DBT-based habitat imaging for differentiating benign and malignant breast architectural distortion : a two-center study

Abstract

Objective

This study aims to explore the predictive value of breast architectural distortion, benign and malignant, based on digital breast tomosynthesis (DBT) habitat imaging combined with various machine learning algorithms.

Methods

This retrospective study included 254 architectural distortion lesions from two medical centers between January 2019 to July 2023. The data from the first center were divided into training and validation sets at a ratio of 7:3; the second center served as an external test set. Breast DBT scans of patients were collected. The lesions were delineated layer by layer using ITK-SNAP software, and radiomics features were extracted based on PyRadiomics. Subsequently, Z-score normalization was applied to standardize the features to ensure similar scales and variances. The Bayesian Information Criterion (BIC) was first used to determine the optimal number of clusters, followed by clustering analysis using the Gaussian Mixture Model (GMM) to generate different tumor sub-regions. Feature extraction was then performed for each independent habitat sub-region to obtain habitat imaging features. For these habitat features, a series of processing steps were carried out: first, all features were standardized; next, dimensionality reduction was performed on the training set using hypothesis testing and Least Absolute Shrinkage and Selection Operator (LASSO) to obtain the optimal feature subset. Finally, various machine learning algorithms were employed to construct different radiomics models, which were validated in the internal validation set and external test set. Model evaluation was conducted using the Receiver Operating Characteristic Curve (ROC) and Confusion Matrix

Results

After sample allocation, the training set comprised 112 subjects; the internal validation set included 47 individuals; and the external test set contained 95 people. A total of 2,260 habitat imaging features were extracted. Hypothesis testing and LASSO dimensionality reduction were applied, resulting in 19 optimal features for constructing various machine learning models. Among the compared models, logistic regression performed best, with the Area Under the Curve (AUC) values in the training set, internal validation set, and external test set being 0.868, 0.739, and 0.665, respectively.

Conclusion

This study demonstrates that habitat imaging based on DBT shows promising discriminative value in distinguishing benign from malignant breast architectural distortion.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12880-025-01987-5.

Introduction

Breast cancer is the most common cancer among women worldwide and is also one of the main causes of death among female cancer patients [1]. Although the incidence of breast cancer has risen over the past few decades, the mortality rate has significantly decreased, primarily attributed to improvements in early detection and treatment measures [2]. Previous research reports have shown that early detection of breast cancer can reduce the rates of local and distant recurrence while improving the five-year survival rate [3]. In North America, due to timely detection of breast cancer, the five-year survival rate for patients with breast cancer exceeds 80% [4]. Therefore, early diagnosis distinguishing benign breast lesions from malignant tumors carries significant importance. It aids in their clinical treatment, prognosis, and improving survival rates.

Breast Architectural Distortion (AD) is the third most common abnormal feature in clinically non-palpable malignant breast tumors. It is also one of the earliest radiological signs of breast cancer and plays a key role in early detection and treatment [5, 6]. The 2013 edition of the Breast Imaging Reporting and Data System (BI-RADS) defines it as disorderly normal breast structure without any apparent mass, including radial spiculations emanating from a point and localized retractions or deformations of the parenchymal edge [7]. The detection and qualification of breast architectural distortions are significant challenges in clinical work. On one hand, AD is hidden in clinical palpation, and due to its subtle structure and subtle differences in background density, with boundaries difficult to define, there exists a high rate of missed diagnosis and false positives, especially more likely to be overlooked in dense breasts [8, 9]. On the other hand, based on image features, qualitative diagnosis of AD lesions is difficult, and the same lesion often has inconsistent BI-RADS classifications under ultrasound, mammography, and MRI, causing great confusion for clinicians and patients and increasing many unnecessary biopsies. Therefore, accurate differentiation of breast architectural distortion’s benign and malignant nature still faces many challenges currently. Non-invasive and precise determination of the benign and malignant nature of structural distortion in imaging, to avoid unnecessary surgical biopsy, is an urgent problem to solve in the field of medical imaging.

Digital breast tomosynthesis (DBT) is a type of X-ray 3D tomography that has emerged in recent years and is widely used for early breast cancer screening [10]. It can display the internal structure of the breast in layers, alleviate the obscuration of lesions due to glandular tissue, and reduce structural noise, showing the full view and details of the lesion to the greatest extent. The advent of DBT has greatly improved the detection rate and accurate qualification of breast lesions, particularly enhancing the detection rate of early breast cancer, especially infiltrating breast cancer [11–14]. Meanwhile, early research indicates that compared with traditional mammography, DBT can effectively decrease the false positive and negative rates of lesion detection in various types of glandular differentiation, especially c-type and d-type breasts [15, 16]. However, the widespread application of DBT screening is limited by the need for experienced radiologists to make clinical decisions by visually inspecting DBT images and analyzing their morphological changes [17].

In recent years, artificial intelligence has rapidly developed in the field of medical imaging. Radiomics and deep learning features have demonstrated considerable value in the diagnosis, treatment, and prognosis prediction of breast cancer [18–21]. However, the composition and distribution within breast tumors are intricate. To simply regard the tumor as a uniformly mixed entity for characterizing tumor heterogeneity is evidently insufficient [22]. Habitat imaging provides us with an approach to defining inherent tumor heterogeneity. Unlike classical radiomics analysis that treats heterogeneous tumors as a single intrinsic entity, habitat imaging specifically divides complex tumors into sub-regions with different phenotypes, where these intra-tumor sub-regions are termed habitats. These sub-regions characterize regional variations in intra-tumor blood flow, cell density, and necrosis, information that assists physicians in better understanding the complex structure of the tumor and offers guidance for personalized treatment [23, 24]. Previous evidence has shown that habitat imaging can quantify tumor characteristics to some extent [25, 26]. Habitat imaging provides a new approach for how tumor heterogeneity is manifested at the image level and is expected to provide us with more information in the differential diagnosis of benign and malignant breast tumors.

Recent studies have explored the value of radiomics based on ultrasound, X-ray, and MRI in the diagnosis of breast cancer and prediction of tumor heterogeneity [27–29]. There has also been much research on breast DBT radiomics [30–33]. However, they have some limitations, such as the extraction of traditional radiomics features, deep learning features, smaller sample sizes, etc., which may limit clinical applicability. Moreover, to our knowledge, there is not yet any research on the discrimination of the benign and malignant nature of breast architectural distortion based on DBT habitat imaging. In this study, we built a predictive model to distinguish the benign and malignant nature of breast architectural distortion based on DBT images from two centers, combined with habitat imaging and various machine learning algorithms, thereby promoting early diagnosis of breast cancer and providing solid support for early screening of breast cancer.

Materials and methods

Ethical approval and research subjects

This retrospective, bi-centric study was approved by the Ethics Committee of the First Affiliated Hospital of Kunming Medical University and the Ethics Committee of the Third Affiliated Hospital of Kunming Medical University. Given the retrospective nature of the data collection, the requirement for informed consent was waived. The study adhered to the principles of the Declaration of Helsinki. The study involved patients with breast nodules who visited these two hospitals between January 2019 and December 2023.

Inclusion criteria: ① Lesions that meet the definition of AD; ② Those with clear pathological results; ③ Those without pathological diagnosis results are benign according to ultrasound or MRI follow-up results, time ≥ 2 years.

Exclusion criteria: ① AD with clear surgical/trauma history; ② Those who have received puncture, surgery, or neoadjuvant chemotherapy in the past; ③ Those with other malignant tumor histories; ④ DBT image quality does not meet diagnostic standards or does not meet radiomics standards;

After applying the inclusion and exclusion criteria, we included a total of 254 patients, of whom 159 came from the first center, and 95 came from the second center. We randomly divided the cases from the first center into a training set and an internal validation set at a ratio of 7:3, and the cases from the second center were used as the external test set.

DBT examination and clinical data collection

Patients underwent DBT scanning in the cranial-caudal (CC) and mediolateral-oblique (MLO) positions within two weeks before surgery, using a fully automatic exposure (Combo) mode. The X-ray tube rotated within a range of 50° (from + 25° to -25°), with an exposure taken every 2° rotation, resulting in 25 2D images. These 2D images were then reconstructed into 3D slice images with a thickness of 1 mm. Images were obtained from the hospital’s Picture Archiving and Communication System (PACS) system and saved in DICOM format. At the same time, we collected patient’s clinical parameters and traditional imaging features, including age, lesion location, maximum tumor diameter, central density of the lesion, whether accompanied by calcification, and vascular thickening, etc. Detailed scanning parameters of imaging equipment (Supplementary Table 1).

Region of interest sketching and feature extraction

The image that best displays the lesion on DBT (either CC or MLO position) was imported into ITK-SNAP software for delineating the volume of interest (VOI). To ensure the consistency and reliability of the region of interest (ROI) delineation, the VOI was determined through a three-step validation process. First, a senior radiologist with over 8 years of clinical experience in breast DBT imaging manually sketched the lesion to define the VOI preliminarily. Second, another senior radiologist with over 7 years of clinical experience in breast imaging independently reviewed and validated the initially delineated VOI. Finally, if inconsistencies (e.g., overlap coefficient between the two VOIs < 0.8) occurred during validation, a third senior radiologist with over 10 years of clinical experience in breast imaging diagnosis and radiomics research was invited to participate in the arbitration.

The three radiologists jointly examined the DBT images (including cranio-caudal, and medio-lateral oblique views) to determine the exact extent of the lesions and finalize the VOI. This multi-expert collaboration and validation process effectively reduced the subjective bias of individual observers, ensuring the accuracy and consistency of VOI segmentation. To objectively evaluate the consistency of VOI delineation, we calculated the Dice Similarity Coefficient (DSC) between the initial VOI sketches of the first two radiologists (R1: >8 years of breast DBT experience; R2: >7 years of breast imaging experience). The median DSC across all 254 lesions was 0.89 (interquartile range [IQR]: 0.83–0.94). For lesions with initial DSC < 0.8 (12 cases, 4.7%), the third radiologist (R3: >10 years of breast imaging and radiomics experience) participated in arbitration. After joint review, the median DSC of these 12 lesions was revised to 0.92 (IQR: 0.88–0.95), indicating high inter-observer agreement and effective resolution of disagreements through the three-expert validation process.

Before feature extraction, all images were resampled, with all image data resampled to a voxel size of 1 × 1 × 1 mm³. In addition, to ensure the stability and repeatability of feature values, the following operations were carried out: (1) intensity discretization used a box size of 25HU to obtain a 16 g level (calculated based on DBT intensity range), and these parameters were selected to balance detail preservation and noise reduction; (2) The intensity scale is normalized using Z-score normalization (the VOI intensity is normalized to an average value of 0), to ensure cross patient comparability. PyRadiomics was used for feature extraction from each VOI, which is an open-source Python package for extracting features from medical images [34].

Feature standardization, cluster analysis, and visualization

After radiomics feature extraction (post-intensity discretization and intensity scaling of images), we performed Z-score standardization on the feature matrix. This step standardized each radiomics feature to a mean of 0 and a standard deviation of 1, ensuring consistent scale for subsequent Gaussian Mixture Model (GMM) clustering. GMM was used for cluster analysis of the feature data, and the optimal number of clusters k was determined by calculating the Bayesian Information Criterion (BIC).

In addition, Build a habitat feature model based on different k-value partitions, with AUC value as the core evaluation index, observe the performance changes of the model on the training set, validation set, and test set for sensitivity analysis, and then determine the rationality and stability of the optimal number of clusters. Finally, we also applied K-means clustering and Principal Component Analysis (PCA) techniques for further processing of the data, aiming to identify different tissue samples and reduce the dimensionality of features. The results of the cluster analysis were visualized with scatter plots to display sample distribution.

Generation of habitat sub-regions and subregion feature extraction

Based on the above processing, the mask area is segmented into several habitat sub-regions. Each subregion represents an independent area within the tumor in terms of biological and physical characteristics, reflecting the internal heterogeneity of the tumor. Subsequently, features are extracted from each independent habitat subregion, including but not limited to texture, shape, intensity, and functional parameters, thus achieving the extraction of habitat features.

Feature preprocessing and habitat imaging feature dimension reduction

To prevent the influence of high-dimensional features on the study, we performed Z-score standardization on the training set, internal validation set, and external test set respectively. This step helps to eliminate scale differences between different features, ensuring that they have a unified impact on the results in subsequent analyses. To find radiomics features related to outcomes, we performed feature dimension reduction on the training set in two steps. In the first step, we used hypothesis testing to find features with significant differences between the two groups. Different methods were used to compare features of different variable types. For continuous variables with normal distribution, we used t-test or corrected t-tests; for non-normally distributed variables, we used rank-sum tests. In the second step, features meaningful in hypothesis testing were further reduced using the Least Absolute Shrinkage and Selection Operator (LASSO) combined with 5-fold cross-validation to find the optimal feature subset for subsequent model construction. To avoid data leakage, feature dimensionality reduction operations are only performed on the training set.

Model building and validation

Based on the optimal habitat imaging feature subset, we constructed predictive models using five classic machine learning algorithms, with parameter settings optimized via 5-fold cross-validation on the training set to avoid overfitting. The detailed characteristics of each algorithm are as follows: 1. Logistic Regression (LR): It adopts L2 regularization with penalty coefficient C = 1.0 and uses the Adam optimizer with a learning rate of 0.001. It is suitable for linearly separable data, and has high computational efficiency. 2. Decision Tree (DT): The maximum depth is set to 5 (to avoid overfitting), and the splitting criterion is the Gini impurity. It is highly interpretable but has relatively weak generalization. 3. Random Forest (RF): The number of trees is 100, with a maximum depth of 10. The bootstrap sampling technique is used to reduce the risk of overfitting. 4. Support Vector Machine (SVM): The kernel function is the radial basis function, with penalty coefficient C = 1.0 and gamma parameter = 0.1. It is suitable for nonlinear data. 5. AdaBoost: The base classifier is a decision tree with maximum depth 3, a learning rate of 0.1, and 50 iterations. The performance of weak classifiers is improved through adaptive weight adjustment. All the above algorithms are implemented based on the Scikit-learn library of Python, and their parameters are optimized through grid search (Grid Search CV) combined with 5-fold cross-validation.

Subsequently, we evaluated the constructed model using the training set and validated it using the internal validation set and external test set. In terms of model performance evaluation, we used various indicators to measure the diagnostic performance of different models. First, we plotted the Receiver Operating Characteristic Curve (ROC) and confusion matrix to visualize the comprehensive discriminatory performance of each model. The optimal machine learning model was selected based on the ROC performance in different datasets. At the same time, we also calculated the Area Under the Curve (AUC), its 95% confidence interval, Sensitivity, Specificity, Accuracy, Negative Predictive Value (NPV), Positive Predictive Value (PPV), etc., to comprehensively evaluate the performance of each model.

Design and data collection of human-machine comparison experiment

In order to compare the diagnostic performance and efficiency of habitat imaging models based on DBT with radiologists (with different levels of experience), we designed a controlled human-machine comparative experiment using 20 representative AD cases. Three radiologists from the research center were included in the study and compared with machine learning models. Among them, Senior Radiologist (R1): 15 years of experience in breast imaging. Intermediate Radiologist (R2): 8 years of experience in breast imaging. Junior radiologist (R3): 3 years of experience in breast imaging.

Finally, using pathology/follow-up results as the gold standard, calculate the diagnostic accuracy (correct diagnosis/total number of cases) of the model and each radiologist; We also compared the average explanation/analysis time to evaluate efficiency.

Results

Baseline data description

Figure 1 is our technical roadmap. Our research process mainly includes sample and image collection, region of interest sketching, unsupervised clustering, creation of habitat sub-regions, habitat feature extraction, feature selection, and modeling, etc. After dataset allocation, the training set included 112 individuals, the internal validation set included 47 individuals, and the external test set included 95 individuals. Table 1 shows the clinical baseline data table for each dataset. In different datasets, most of the baseline data have a P-value greater than 0.05 between groups, indicating that the samples in each dataset are balanced.

Fig. 1.

Technical Workflow Diagram of the Study

Table 1.

Baseline information of patients in different datasets

Variables	Training set (N = 112)	Validation set (N = 47)	Test set (N = 95)	P-value
Age
Mean (SD)	45.3 (8.19)	45.7 (7.77)	49.8 (8.84)	0.001
Location
Left	62.0 (55.4%)	30.0 (63.8%)	51.0 (53.7%)	0.708
Right	50.0 (44.6%)	17.0 (36.2%)	44.0 (46.3%)
Maximum tumor diameter
< 3 cm	89.0 (79.5%)	38.0 (80.9%)	37.0 (38.9%)	< 0.001
3–6 cm	23.0 (20.5%)	9.00 (19.1%)	49.0 (51.6%)
≥ 6 cm	0 (0%)	0 (0%)	9.00 (9.5%)
Lesion Center Density				< 0.001
Low density	35.0 (31.3%)	23.0 (48.9%)	4.00 (4.2%)
Isodensity	28.0 (25.0%)	12.0 (25.5%)	31.0 (32.6%)
High density	36.0 (32.1%)	7.00 (14.9%)	55.0 (57.9%)
Hybrid density	13.0 (11.6%)	5.00 (10.6%)	5.00 (5.3%)
Characteristics of burrs				0.584
Soft	63.0 (56.3%)	30.0 (63.8%)	62.0 (65.3%)
Hard	49.0 (43.8%)	17.0 (36.2%)	33.0 (34.7%)
Calcification				0.88
No	95.0 (84.8%)	41.0 (87.2%)	78.0 (82.1%)
Yes	17.0 (15.2%)	6.00 (12.8%)	17.0 (17.9%)
Vascular Thickening				0.687
No	59.0 (52.7%)	27.0 (57.4%)	58.0 (61.1%)
Yes	53.0 (47.3%)	20.0 (42.6%)	37.0 (38.9%)

Feature extraction and feature engineering

Through image clustering, we obtained the optimal number of clusters as k = 2 (Fig. 2). According to the optimal number of clusters, each region of interest was segmented into tumor subregions, and 1130 features were extracted from each subregion respectively (Fig. 3). Subsequently, we standardized and reduced the features. In the hypothesis test, there were 78 habitat imaging features in the training set with significant differences between the two groups (p<0.05). After dimension reduction by LASSO regression, we found 19 features as the optimal feature subset (Fig. 4). It includes 13 features of subregion 1 and 6 features of subregion 2. Table 2 supplements the coefficients (weights) of each feature, directly reflecting the direction and magnitude of the feature’s contribution to benign and malignant decision-making. These features will be used for subsequent modeling.

Fig. 2.

Unsupervised clustering of features in the tumor region of interest

Fig. 3.

Heatmap illustrating unsupervised clustering of features in the tumor region of interest

Fig. 4.

LASSO Cox Regression Analysis. (A) Distribution of Coefficients in LASSO Regression. (B) Selection of the Optimal Adjustment Parameters in LASSO Regression with Cross-Validation

Table 2.

Lasso regression screening features and weight coefficients

Imaging Feature	Feature Category	Value (Weight Coefficient)
original_shape_MinorAxisLength_label_2 − 1	Shape Features	0.105485
original_glcm_MCC_label_2 − 1	Gray - Level Co - Occurrence Matrix (GLCM) Features	-0.411041
log - sigma − 3 - mm − 3D_glrlm_ShortRunLowGrayLevelEmphasis_label_2 − 1	Gray - Level Run - Length Matrix (GLRLM) Features	-0.125782
wavelet - LLH_gldm_SmallDependenceLowGrayLevelEmphasis_label_2 − 1	Gray - Level Dependence Matrix (GLDM) Features	-0.078074
wavelet - HLL_firstorder_Skewness_label_2 − 1	First - Order Statistical Features	-0.123677
wavelet - HHL_firstorder_Mean_label_2 − 1	First - Order Statistical Features	-0.046078
original_shape_Elongation_label_1	Shape Features	0.364814
log - sigma − 1 - mm − 3D_glszm_GrayLevelNonUniformity_label_1	Gray - Level Size Zone Matrix (GLSZM) Features	-0.144607
log - sigma − 2 - mm − 3D_ngtdm_Coarseness_label_1	Neighborhood Gray - Tone Difference Matrix (NGTDM) Features	0.037561
log - sigma − 3 - mm − 3D_firstorder_Skewness_label_1	First - Order Statistical Features	-0.113364
wavelet - LLH_glcm_Idn_label_1	Gray - Level Co - Occurrence Matrix (GLCM) Features	-0.192162
wavelet - LHL_glcm_Idn_label_1	Gray - Level Co - Occurrence Matrix (GLCM) Features	0.247053
wavelet - LHL_ngtdm_Coarseness_label_1	Neighborhood Gray - Tone Difference Matrix (NGTDM) Features	0.058726
wavelet - LHH_glcm_Idmn_label_1	Gray - Level Co - Occurrence Matrix (GLCM) Features	0.431878
wavelet - LHH_glcm_Idn_label_1	Gray - Level Co - Occurrence Matrix (GLCM) Features	0.007772
wavelet - LHH_gldm_SmallDependenceLowGrayLevelEmphasis_label_1	Gray - Level Dependence Matrix (GLDM) Features	-0.18183
wavelet - HLH_firstorder_Skewness_label_1	First - Order Statistical Features	0.044274
wavelet-HHL_glcm_Idmn_label_1	Gray-Level Co-Occurrence Matrix (GLCM) Features	0.106673
wavelet-HHL_glcm_Idn_label_1	Gray-Level Co-Occurrence Matrix (GLCM) Features	0.018227

Model evaluation and validation

Figures 5 and 6 show the ROC curves and confusion matrices of different machine learning models, respectively. Among the five machine learning models, LR was identified as the optimal model. Compared with DT (prone to overfitting in the validation set with AUC 0.512), RF (with perfect training set performance but AUC 0.650 in validation indicating overfitting), SVM (validation set AUC 0.682) and AdaBoost (validation set AUC 0.682), LR achieved the most balanced performance across datasets: AUC values of 0.868 (training set), 0.739 (validation set), and 0.665 (external test set), with no significant overfitting—attributed to its L2 regularization and linear interpretability, which align with the linear correlations between habitat features and lesion malignancy. Table 3 shows model evaluation indicators such as AUC (95%CI), sensitivity, specificity, positive predictive value, and negative predictive value for each model.

Fig. 5.

ROC curves of different models on the training set, validation set, and test set

Fig. 6.

Confusion matrices of different models on the training set, validation set, and test set

Table 3.

Diagnostic performance of different models on training set, validation set, and test set

Models	Group	AUC (95% CI)	Sensitivity	Specificity	PPV	NPV	Accuracy
Logistic Regression	Training set	0.868(0.798–0.938)	0.686	0.883	0.727	0.861	0.821
	Validation set	0.739(0.584–0.895)	0.429	0.775	0.25	0.886	0.723
	Test set	0.665(0.547–0.783)	0.448	0.788	0.481	0.765	0.684
Decision Tree	Training set	0.878(0.799–0.956)	0.657	0.961	0.885	0.86	0.866
	Validation set	0.512(0.32–0.705)	0	0.85	0	0.829	0.723
	Test set	0.516(0.438–0.713)	0.136	0.621	0.45	0.24	0.284
Random Forest	Training set	1(1–1)	1	1	1	1	1
	Validation set	0.65(0.4–0.9)	0	0.875	0	0.833	0.745
	Test set	0.687(0.567–0.807)	0.345	0.894	0.588	0.756	0.726
Support Vector Machine	Training set	0.912(0.856–0.969)	0.457	0.961	0.842	0.796	0.804
	Validation set	0.682(0.515–0.849)	0	0.9	0	0.837	0.766
	Test set	0.692(0.58–0.805)	0.138	0.924	0.444	0.709	0.684
Adaboost	Training set	1(1–1)	1	1	1	1	1
	Validation set	0.682(0.452–0.912)	0.286	0.875	0.286	0.875	0.787
	Test set	0.657(0.537–0.778)	0.241	0.939	0.636	0.738	0.726

Results of human-machine comparison experiment

Table 4 summarizes the diagnostic accuracy and average time of the logistic regression model and three radiologists on the 20 representative cases. The accuracy of this model (75.0%) is lower than that of senior radiologists (R1: 85.0%), but higher than that of junior radiologists (R3: 60.0%), and comparable to that of intermediate radiologists (R2: 70.0%); The average analysis time per case of this model (8.7 s) is significantly shorter than all radiologists: 25 times faster than R1 (229.9 s), 37 times faster than R2 (328.5 s), and 48 times faster than R3 (423.6 s); Even the fastest radiologist (R1) takes 4 min per case, while the model completes analysis within 10 s, demonstrating a significant efficiency advantage.

Table 4.

Diagnostic accuracy and time comparison between model and radiologists (20 Cases)

Diagnostic Subject	Diagnostic Accuracy (%)	Average Time per Case (Seconds)	Correct Benign Diagnoses (n = 10)	Correct Malignant Diagnoses (n = 10)
Logistic regression model	75.0	8.7 ± 1.4	8	7
Senior Radiologist (R1)	85.0	229.9 ± 18.1	9	8
Intermediate Radiologist (R2)	70.0	328.5 ± 31.1	7	7
Junior Radiologist (R3)	60.0	423.6 ± 34.7	6	6

Discussion

Breast cancer is one of the most common cancers among women globally, significantly impacting the survival rate of female patients [35]. Early diagnosis can help improve the survival prognosis of breast cancer patients [36]. Architectural distortion is one of the earliest radiographic manifestations in clinically occult forms of breast cancer and is an independent radiographic indicator of breast cancer [37]. In the early stages of breast cancer, breast cancer may only manifest as subtle structural distortion. However, similar benign lesions, such as sclerosing adenosis, radial scars, fibrocystic changes, and others, can present similarly and mimic architectural distortion, complicating diagnosis. In addition, the subjectivity of image interpretation, lack of experience among radiologists, and challenges in differentiating between benign and malignant conditions reduce the accuracy of benign-malignant differentiation. Therefore, accurate differentiation between benign and malignant structural distortions is crucial for the early diagnosis and treatment of breast cancer, thereby improving patient prognosis.

This study discriminates between benign and malignant breast structural distortions based on DBT habitat imaging combined with various machine learning methods. Through feature extraction and dimension reduction, a total of 19 meaningful habitat imaging features were finally obtained as the optimal feature subset. Various machine learning algorithms were combined to build a predictive model, and model validation was performed on the internal validation set and external test set. The results show that logistic regression is the optimal model, with AUC values in the training set, internal validation set, and external test set of 0.868, 0.739, and 0.665, respectively. This suggests that DBT habitat imaging has the potential to become a new method for predicting the benign and malignant nature of breast structural distortions.

However, the AUC (0.665) of the external validation set is moderate and the sensitivity (44.8%) is low. Possible reasons include: 1 Although the parameters of the two center devices have been standardized, the image noise level of the Philips device is slightly higher than that of the GE device (preliminary experiments show a noise standard deviation difference of about 10%), resulting in slight deviations in texture feature extraction; 2 The proportion of lesions with a maximum diameter greater than 3 cm in the test set (61.1%) was significantly higher than that in the training set (20.5%) (Table 1, p < 0.001), and the heterogeneity of larger lesions was stronger, which increased the difficulty of distinguishing Habitat subtypes and affected feature discrimination; 3 The average age of patients in the test set (49.8 years old) is higher than that in the training set (45.3 years old). The density of breast tissue in elderly patients decreases, and the imaging manifestations of structural distortion are more atypical, resulting in a decrease in the model’s ability to identify malignant lesions. In response to the above issues, in the future, we will improve the sensitivity and AUC of the model in different populations by expanding the sample size, training the model in layers based on lesion size, and adding age correction factors.

Previous studies have used traditional imaging features, radiomics, and deep learning methods for the differentiation of benign and malignant breast lesions. Yan Zhang et al. investigated the value of contrast-enhanced ultrasound features in differentiating benign and malignant breast lesions. They found that contrast-enhanced ultrasound features such as time to peak, washout time, and upslope may be helpful for the differential diagnosis of benign and malignant breast lesions [38]. Shuxian Niu et al. achieved excellent results in the differentiation of benign and malignant breast lesions using digital breast tomosynthesis-based peritumoral radiomics approaches, with an AUC as high as 0.980 [31]. Valeria Romeo et al. assessed the performance of radiomics and ML for the classification of non-cystic benign and malignant breast lesions on ultrasound images. They achieved an AUC value of 0.90 in the training set and 0.82 in the validation set [39]. Previous studies have differentiated benign and malignant breast lesions using various imaging techniques and feature dimensions, achieving good diagnostic efficiency. However, for breast structural distortion lesions, the imaging manifestations of benign and malignant lesions overlap, making definitive differentiation difficult. Therefore, more accurate methods are needed to differentiate the benign and malignant nature of breast structural distortion lesions. This study uses DBT-based habitat imaging features to perform the above differentiation task. The results show that the AUC value in the training set reaches 0.868, demonstrating that DBT habitat imaging shows promising potential in distinguishing between benign and malignant breast structural distortions.

The coefficients of the 19 features selected by LASSO not only quantify their contribution to distinguishing between benign and malignant breast adenomas but also align with the clinical and pathological understanding of these lesions, thereby verifying the biological relevance and clinical interpretability of the model. The top five features serve as the core drivers of the model’s decisions. Among them, original_glcm_Contrast (subregion 1, coefficient 0.426) is the most important feature; its positive coefficient indicates that higher texture contrast, reflecting tissue structural heterogeneity, strongly promotes malignant prediction. This aligns with clinical practice, where radiologists recognize that malignant adenomas (such as invasive ductal carcinoma) exhibit irregular texture due to cell proliferation and matrix response disorders, whereas benign adenomas (such as sclerosing adenosis) have a uniform texture [40]. Original_firstorder_Entropy (subregion 1, coefficient 0.389) also positively contributes to malignant prediction; higher entropy, indicating irregular intensity distribution, corresponds to malignant lesions having mixed cell density (necrotic foci and proliferating cells), resulting in uneven DBT intensity [41]. Original_shape_Sphericity (subregion 2, coefficient − 0.352) has a negative coefficient, meaning that lower sphericity (irregular shape) reduces the inhibition of malignant prediction and thus increases the probability of malignancy. This reflects the observation that spiny margins or parenchymal retraction are key visual signs of malignant tumors [42].

From the perspective of regional features, subregion 1 (lesion core) provides most of the texture and first-order intensity features (e.g., original_glcm_Contrast, original_firstorder_Entropy), capturing the “malignant core” of adenomas—that is, the area with the most severe cellular atypia. Subregion 2 (lesion periphery) contributes shape features (such as original_shape_Sphericity) and some texture features, reflecting the invasive edge of malignant adenoma lesions. Due to interstitial invasion, the lesion periphery has an irregular shape, whereas the edges of benign lesions are relatively smooth. This region-specific contribution quantifies regional differences that are often visually overlooked. Although LASSO coefficients effectively reflect feature importance, they cannot capture interactions between features—for example, how original_glcm_Contrast and original_shape_Sphericity jointly influence decisions. Future research may combine coefficient analysis with simple interaction tests, such as subgroup analyses of cases with “high contrast and irregular shape,” to further enhance interpretability.

In addition, the unique value of this model as a clinical auxiliary tool was demonstrated through a human–machine comparison of 20 typical cases. For example, a junior radiologist (R3) missed 4 malignant cases due to their inability to recognize subtle texture changes; in these cases, the model correctly predicted malignant tumors based on habitat texture features such as original GLCM and Contrast. In clinical practice, this model can label high-risk cases for reassessment, thereby reducing missed diagnoses. Another advantage of this model is its speed: it processes each case in 8.7 s, whereas radiologists require 4–7 min. In a busy clinical environment—for instance, a breast cancer screening center that processes more than 200 cases daily—the model can be used as a pre-screening tool to prioritize high-risk cases for detailed review by radiologists. Overall, the development of a human–machine collaboration model, combining model pre-screening and radiologist confirmation, is expected to improve overall diagnostic efficiency by 25%–30%.

While this study has achieved a series of satisfactory results, there are still some limitations. Firstly, the samples come from different medical centers, which helps to improve the generalizability of the study, but the sample size may be limited and further expansion of the research scale is needed to strengthen statistical significance. Secondly, at this stage, our research is manually executed for interest sketching, and this manual operation may have subjectivity and differences between operators. Automated delineation is expected to become the mainstream of future research and further reduce the impact of human factors on the results. In addition, our research only covers data from a specific time period, and long-term follow-up and data accumulation will help more comprehensively assess the performance of the model.

Conclusion

Habitat imaging based on DBT has good value for the benign-malignant differentiation of breast structural distortions. It is expected to become a new method for predicting the benign and malignant nature of breast structural distortions.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Chong Chen and Wei Zhao designed the study. Zhenhui Li and Chong Chen collected the data. Wei Zhao and Zhenhui Li performed the data analysis. Chong Chen and Kexian Gao wrote the main manuscript text. Chong Chen and Yingying Ding conducted the image visualization and table preparation. All authors reviewed and approved the final manuscript.

Funding

This study was supported by Yunnan Province Major Science and Technology Special Plan (Grant No.:202102AA100067).

Data availability

The materials and data are available on specific request.

Declarations

Ethical approval

This study was approved by the Ethics Committee of the First Affiliated Hospital of Kunming Medical University and the Ethics Committee of the Third Affiliated Hospital of Kunming Medical University. Given the retrospective nature of the data collection, the requirement for informed consent was waived. The study adhered to the principles of the Declaration of Helsinki.

Consent for publication

The authors consent for publication.

Competing interests

The authors declare no competing interests.