF); Beijing Medical Award Foundation (Project No. YXJL-2024-0350-0147); National Key Clinical Specialty Construction Discipline (Project No: HBCHBCC-C05); Hubei Province “Chutian Talents Program” medical and health personnel project; Scientific Research Projects of Hubei Cancer Hospital (Project No. 2025HBCHYN42).

Data availability

The datasets generated and/or analyzed during the current study are available from the corresponding author upon reasonable request.

Declarations

Ethical approval and consent to participate

This study followed all dictates of the Declaration of Helsinki and was approved by the Ethics Committee of China-Japan Union Hospital of Jilin University. Written informed consent was waived due to the retrospective nature of this study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.