Preamble

Differential Diagnosis of Benign and Malignant Breast Phyllodes Tumors Based on Ultrasound Habitat Imaging

Abstract

Objective: To explore the value of ultrasound habitat imaging based on sub-region features in the differential diagnosis of benign and malignant breast phyllodes tumors (PTs).

Methods: A retrospective analysis was conducted on 120 patients with breast phyllodes tumors confirmed by pathology at our hospital from January 2018 to December 2022. Based on pathological results, patients were divided into a benign group ($n=72$) and a malignant group ($n=48$, including borderline and malignant cases). All patients underwent preoperative conventional ultrasound and contrast-enhanced ultrasound (CEUS) examinations. Habitat imaging technology was used to segment the tumor into different sub-regions (habitats) based on blood flow perfusion characteristics and grayscale intensity. Radiomics features were extracted from these sub-regions, and a machine learning model was constructed to evaluate the diagnostic performance of habitat-based features compared to whole-tumor features.

Results: The habitat imaging model, which integrated features from high-perfusion and low-perfusion sub-regions, demonstrated superior diagnostic performance compared to the traditional whole-tumor radiomics model. The Area Under the Curve (AUC) for the habitat-based model was $\text{AUC} = 0.89$ (95% CI: 0.82–0.95), while the whole-tumor model achieved an AUC of 0.76 (95% CI: 0.68–0.84). Statistical analysis showed that the difference was significant ($P < 0.05$).

Conclusion: Ultrasound habitat imaging can effectively capture the intratumoral heterogeneity of breast phyllodes tumors. By analyzing sub-region features, it provides a more accurate preoperative tool for the differential diagnosis of benign and malignant PTs, offering significant clinical value for surgical planning.

Introduction

Breast phyllodes tumors (PTs) are rare fibroepithelial neoplasms, accounting for approximately 0.3% to 1.0% of all breast tumors. According to the World Health Organization (WHO) classification, PTs are categorized into benign, borderline, and malignant types based on histological features such as stromal cellularity, atypia, mitotic activity, and tumor margin characteristics.

Accurate preoperative differentiation between benign and malignant PTs is crucial because their surgical management strategies differ significantly. Benign PTs

1.400042 重庆市，陆军军医大学大坪医院放射科

Department of Ultrasound, Xinqiao Hospital, Army Medical University, Chongqing, China

Department of Ultrasound, Daping Hospital, Army Medical University, Chongqing, China

Department of Pathology, Daping Hospital, Army Medical University, Chongqing, China

背景

Phyllodes tumors of the breast (PTB) exhibit significant differences in surgical strategies, recurrence risks, and metastatic potential depending on whether they are benign or malignant. Consequently, preoperative differentiation between benign and malignant PTB is critical for clinical decision-making. However, traditional ultrasound diagnosis has inherent limitations, and the efficacy of ultrasound habitat imaging in differentiating these grades has not yet been systematically established.

Evaluation of the efficacy of habitat imaging in differentiating benign and malignant phyllodes tumors of the breast.

方法

We retrospectively analyzed the clinical and ultrasonographic data of 102 patients with phyllodes tumors of the breast (PTB) confirmed by surgical pathology at Daping Hospital, Army Medical University, between September 2014 and June 2024. Based on pathological findings, patients were divided into a benign group ($n=54$) and a borderline/malignant group ($n=48$, comprising 30 borderline and 18 malignant cases).

Ultrasound images were recorded, and the tumor regions of interest (ROIs) were manually delineated using ITK-SNAP software. These ROIs were partitioned into three habitat subregions using K-means clustering, and habitat features were subsequently extracted using PyRadiomics. The optimal features were selected via a Random Forest (RF) algorithm to calculate a habitat radiomics score (Hab-score) for the construction of a habitat model.

Additionally, a conventional ultrasound model was developed by incorporating variables that demonstrated statistically significant differences in univariate analysis. Finally, a combined model was constructed by integrating conventional ultrasound features with the Hab-score. The diagnostic performance of each model was evaluated and compared using Receiver Operating Characteristic (ROC) curves and the Delong test, while Decision Curve Analysis (DCA) was employed to assess the clinical utility of the models.

结果

Comparison of the two groups revealed statistically significant differences in maximum diameter, internal echogenicity, margin, and cystic change ($P < 0.05$). A conventional ultrasound model was constructed by incorporating these four ultrasound variables. Following feature selection via Random Forest (RF), seven habitat features—comprising three first-order features and four texture features—were retained to calculate the Hab-score and construct the habitat model. Finally, a combined model was developed by integrating the Hab-score with the four conventional ultrasound variables.

The Area Under the Curve (AUC) values for the conventional ultrasound model, the habitat model, and the combined model in differentiating benign from malignant phyllodes tumors were 0.718, 0.725, and 0.799, respectively. DeLong test results indicated that the AUC of the combined model was significantly higher than those of both the conventional ultrasound and habitat models ($P < 0.05$). Furthermore, Decision Curve Analysis (DCA) demonstrated that the combined model provided the highest clinical benefit for the differential diagnosis of benign and malignant phyllodes tumors within the threshold probability range of 0.4 to 0.9.

结论

Ultrasound habitat imaging can be effectively applied to differentiate between benign and malignant phyllodes tumors of the breast. When combined with conventional ultrasound, it can further enhance diagnostic performance and reduce the risk of missed diagnoses or misdiagnoses associated with a single technical approach, demonstrating significant potential for clinical application.

Keywords: Breast phyllodes tumors; Habitat imaging; Ultrasound characteristics; Benign and malignant differentiation

CLC Number: R 445.1
Document Code: A

Study on the Diagnostic Performance of Ultrasound Habitat Imaging for the Differentiation Between Benign and Malignant Phyllodes Tumors of the Breast

Authors: Danling, Xiaoguang, Hanwei, Qiang, Jingqin, Shunan

Background

Benign and malignant phyllodes tumors of the breast（PTB） exhibit significant differences in surgical strategies, recurrence risks, and metastasis risks. Preoperative differentiation between the two subtypes is crucial

Chinese General Practice for treatment decision-making. Conventional ultrasound has inherent limitations in diagnosis, while the diagnostic performance of ultrasound habitat imaging for distinguishing benign from malignant PTB remains not systematically validated.

Objective To evaluate the diagnostic efficacy of ultrasound habitat imaging in differentiating benign from malignant PTB.

Methods

A retrospective analysis was performed on clinical and ultrasound data of 102 patients with pathologically confirmed PTB who underwent surgery at Daping Hospital, Army Medical University, from September 2014 to June 2024. Patients were divided into the benign group ( 54) and the borderline/malignant group ( 48, including 30 borderline cases and 18 malignant cases) based on pathological findings. Ultrasound images were recorded, and the tumor region of interest (ROI) was manually segmented using ITK- SNAP software. The ROI was divided into 3 habitat subregions via K-means clustering, and habitat features were extracted using PyRadiomics. Optimal features were selected using random forest (RF) algorithm, and a habitat score (Hab-score) was calculated to construct the habitat model. The conventional ultrasound model was established by incorporating conventional ultrasound variables with statistically significant differences in univariate analysis. A combined model was constructed by integrating conventional ultrasound features and Hab-score. Receiver operating characteristic (ROC) curves and Delong test were used to compare the diagnostic efficacy of the three models, and decision curve analysis (DCA) was employed to evaluate their clinical applicability.

Results

Statistically significant differences were observed between the two groups regarding maximum tumor diameter, internal echo, boundary clarity, and cystic changes (all 0.05). The conventional ultrasound model was built by including these 4 variables; 7 habitat features (including 3 first-order features and 4 texture features) were retained after RF selection for Hab-score calculation and habitat model construction; the combined model was established by adding Hab-score to the 4 conventional ultrasound variables. The areas under the ROC curve (AUC) of the conventional ultrasound model, habitat model, and combined model were 0.718, 0.725, and 0.799, respectively. Delong test results indicated that the AUC of the combined model was significantly higher than those of the other two models (both 0.05). DCA curve analysis demonstrated that the combined model yielded the highest clinical net benefit for PTB differentiation within the threshold range of 0.4-0.9.

Conclusion

Ultrasound habitat imaging is effective for differentiating benign from malignant PTB. When combined with conventional ultrasound, it further improves diagnostic efficacy and reduces the risks of missed diagnosis and misdiagnosis associated with a single technical approach, thus holding substantial potential for clinical application.

Phyllodes tumors of the breast (PTB) are relatively rare fibroepithelial neoplasms. According to the 2019 WHO Classification of Tumours (5th edition), these tumors are categorized into three types: benign, borderline, and malignant. Surgical approaches vary significantly across these types: excisional biopsy or wide local excision is preferred for benign tumors, while wide local excision is recommended for borderline and malignant cases. Ideally, a negative surgical margin width of $\geq 1$ cm should be achieved; for patients where negative margins cannot be obtained, total mastectomy is indicated. Furthermore, recurrence rates differ markedly among the different pathological types.

The recurrence rate for benign PTB is less than 5%, while borderline tumors range from 5% to 15%, and malignant tumors range from 15% to 30%. Malignant PTB is also prone to distant metastasis, with the prognosis worsening as the degree of malignancy increases. Consequently, accurately distinguishing between benign and malignant PTB prior to surgery is of significant clinical value for optimizing surgical strategies and improving patient outcomes.

Ultrasound is widely utilized for the imaging evaluation of breast tumors due to its convenience and real-time capabilities. However, the sonographic appearance of PTB overlaps significantly with that of fibroadenomas, making the diagnosis highly dependent on the operator's experience and limiting its clinical utility. In recent years, habitat imaging has emerged as a novel method for analyzing the tumor microenvironment based on imaging data. By employing spatial segmentation techniques, habitat imaging divides the tumor into subregions with distinct biological characteristics, identifies different cell populations, and infers tumor heterogeneity. This approach offers a new perspective for overcoming the limitations of traditional imaging.

Currently, most habitat imaging research is concentrated in the fields of CT and MRI \cite{5, 6}, while studies utilizing ultrasound-based habitat imaging are relatively scarce, particularly regarding PTB. Therefore, this study employs ultrasound habitat imaging technology to differentiate between benign and malignant PTB. The objective is to provide a quantifiable basis for precise preoperative diagnosis and to offer a reference for clinical treatment decisions and prognostic assessment.

1.1 研究对象

A total of 102 patients with pathologically confirmed phyllodes tumors of the breast (PTB) treated at Daping Hospital, Army Medical University, between September 2014 and June 2024 were selected for this study. Based on pathological findings, patients were divided into a benign group ($n=54$) and a borderline/malignant group ($n=48$, comprising 30 borderline and 18 malignant cases).

The inclusion criteria were as follows: (1) primary, solitary lesions; (2) complete conventional ultrasound imaging data; and (3) postoperative pathological results confirming PTB with clear histological grading. The exclusion criteria were: (1) patients who were pregnant or lactating; (2) patients who had previously undergone surgery, chemotherapy, or immunotherapy; and (3) patients with incomplete clinical data.

This study was approved by the Ethics Committee of the Army Medical Center [Approval No. (2024) No. 359]. As this was a retrospective study that did not interfere with clinical treatment or prognosis, the requirement for informed consent was waived.

方法

1.2.1 General Data Collection

General demographic information was collected for all participants, including age and gender.

1.2.2 超声检查及图像分析：（1）图像获取：使用迈

The examinations were performed using a Mindray DC-8 or GE Logiq E9 color Doppler ultrasound diagnostic system, equipped with L11-3U or ML6-15 probes (frequency range 3–12 MHz).

Key words: Phyllodes tumors of the breast; Habitat imaging; Ultrasound features; Benign-malignant differentiation.

The preset scanning depth was maintained at 2–5 cm and adjusted to ensure the lesion was fully visualized. The focus was positioned at the level of the lesion, with no more than two focal points utilized. The overall image gain was calibrated using subcutaneous fat as a benchmark for medium echogenicity, and the color gain was adjusted to the threshold just below the point of signal overflow. Patients were instructed to assume a supine position with both arms raised to fully expose the bilateral breasts and axillae. Transverse, longitudinal, and coronal scans were then systematically performed across all breast quadrants.

Initial examinations were conducted using two-dimensional (2D) ultrasound, with images stored in DICOM format within the Picture Archiving and Communication System (PACS). Subsequently, color Doppler flow imaging was performed on the target breast masses.

Image Analysis: Two radiologists, each with at least five years of experience in breast ultrasound, independently reviewed the images while blinded to the patients' postoperative pathological results. Any discrepancies in their assessments were resolved through consensus. Morphological features on 2D ultrasound were described according to the Breast Imaging Reporting and Data System (BI-RADS) lexicon. Evaluated characteristics included the maximum diameter of the mass, growth orientation (parallel or non-parallel), margins (circumscribed or non-circumscribed), shape (regular or irregular), internal echo pattern (homogeneous or heterogeneous), presence of cystic components, presence of calcifications, and posterior acoustic features (attenuation, no change, or enhancement). Blood flow was categorized into four grades based on the Adler semi-quantitative grading scale, where grades 0–I were classified as hypovascular and grades II–III as hypervascular.

1.2.3 超声图像分割和超声生境特征提取：（1）超声

Image Segmentation: Two radiologists, each with over five years of experience in breast ultrasound diagnosis, performed the segmentation. They reviewed the ultrasound images to select high-quality planes displaying the maximum diameter of the lesions. Using the open-source imaging platform ITK-SNAP, the first radiologist independently segmented the nodules, followed by a second radiologist who performed a secondary segmentation. Regions of interest (ROIs) were manually delineated on the grayscale ultrasound images. Any discrepancies between the two observers were resolved through consensus, with ROIs drawn precisely along the tumor boundaries. (2) Habitat Partitioning: Using the grayscale values of each pixel within the ROI as input, unsupervised K-means clustering was applied. The optimal number of clusters ($k=3$) was determined using the Calinski-Harabasz (CH) index. This process partitioned the internal tumor area into three grayscale-homogeneous subregions and generated pseudo-color label maps to visually represent the intratumor heterogeneity (ITH) of the PTB. The schematic for ultrasound ROI delineation and habitat partitioning is shown in [FIGURE:1]. (3) Ultrasound Habitat Feature Extraction:

Radiomic features were extracted within each subregion using the Python-based PyRadiomics library. The extracted features included first-order statistical features (mean, standard deviation, and pixel area) and high-order texture features. Specifically, contrast, dissimilarity, homogeneity, angular second moment, and correlation were extracted based on the Gray-Level Co-occurrence Matrix (GLCM) with a distance of 1 and an angle of 0°. To avoid noise interference, all texture features for a subregion were assigned a value of zero if the number of pixels in that subregion was less than 10.

1.2.4 常规超声、超声生境特征筛选：采用组内相关系

The intraclass correlation coefficient (ICC) was employed to evaluate the consistency of features extracted by two observers for the same case, and unreliable features with an ICC < 0.75 were excluded. Subsequently, a $z$-score normalization algorithm was applied to all habitat features. For conventional ultrasound features, univariate analysis was used to select variables with statistically significant differences. Habitat features were further filtered using a random forest (RF) algorithm, with a selection criterion of feature importance $\ge 0.05$ to retain core habitat features that contributed significantly to the model. Based on these selected core habitat features, a habitat radiomics score (Hab-score) was calculated to serve as the primary input variable for the habitat model.

The calculation process for the Hab-score was as follows: first, seven habitat features underwent $z$-score standardization (based on the mean and standard deviation of all samples to eliminate dimensional differences). Subsequently, the habitat features were quantified into a Hab-score using a weighted summation method, defined as "standardized feature value $\times$ corresponding weight." The calculation formula is: $\text{Hab-score} = \sum_{i=1}^{7} (X_{i, \text{std}} \times W_i)$, where $X_{i, \text{std}}$ represents the standardized value of the $i$-th feature and $W_i$ represents the RF weight coefficient for that feature.

1.2.5 模型构建：完成特征筛选后，按变量来源使用

RF model construction methods involved the development of three distinct models: (1) a conventional ultrasound model (incorporating variables with statistically significant differences identified through univariate analysis of ultrasound features); (2) a habitat model (incorporating the Hab-score, calculated based on core habitat features screened via random forest with a feature importance $\geq 0.05$); and (3) a combined model (incorporating both the statistically significant ultrasound features and the Hab-score).

Statistical Methods: Data statistical analysis was performed using Python (version 3.11.8). Continuous variables following a normal distribution are expressed as ($\bar{x} \pm s$) and were analyzed using independent samples $t$-tests. Non-normally distributed continuous variables are expressed as $M (Q_1, Q_3)$, with intergroup comparisons conducted using the Mann-Whitney $U$ test. Categorical data are presented as constituent ratios, and intergroup comparisons were performed using the Chi-square test. Receiver Operating Characteristic (ROC) curves were plotted to evaluate the predictive performance of each model, and the DeLong test was employed to compare differences in the Area Under the Curve (AUC) between models. Decision Curve Analysis (DCA) was utilized to assess the clinical utility of the models. Statistical significance was defined as $P < 0.05$.

0.05 为

The difference was statistically significant.

2.1 良性组与交界性

Comparison of General Data and Ultrasound Characteristics Between Groups

A total of 102 patients with phyllodes tumors of the breast (PTB) were included in this study, all of whom were female, with a mean age of $45.7 \pm 12.1$ years. The cohort consisted of 54 cases in the benign group and 48 cases in the borderline/malignant group (comprising 30 borderline and 18 malignant cases). Statistically significant differences were observed between the two groups regarding maximum tumor diameter, boundary characteristics, internal echo patterns, and cystic changes ($P < 0.05$), as detailed in [TABLE:1].

Habitat Feature Selection and Model Construction

A conventional ultrasound model was constructed by incorporating four ultrasound variables that showed significant differences in the univariate analysis: maximum tumor diameter, boundary, internal echo, and cystic changes. Three subregions (habitat_1, habitat_2, and habitat_3) were extracted from the intratumoral region of interest (ROI), yielding a total of 24 habitat features. Following Random Forest (RF) screening, seven habitat features (three first-order features and four texture features) were ultimately retained to calculate the Hab-score and construct the habitat model. Among these, habitat_1_homogeneity contributed the highest weight (+0.081), identifying it as the most significant habitat feature for differentiation.

Note: A–C represent the conventional ultrasound image, ROI, and habitat image of a benign phyllodes tumor, respectively; D–F represent the conventional ultrasound image, ROI, and habitat image of a borderline phyllodes tumor; G–I represent the conventional ultrasound image, ROI, and habitat image of a malignant phyllodes tumor. Habitat_1–3 refer to the tumor subregions 1–3 segmented via K-means clustering analysis.

Examples of ultrasound image segmentation and ultrasound habitat feature extraction. Comparison of general data and ultrasound image features between the benign group and the borderline/malignant group: Parallel growth [n (%)], Boundary [n (%)], Morphology [n (%)], Internal echo [n (%)]. Benign group: $45.9 \pm 12.9$ years, $3.6 (2.6, 4.6)$ cm, 54 (100.0%). Borderline/malignant group: $43.9 \pm 11.4$ years, $5.2 (3.3, 7.4)$ cm, 47 (97.9%), 1 (2.1%), 38 (79.2%), 10 (20.8%), 28 (58.3%), 20 (41.7%), 9 (18.8%), 39 (81.2%). Cystic change [n (%)], Calcification [n (%)], Posterior echo [n (%)], Blood flow signal [n (%)]. Values are presented as mean $\pm$ SD or median (interquartile range); test statistics are $\chi^2$ values unless otherwise noted; "-" indicates Fisher's exact test.

Other prominent habitat features included habitat_3_mean (+0.064) and habitat_2_mean (+0.064), as shown in [TABLE:2].

A comparison of the optimal habitat features and overall habitat characteristics between the two groups revealed that the median values of habitat_2_mean and habitat_3_std were higher in the benign group than in the borderline/malignant group. Conversely, the median values for habitat_1_homogeneity, habitat_3_mean, habitat_2_correlation, habitat_1_correlation, and habitat_1_ASM were lower in the benign group compared to the borderline/malignant group ($P < 0.05$). Furthermore, the median Hab-score for the benign group was significantly lower than that of the borderline/malignant group ($P < 0.05$), as illustrated in [FIGURE:2]. Based on the four conventional ultrasound variables mentioned previously, the Hab-score was further integrated to construct a combined model.

Performance Evaluation of the Models

Receiver Operating Characteristic (ROC) curves were plotted to evaluate the performance of the conventional ultrasound model, the habitat model, and the combined model in differentiating benign from borderline/malignant phyllodes tumors. The results showed that the conventional ultrasound model achieved an Area Under the Curve (AUC) of 0.718 (95% CI = 0.610–0.816), with an accuracy of 67.7%, sensitivity of 58.3%, and specificity of 75.9%. The habitat model yielded an AUC of 0.725 (95% CI = 0.620–0.814), with an accuracy of 61.8%, sensitivity of 68.8%, and specificity of 55.6%. The combined model demonstrated the highest performance, with an AUC of 0.799 (95% CI = 0.712–0.876), an accuracy of 70.6%, sensitivity of 62.5%, and specificity of 77.8%, as shown in [TABLE:3] and [FIGURE:3]. Delong test results indicated that the AUC of the combined model was significantly higher than those of both the conventional ultrasound and habitat models ($P < 0.05$).

Decision Curve Analysis (DCA) indicated that the combined model provided the highest clinical benefit for differentiating benign and malignant PTB within the threshold range of 0.4 to 0.9, as shown in [FIGURE:4].

3 讨论

In recent years, conventional ultrasound has become one of the most widely utilized imaging techniques in clinical practice due to its real-time and non-invasive advantages. However, its potential for quantifying the heterogeneity of phyllodes tumors of the breast (PTB) has not yet been fully explored. Habitat imaging, an innovative technology for quantifying the spatial heterogeneity of the tumor microenvironment, has been extensively applied in the fields of CT and MRI. Note: habitat_1_homogeneity represents the homogeneity of habitat 1; habitat_3_mean, habitat_2_mean, and habitat_3_std represent the mean of habitat 3, the mean of habitat 2, and the standard deviation of habitat 3, respectively ($p < 0.05$).

Boxplots comparing the optimal habitat features and overall habitat features between the benign group and the borderline/malignant group.

Chinese General Practice

Comparison of the performance evaluation of three models in differentiating benign and malignant phyllodes tumors

Note: AUC = Area Under the Receiver Operating Characteristic Curve.

Comparison of AUC values among the conventional ultrasound model, habitat model, and combined model for the diagnosis of benign and malignant phyllodes tumors.

Conventional ultrasound model: 1 - Specificity.

ROC curves of the conventional ultrasound model, habitat model, and combined model for distinguishing benign and malignant phyllodes tumors.

Habitat imaging has been confirmed as an effective auxiliary tool for differentiating benign from malignant tumors and evaluating biological behavior, gaining widespread attention in the field of medical imaging. However, the application of ultrasound habitat imaging in the differentiation of phyllodes tumors of the breast (PTB) has not yet been systematically explored. This study utilized ultrasound images to extract habitat features through sub-region segmentation. By establishing a combined model based on conventional ultrasound features and habitat characteristics, we aimed to improve the diagnostic performance of ultrasound in distinguishing benign and malignant PTB, providing a non-invasive assessment protocol to guide preoperative clinical decision-making.

In the univariate analysis of conventional ultrasound features in this study, significant differences were observed in maximum diameter, internal echoes, margins, and cystic changes ($P < 0.05$). Malignant breast tumors generally exhibit faster growth rates and larger diameters than benign ones. Rapid growth often leads to disordered internal structures, manifested as intratumoral necrosis, hemorrhage, or fibrosis, which appear as heterogeneous echogenicity on ultrasound. [FIGURE: DCA Curves of conventional ultrasound model, habitat model, and combined model for distinguishing benign and malignant phyllodes tumors] Benign PTBs typically possess a complete fibrous capsule, whereas malignant PTBs exhibit stromal overgrowth, significant nuclear atypia, and a mitotic count $\geq 10/10$ high-power fields (HPF). Their infiltrative growth pattern allows them to invade surrounding mammary ducts, lobules, or glandular tissues and breach the capsule, resulting in indistinct margins. According to a study by Cao Yimin on 107 PTB cases, approximately 88.89% of benign PTBs had clear margins, while only 11.11% showed blurred margins, which is consistent with our findings. Furthermore, the rapid proliferation of malignant PTBs leads to an imbalance in blood supply, making them prone to micro-necrosis and hemorrhage; these pathological changes manifest as internal cystic components on ultrasound. Li et al. used logistic regression to evaluate ultrasound features for predicting the malignancy of 79 PTB cases, finding significant differences in age, lesion size, morphology, internal echoes, liquefaction, and blood flow ($P < 0.05$). This differs slightly from our univariate analysis. Given that their sample size ($n=79$) was smaller than ours ($n=102$), multicenter data is still needed for support. Regarding vascularity, the malignant group in our study showed richer blood flow signals (higher proportion of Grade II–III). We hypothesize that stromal overgrowth in borderline/malignant phyllodes tumors causes disordered cell arrangement and irregular internal structures, destroying the original vascular bed and inducing significant immature angiogenesis. While this results in richer signals on Color Doppler, the precise identification of intratumoral blood flow requires further research.

In this study, the tumor ROI was divided into three habitat sub-regions using K-means clustering, and seven optimal habitat features were ultimately retained: habitat_1_homogeneity, habitat_3_mean, habitat_2_mean, habitat_3_std, habitat_2_correlation, habitat_1_correlation, and habitat_1_ASM. Further analysis of the feature distribution revealed that habitat_1_ASM is used to quantify the uniformity and concentration of image texture; higher values indicate greater texture consistency and regularity. This correlates with the relatively slower growth rate, more mature structural differentiation, and more orderly arrangement of benign tumor cells. Furthermore, an increase in habitat_3_std reflects a higher dispersion of grayscale values within malignant lesions.

Chinese General Practice. This higher dispersion may be attributed to the coexistence of malignant tumor cells with internal necrosis and hemorrhage, leading to more pronounced grayscale differences. The elevation of habitat_3_mean and habitat_2_mean indicates higher grayscale levels in the malignant group, which is associated with signal enhancement caused by intensified malignant cell proliferation. Increased levels of habitat_1_homogeneity, habitat_2_correlation, and habitat_1_correlation reflect stronger spatial correlation of grayscale values in malignant tumors. This may be due to continuous abnormal proliferation forming pathological structures (such as tumor nests or abnormal vascular networks), resulting in an abnormal but regular vascular distribution that tends to unify the image texture.

The habitat features mentioned above are all related to the increased tissue heterogeneity caused by disordered proliferation and invasive growth of malignant cells, suggesting that feature differences across different habitat sub-regions can reflect intratumoral heterogeneity. Zhang et al. conducted a retrospective analysis of multi-parametric MRI images from 76 breast cancer patients, quantitatively visualizing differences in vascular distribution and cellular structure within lesions through habitat features. They found that habitat imaging could identify breast cancer heterogeneity, which aligns with our results. However, that study deeply analyzed differences in the volume proportions of different habitats across various HER2 expression levels, whereas ultrasound habitat analysis associated with pathological markers requires further investigation.

This study compared the performance of the conventional ultrasound model, the habitat model, and the combined model in differentiating benign and malignant PTB. The combined model performed the best, surpassing both the conventional ultrasound and habitat models. The DeLong test further confirmed these findings.

The differences in AUC between the combined model and both the conventional ultrasound and habitat models were statistically significant ($P < 0.05$). However, there was no statistically significant difference between the habitat model and the conventional ultrasound model ($P > 0.05$). This suggests that the diagnostic value of relying solely on conventional ultrasound or habitat features is limited; a breakthrough in diagnostic performance requires the integration of multiple features. In a study by Xu et al. using habitat analysis on MRI images of 142 breast cancer patients to identify triple-negative breast cancer, the combined model achieved an AUC of 0.951, demonstrating extremely high diagnostic efficacy. This differs from our results, likely due to the inherent characteristics of the imaging modalities. As a multi-parametric functional imaging technique based on tumor microcirculation perfusion and water molecule diffusion, MRI can quantify the physiological state of the tumor microenvironment at a microscopic level. Thus, MRI habitat partitioning more closely reflects the essential characteristics of tumor biological behavior. In contrast, ultrasound imaging is based on the principles of acoustic reflection and scattering; its grayscale texture features only indirectly reflect differences in tissue density and structural arrangement. Furthermore, Color Doppler parameters can only capture blood flow in vessels with a diameter $> 0.1$ mm, failing to directly assess microscopic angiogenesis and cellular/molecular features, which results in lower biological specificity of ultrasound habitat features compared to MRI. Additionally, the sensitivity of the conventional ultrasound model was only 58.3%, reflecting a risk of underdiagnosing borderline/malignant PTB. While the habitat model's sensitivity improved to 68.8%—suggesting that quantifying intratumoral habitat features can more sensitively capture the biological characteristics of increased heterogeneity in malignant tumors—its specificity was only 55.6%. This indicates that some benign PTBs may be confused with malignant ones due to internal changes like cystic degeneration. Therefore, the complementary strengths of the two single models highlight the significant advantage of feature integration in the combined model. Yu Lihui et al. constructed an ultrasound + texture feature model based on 298 PTB patients, achieving AUCs of 0.825 and 0.818 in the training and testing sets, respectively. These are slightly higher than the AUC of our combined model (0.799), likely because their significantly larger sample size reduced random error and improved the model's ability to capture independent predictors. Furthermore, unlike our study, Yu Lihui et al. utilized internal validation. Despite these different strategies, our combined model's AUC of 0.799 remains highly comparable, reflecting its good discriminatory ability. From a clinical perspective, when palpation or conventional ultrasound detects a breast mass suspicious for PTB that is difficult to classify as benign or malignant, ultrasound habitat imaging can be introduced as a combined diagnostic tool to overcome the technical limitations of traditional ultrasound in assessing tumor heterogeneity. However, this study has limitations: (1) The sample size is limited and based on a single-center retrospective analysis; larger, multicenter studies are needed to verify the model's generalizability. (2) The ultrasound images were two-dimensional; constructing a three-dimensional model could enrich habitat features and more comprehensively reflect PTB heterogeneity. (3) Images were acquired using Mindray DC-8 and GE Logiq E9 systems. Although core parameters like probe frequency (3–12 MHz) and depth were standardized, inherent differences in resolution, grayscale processing algorithms, and color gain optimization exist between brands. Additionally, subtle adjustments by different operators during image acquisition may introduce bias.

In conclusion, this study effectively applied ultrasound habitat imaging to differentiate benign and malignant phyllodes tumors of the breast. By combining these with conventional ultrasound features, diagnostic performance was further enhanced, reducing the risk of missed diagnoses or misdiagnoses associated with a single technical approach. This provides a new non-invasive technical pathway for the ultrasound assessment of PTB heterogeneity.

Author Contributions: Xie Danling was responsible for the implementation of the study and writing the manuscript; Xie Danling and Liu Boya performed ultrasound data collection, organization, statistical processing, figure preparation, and manuscript revision; Li Xiaoguang participated in the experimental design; Wang Hanwei participated in statistical analysis; Ma Qiang was responsible for providing pathological data; Fang Jingqin participated in the revision of the research protocol and manuscript; Wang Shunan was responsible for experimental design, quality control, review, and supervision.

The authors declare no conflicts of interest.

参考文献

[1] TAN P H, ELLIS I, ALLISON K, et al. The 2019 World Health Organization classification of tumours of the breast[J]. Histopathology, 2020, 77(2): 181-185. DOI: 10.1111/his.14091. [2] CHEN C, SUN Q, LI Y. Expert consensus on the diagnosis and treatment of phyllodes tumors of the breast in Chinese women[J]. Chinese Journal of Breast Disease (Electronic Edition), 2023, 17(4): 201-209. DOI: 10.3877/cma.j.issn.1674-0807.2023.04.001. [3] NIU S H, HUANG J H, LI J, et al. Differential diagnosis between small breast Phyllodes tumors and fibroadenomas using artificial intelligence and ultrasound data[J]. Quant Imaging Med Surg, 2021, 11(5): 2052-2061. DOI: 10.21037/qims-20-919.

[4] CHENG W Q, QI X, YANG H K, et al. Identification of Luminal and non-Luminal breast cancer based on multi-parameter MR habitat imaging[J]. Chinese Journal of Magnetic Resonance Imaging, 2025, 16(05): 170-180. DOI: 10.12015/issn.1674-8034.2025.05.026. [5] LIU Z N, MOUNI D, ZHANG S M, et al. Predicting the early response to neoadjuvant chemotherapy in high-grade serous ovarian cancer by intratumoral habitat heterogeneity based on 18F-FDG PET/CT[J]. Eur J Nucl Med Mol Imaging, 2025. DOI: 10.1007/s00259-025-07480-z. [6] GONG Z J, LIU Z X, HUANG K Y, et al. Habitat analysis based on magnetic resonance imaging for the prediction of prostate cancer: a dual-center study[J]. Quant Imaging Med Surg, 2025, 15(9): 8395-8408. DOI: 10.21037/qims-2025-223. [7] WANG Y, DAI A, WEN Y, et al. Prediction of high-risk capsule characteristics for recurrence of pleomorphic adenoma in the parotid gland based on habitat imaging and peritumoral radiomics: a two-center study[J]. Acad Radiol, 2025, 32(7): 4134-4145. DOI: 10.1016/j.acra.2024.11.025. [8] ADLER D D, CARSON P L, RUBIN J M, et al. Doppler ultrasound color flow imaging in the study of breast cancer: preliminary findings[J]. Ultrasound Med Biol, 1990, 16(6): 553-559. DOI: 10.1016/0301-5629(90)90020-d. [9] HUANG Y J, CHEN Y J, ZENG P Y, et al. Comparative study of ultrasound images of different pathological subtypes of breast phyllodes tumors[J]. Chinese Journal of Ultrasound Medicine, 2024, 40(04): 410-413. [10] PACKER M D C, LESTER S C. Current understanding of phyllodes tumors of the breast: Tumor classification, molecular landscape, and best pathology practice[J]. Hum Pathol, 2025, 162: 105863. DOI: 10.1016/j.humpath.2024.105863. [11] BASARA AKIN I, OZGUL H, SIMSEK K, et al. Texture analysis of ultrasound images to differentiate simple fibroadenomas from complex fibroadenomas and benign Phyllodes tumors[J]. J Ultrasound Med, 2020, 39(10): 1993-2003. DOI: 10.1002/jum.15304. [12] CAO Y M, ZHOU S P, HU T, et al. Analysis of ultrasound imaging characteristics of benign and malignant breast phyllodes tumors[J]. Medical Recapitulate, 2018, 24(24): 4970-4973. DOI: 10.14033/j.cnki.cfmr.2018.24.038. [13] LI T, LI Y, YANG Y, et al. Logistic regression analysis of ultrasound findings in predicting the malignant and benign phyllodes tumor of the breast[J]. PLoS One, 2022, 17(4): e0265952. DOI: 10.1371/journal.pone.0265952. [14] TSUCHIYA M, MASUI T, TERAUCHI K, et al. MRI-based radiomics analysis for differentiating phyllodes tumors of the breast from fibroadenomas[J]. Eur Radiol, 2022, 32(6): 4090-4100. DOI: 10.1007/s00330-021-08510-8. [15] JING S, WANG H, LIN P, et al. Quantifying and interpreting biologically meaningful spatial signatures within tumor microenvironments[J]. NPJ Precis Oncol, 2025, 9(1): 68. DOI: 10.1038/s41698-025-00857-1. [16] ZHANG X, CHEN X, FU Y, et al. Study on heterogeneity of vascularity and cellularity via multiparametric MRI habitat imaging in breast cancer[J]. BMC Medical Imaging, 2025, 25(1). DOI: 10.1186/s12880-025-01698-x. [17] XU R, YU D, LUO P, et al. Do habitat MRI and fractal analysis help distinguish triple-negative breast cancer from non-triple-negative breast carcinoma[J]. Can Assoc Radiol J, 2024, 75(3): 584-592. DOI: 10.1177/08465371241231573. [18] YU L H, WANG N N, CHEN Y Y, et al. Predictive value of ultrasound features combined with texture analysis for the malignancy risk of breast phyllodes tumors[J]. Chinese Journal of Ultrasound Medicine, 2025, 41(02): 150-154. (Received: 2025-10-10; Revised: 2025-10-31) (Editor: LI Weixia)