Prediction of neoadjuvant chemotherapy efficacy in patients with HER2-low breast cancer based on ultrasound radiomics

Abstract

Background

Neoadjuvant chemotherapy (NAC) is a crucial therapeutic approach for treating breast cancer, yet accurately predicting treatment response remains a significant clinical challenge. Conventional ultrasound plays a vital role in assessing tumor morphology but lacks the ability to quantitatively capture intratumoral heterogeneity. Ultrasound radiomics, which extracts high-throughput quantitative imaging features, offers a novel approach to enhance NAC response prediction. This study aims to evaluate the predictive efficacy of ultrasound radiomics models based on pre-treatment, post-treatment, and combined imaging features for assessing the NAC response in patients with HER2-low breast cancer.

Methods

This retrospective multicenter study included 359 patients with HER2-low breast cancer who underwent NAC between January 1, 2016, and December 31, 2020. A total of 488 radiomic features were extracted from pre- and post-treatment ultrasound images. Feature selection was conducted in two stages: first, Pearson correlation analysis (threshold: 0.65) was applied to remove highly correlated features and reduce redundancy; then, Recursive Feature Elimination with Cross-Validation (RFECV) was employed to identify the optimal feature subset for model construction. The dataset was divided into a training set (244 patients) and an external validation set (115 patients from independent centers). Model performance was assessed via the area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and F1 score.

Results

Three models were initially developed: (1) a pre-treatment model (AUC = 0.716), (2) a post-treatment model (AUC = 0.772), and (3) a combined pre- and post-treatment model (AUC = 0.762).To enhance feature selection, Recursive Feature Elimination with Cross-Validation was applied, resulting in optimized models with reduced feature sets: (1) the pre-treatment model (AUC = 0.746), (2) the post-treatment model (AUC = 0.712), and (3) the combined model (AUC = 0.759).

Conclusions

Ultrasound radiomics is a non-invasive and promising approach for predicting response to neoadjuvant chemotherapy in HER2-low breast cancer. The pre-treatment model yielded reliable performance after feature selection. While the combined model did not substantially enhance predictive accuracy, its stable performance suggests that longitudinal ultrasound imaging may help capture treatment-induced phenotypic changes. These findings offer preliminary support for individualized therapeutic decision-making.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40644-025-00934-5.

Introduction

Breast cancer remains the most prevalent malignancy among women worldwide, with increasing incidence and mortality rates, particularly in resource-limited regions [1, 2]. It is a heterogeneous disease comprising distinct molecular subtypes [3], each exhibiting unique biological behaviors and therapeutic responses. Among these, HER2 expression has garnered particular attention due to its strong association with variations in clinical management strategies and survival outcomes [4, 5].

Traditionally, HER2 expression in breast cancer has been dichotomized as positive (IHC 3 + or IHC 2+/ISH+) or negative (IHC 0, 1+, or 2+/ISH−). However, the recent approval of antibody-drug conjugates (ADCs), particularly trastuzumab deruxtecan (T-DXd), has led to the recognition of HER2-low breast cancer as a distinct clinical entity, accounting for approximately 45–55% of all breast cancer cases [6–9]. In addition to their limited response to traditional anti-HER2 therapies, HER2-low tumors exhibit distinct clinicopathological characteristics and treatment outcomes compared to HER2-0 tumors. Importantly, several studies have reported significantly different rates of pathologic complete response (pCR) and long-term survival between HER2-low and HER2-0 patients undergoing neoadjuvant chemotherapy (NAC) [6, 10]. These findings highlight the clinical importance of NAC in HER2-low breast cancer and suggest that HER2-low tumors may represent a unique biological subset requiring tailored treatment evaluation. This underscores the need for more accurate and noninvasive tools to predict NAC response specifically in this population, which is the focus of the present study.

NAC is a standard treatment strategy for locally advanced and operable breast cancer [11, 12], offering the dual benefits of tumor downsizing to facilitate breast-conserving surgery and providing early prognostic insights on the basis of treatment response. However, current assessment methods—including clinical markers and conventional imaging—often fail to reliably predict NAC response, particularly in biologically diverse subtypes such as HER2-low breast cancer [13]. In China, ultrasound is the most widely used and accessible modality for breast cancer screening. It is commonly applied either alone or in combination with Magnetic Resonance Imaging (MRI) to evaluate the response to neoadjuvant chemotherapy. Owing to its advantages of low cost, convenience, absence of ionizing radiation, and high specificity, ultrasound is widely adopted, especially for patients who may not undergo timely MRI evaluation due to financial or logistical constraints. It allows dynamic and accurate assessment of tumor location, size, morphology, margins, and vascularity.

Despite its widespread use, ultrasound’s diagnostic accuracy is limited by operator dependency and subjective interpretation—highlighting the need for objective, reproducible analytic approaches [14].

In addition, the imaging characteristics of HER2-low tumors remain poorly understood, and there is limited research on noninvasive biomarkers to predict treatment efficacy. Recent studies in radiomics, particularly ultrasound-based radiomics, present a promising approach for overcoming these limitations [15]. Radiomics involves the high-throughput extraction of quantitative imaging features that describe tumor texture, morphology, and functional properties [16, 17]. By quantifying tumor heterogeneity and treatment-induced changes, radiomics facilitates an objective assessment of therapeutic response [18]. Emerging evidence suggests that ultrasound radiomics features, including echogenicity, vascular patterns, and elasticity-related parameters, correlate with NAC response [19, 20]. Nonetheless, clinical implementation of ultrasound radiomics remains limited by heterogeneity in imaging protocols, suboptimal feature selection methods, and a lack of external validation across large patient cohorts [13].

Based on the above clinical background, this multicenter retrospective study leverages a dataset of 359 patients with HER2-low breast cancer with pre- and post-NAC US images to develop a robust radiomics-based predictive model. Using recursive feature elimination with cross-validation, we identified optimal feature subsets that capture both baseline tumor biology and dynamic treatment responses [21, 22]. Our goal is to standardize radiomics workflows, enhance predictive accuracy, and facilitate personalized NAC treatment strategies, ultimately reducing overtreatment and improving patient outcomes [23, 24].

Methods

Patient population

This study retrospectively analysed ultrasound images from 359 patients with HER2-low breast cancer who underwent neoadjuvant chemotherapy between January 1, 2016, and December 31, 2020. Patients were categorized into two groups on the basis of their Overall Response Rate (ORR): the good response group (complete response [CR] and partial response [PR]) and the poor response group (progressive disease [PD] and stable disease [SD]) [25–27]. The dataset was divided into a training set and an external validation set, ensuring the model’s generalizability across independent institutions. The training dataset was obtained from the First Hospital of China Medical University. The external validation dataset was collected from three independent institutions: the Fourth Affiliated Hospital of China Medical University, the First Affiliated Hospital of Jinzhou Medical University, and Affiliated Zhongshan Hospital of Dalian University.

Patients were included in this study if they met all of the following conditions: (1) had undergone surgical treatment for breast cancer; (2) had complete clinical information, including postoperative pathological findings and immunohistochemical results; and (3) had clear, high-quality, and diagnostically interpretable ultrasound images obtained both pre- and post- neoadjuvant chemotherapy. The exclusion criteria were as follows: (1) patients with bilateral invasive breast cancer; (2) patients with discordant HER2 status between initial diagnosis and postoperative biopsy; (3) patients who discontinued NAC due to severe neutropenia and proceeded directly to surgery; (4) patients who received additional neoadjuvant endocrine therapy alongside NAC; (5) patients diagnosed with breast cancer subtypes other than invasive ductal carcinoma or invasive lobular carcinoma; and (6) patients without postoperative follow-up data or those who did not provide informed consent.

Standardized ultrasound examinations were performed by board-certified sonographers (≥ 5 years of experience) via four different ultrasound systems: Philips Epiq7 (5–12 MHz), Canon Aplio i900 (5–18 MHz), Supersonic Imagine (4–15 MHz), and Hitachi Ascendus (5–13 MHz). Pre- and post-NAC grayscale images of the breast tumors were retrieved from picture archiving and communication system (PACS) workstations with operators blinded to the pathological outcomes. The tumor margins were manually delineated by a board-certified radiologist via ITK-SNAP software (v3.8.0) [28, 29], incorporating acoustic shadowing regions for volumetric analysis. Longitudinal image pairing was conducted to track tumor morphological evolution throughout NAC cycles.

The image preprocessing protocol included the following standardization steps: (1) Unification of the DICOM format to maintain consistency across multicenter datasets; (2) Rigorous quality control to exclude images with artifacts and incomplete tumor margins (> 10% circumference undefined); and (3) Grayscale normalization via a linear transformation of pixel intensities (0–1 scale) on the basis of dataset-wide extremal values (0.5th-99.5th percentiles), reducing inter-scanner variability while preserving relative intensity distributions. This standardized preprocessing framework ensured quantitative comparability for subsequent radiomic feature analysis.

Feature selection and model development

Radiomic features were extracted from both pre- and post-treatment ultrasound images via the Pyradiomics package (Python) [30, 31]. A total of 488 features were initially extracted, encompassing first-order statistics, shape descriptors, and texture-based features [32, 33] (See Supplementary Material 1 and Supplementary Material 2). Features with zero variance were removed, yielding 486 remaining features. Subsequent feature correlation analysis eliminated highly correlated features (Pearson correlation coefficient > 0.65) [34], resulting in 42 retained features for pre-treatment images and 44 for post-treatment images, forming the basis of the first two experiments.

To optimize feature selection while maintaining model generalizability, Recursive Feature Elimination with Cross-Validation was applied in three experimental settings [35, 36].

In the first experiment, RFECV was used to select diagnostically relevant features from pre-treatment ultrasound images. The second experiment was conducted on post-treatment ultrasound radiomic features to capture distinct biomarker signatures indicative of treatment-induced changes. For the third experiment, RFECV optimization was applied to a combined set of pre- and post-treatment features. This systematic approach ensures robust feature selection, guided by cross-validation performance to prevent premature exclusion of informative predictors. To enhance clarity, a detailed flow diagram illustrating the RFECV process and experimental design is provided in Fig. 1.

Fig. 1.

Workflow of RFECV for ultrasound radiomic feature selection and model building. Overview of the RFECV-based radiomics workflow applied to three modeling scenarios: pre-treatment features, post-treatment features, and combined pre- and post-treatment features. The workflow includes feature extraction from ultrasound images, initial feature filtering, and recursive feature elimination with cross-validation for optimal feature subset selection. This pipeline aims to identify reproducible and predictive radiomic biomarkers to enhance the prediction of neoadjuvant chemotherapy response in HER2-low breast cancer patients

The use of RFECV in this context offers critical advantages: it systematically reduces the high-dimensional feature space by iteratively removing less informative features on the basis of cross-validated model performance, thus mitigating overfitting—a common issue in radiomic data characterized by large feature sets but limited sample sizes. The random forest (RF) model was selected for its robustness in handling high-dimensional data, resistance to overfitting, and suitability for radiomic features with complex interactions [36–38]. RF also provides interpretable feature importance measures crucial for recursive feature elimination. Stratified 5-fold cross-validation was used during feature elimination to balance the class distribution and ensure model stability, offering an effective compromise between computation and robustness given the dataset size [39].

After feature selection, the refined feature subsets were used to retrain Random Forest models for NAC efficacy prediction. The performance of RFECV was assessed by plotting the cross-validation accuracy across different feature subset sizes, while a feature importance ranking was generated to highlight the most predictive features. These steps were performed via Python 3.9.12 (scikit-learn).

Finally, predictive models were constructed via R 4.4.2, where Random Forest classifiers were trained on the selected feature subsets. Model evaluation was based on the AUC, accuracy, precision, recall, F1-score, and confusion matrix analysis [40], ensuring a comprehensive assessment of predictive performance. AUC values and ROC curves were generated in R to further assess classification performance.

Feature reproducibility analysis

To assess the reproducibility of radiomics features extracted from ultrasound images, a test-retest reliability analysis was conducted [41]. Thirty patients with HER2-low breast cancer were randomly selected, and tumor regions were independently delineated by both an experienced radiologist and the researcher. Feature extraction was performed on manually segmented regions of interest (ROIs) from both pre- and post-neoadjuvant chemotherapy images, resulting in 488 radiomics features for each dataset. A schematic overview of the reproducibility filtering process is presented in Fig. 2 to illustrate the independent application of these criteria.

Fig. 2.

Workflow of radiomic feature reproducibility filtering. Workflow summarizing the reproducibility filtering of radiomic features. Feature reproducibility was evaluated using multiple complementary criteria, including mean difference, standard deviation difference, and coefficient of variation. Stable features under each criterion were reported individually. This comprehensive assessment ensures robust and reproducible feature selection for chemotherapy response prediction

The reproducibility of these features was initially evaluated by calculating the mean difference and standard deviation difference between the features derived from the two delineation sets. The mean difference reflects the average deviation between corresponding features, while the SD difference quantifies the variability of these differences. These metrics are critical for understanding the consistency of feature extraction across different delineation methods.

A reproducibility threshold was then applied to identify stable features. Specifically, features with a mean difference exceeding 0.05 or an SD difference greater than 0.1 were classified as unstable and excluded from further analysis [42–44]. This methodology provides a clear and objective criterion for determining feature reliability. In addition, the coefficient of variation for each feature was calculated across both delineation sets [45]. The CV, defined as the ratio of the standard deviation to the mean, is a widely used metric to assess the relative variability of features. Features with a CV value less than 10% were considered reproducible, indicating minimal variability between the two delineation sets [46, 47].

Statistical analysis

All the statistical analyses were performed via Python 3.9.12 and R 4.4.2. Key Python libraries included NumPy, Pandas, and scikit-learn, while R packages such as caret and ggplot2 were used for model evaluation and visualization [48, 49]. Descriptive statistics were computed for patient demographics and clinical data. Performance metrics, including Area Under the Curve, accuracy, precision, recall, and F1-score, were calculated for each experimental model. The Random Forest model’s feature importance was evaluated via scikit-learn’s built-in feature importance function. The optimal model was determined on the basis of the highest AUC value. All analyses followed institutional ethical guidelines and were approved by the relevant institutional review board. Feature reproducibility analysis was completed prior to model development to ensure the reliability of the features used [50].

Results

Patient selection and data acquisition

After applying the inclusion and exclusion criteria, 359 patients with HER2-low breast cancer were included in the final cohort. Patients were classified into a good response group (PR/CR, n = 223) and a poor response group (PD/SD, n = 136) on the basis of the overall response rate [25–27]. The dataset was randomly split into a training set (70%) and an external test set (30%). The training set consisted of 244 patients (155 in the good response group and 89 in the poor response group), whereas the external test set included 115 patients (68 in the good response group and 47 in the poor response group). No significant differences in baseline characteristics, including age and tumor size, were observed between the two sets, ensuring comparability (P > 0.05). The patient enrollment process is illustrated in Fig. 3.

Fig. 3.

Patient enrollment process. Patient enrollment diagram summarizing inclusion criteria and allocation into training and testing cohorts. This flowchart ensures transparent and traceable composition of the study dataset

Model performance

Experiment 1 (Pre-treatment features, 42 features)

The model based on pre-treatment features, consisting of 42 features, achieved an AUC of 0.716, indicating moderate predictive capability (Table 1). The precision-recall AUC (AUC-PR) of 0.803 (Fig. 4a) suggests stable performance in distinguishing responders from non-responders. However, the confusion matrix (Fig. 5a) reveals a relatively high misclassification rate, implying that pre-treatment features alone may not provide sufficient discriminatory power for accurate treatment response prediction. While these features offer initial insights into treatment efficacy, their predictive performance remains limited compared to other models. The ROC curve (Fig. 6a) further underscores this moderate performance, suggesting that pre-treatment features alone may not capture the full variability required for robust neoadjuvant chemotherapy outcome prediction. Specific features are listed in Supplementary Material 3.

Table 1.

Model performance summary pre- and post-RFECV; P values indicate DeLong test results comparing AUCs

Experiment	AUC	Precision	Recall	F1 Score	Accuracy	P
Experiment 1 (Pre-treatment)	0.716 (0.621–0.805)	0.663(0.561–0.761)	0.809(0.710–0.900)	0.729(0.643–0.802)	0.644(0.557–0.730)	Ref
RFECV Experiment 1 (Pre-treatment)	0.746(0.656–0.834)	0.706(0.611–0.803)	0.882(0.791–0.950)	0.784(0.708–0.849)	0.713(0.626–0.791)	0.633
Experiment 2 (Post-treatment)	0.772 (0.691–0.851)	0.674(0.575–0.763)	0.912(0.833–0.972)	0.775(0.694–0.841)	0.687(0.591–0.765)	Ref
RFECV Experiment 2 (Post-treatment)	0.712(0.618–0.780)	0.682(0.583–0.783)	0.882(0.800-0.956)	0.769(0.694–0.843)	0.687(0.600-0.774)	0.249
Experiment 3 (Combined)	0.762(0.669–0.845)	0.677(0.573–0.778)	0.927(0.862–0.985)	0.783(0.705–0.852)	0.696(0.609–0.783)	Ref
RFECV Experiment 3 (Combined)	0.759(0.667–0.843)	0.682(0.583–0.776)	0.882(0.797–0.956)	0.769(0.685–0.842)	0.687(0.600-0.774)	0.955

Fig. 4.

Precision-recall curve. Precision-recall curves for the three experiments, demonstrating the trade-off between precision and recall across classification thresholds. These curves are particularly useful in imbalanced datasets, where the area under the precision-recall curve (AUC-PR) may provide more informative performance evaluation than ROC curves

Fig. 5.

Confusion matrix. Confusion matrices showing the classification performance of each model. The matrices display the numbers of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN), illustrating how well each model differentiates between good and poor responders to neoadjuvant chemotherapy. These visualizations help assess diagnostic accuracy, recall, and misclassification trends

Fig. 6.

Area under the receiver operating characteristic curve. Receiver operating characteristic (ROC) curves for the three experiments, showing the trade-off between sensitivity and specificity. AUC values indicate the discriminative ability of each model. This figure compares the performance of models using pre-treatment, post-treatment, and combined features

Experiment 2 (Post-treatment features, 44 features)

The model incorporating post-treatment features, with a total of 44 features, achieved an improved AUC of 0.772, demonstrating a significant enhancement in predictive performance (Table 1). The precision-recall curve (Fig. 4b) indicated an increased AUC-PR of 0.819, reflecting a greater ability to distinguish responders from non-responders. Additionally, the confusion matrix (Fig. 5b) revealed a reduction in misclassification, underscoring the superior predictive capacity of post-treatment imaging in assessing treatment efficacy. The ROC curve (Fig. 6b) further corroborated these findings, reaffirming that post-treatment features offer greater discriminative power in predicting therapeutic outcomes. These results highlight the importance of post-treatment imaging in capturing tumor response dynamics, which is crucial for evaluating the efficacy of neoadjuvant chemotherapy. Specific features are listed in Supplementary Material 4.

Experiment 3 (Combined pre- and post-treatment features, 86 features)

The combined model, which integrates both pre-treatment and post-treatment features (86 features in total), achieved an AUC of 0.762 (Table 1) and an AUC-PR of 0.808 (Fig. 4c). These metrics are slightly lower than those of the post-treatment-only model, indicating no clear performance advantage from combining feature sets. While the confusion matrix (Fig. 5c) shows some improvement in classification accuracy compared to the pre-treatment-only model, this gain is less pronounced than that observed in the post-treatment-only model, suggesting that pre-treatment features may contribute limited additional predictive value. The ROC curve (Fig. 6c) further reflects a slightly lower AUC compared to the post-treatment-only model. Nonetheless, the combined model may still offer a more comprehensive perspective by incorporating temporal information from both imaging timepoints, which could have potential utility in personalized treatment strategies. A detailed list of the combined features is provided in Supplementary Material 5.

To further refine the model’s performance, Recursive Feature Elimination with Cross-Validation (RFECV) was applied to each experimental configuration.

RFECV experiment 1 (Pre-treatment, 29 selected features)

RFECV applied to the pre-treatment model, selecting 29 key features, resulted in an improved AUC from 0.716 to 0.746 (Table 1). Although the AUC-PR exhibited a slight decline from 0.803 to 0.775 (Fig. 4a vs. Fig. 4d), this finding indicates a trade-off between overall discriminatory power and precision-recall balance. The confusion matrix comparison (Fig. 5a vs. Fig. 5d) demonstrated a reduction in misclassification rates, suggesting enhanced model robustness and stability following feature selection. The ROC curves (Fig. 6a vs. Fig. 6d) supported the observed improvement, with the DeLong test confirming no statistically significant difference (P = 0.633), emphasizing that feature selection preserved predictive performance while potentially improving generalizability. These results underscore the value of carefully selecting the most informative features to refine the pre-treatment model for predicting neoadjuvant chemotherapy response. Specific features are listed in Supplementary Material 6.

RFECV experiment 2 (Post-treatment, 21 selected Features)

RFECV applied to the post-treatment model, resulting in 21 selected features, led to a decrease in the AUC from 0.772 to 0.712 (Table 1). The AUC-PR also decreased from 0.792 to 0.762 (Fig. 4b vs. Fig. 4e). The confusion matrix (Fig. 5b vs. Fig. 5e) revealed an increase in misclassification, and the ROC curve comparison (Fig. 6b vs. Fig. 6e) confirmed the reduction in the AUC, with a non-significant difference indicated by the DeLong test (P = 0.249). These results indicate a decline in performance following feature selection. Specific features are listed in Supplementary Material 7.

RFECV experiment 3 (Combined, 67 selected Features)

The combined model refined through RFECV, incorporating 67 selected features from both pre-treatment and post-treatment datasets, achieved an AUC of 0.759 (Table 1), representing the highest performance among the three experimental groups. The area under the precision-recall curve (AUC-PR) also increased from 0.808 to 0.812 (Fig. 4c vs. Fig. 4f), indicating improved balance between precision and recall. The confusion matrix (Fig. 5c vs. Fig. 5f) showed a further reduction in misclassification compared to the model using the full set of 86 features. The ROC curve comparison (Fig. 6c vs. Fig. 6f) confirmed the stability of the model after feature selection, with no statistically significant difference observed (P = 0.955, DeLong test). These results suggest that the integration of both time points, when optimized through feature selection, can contribute to more consistent classification performance. Specific features are listed in Supplementary Material 8.

Feature importance

To assess the impact of feature selection on model performance, Figs. 7 and 8 presents two visualizations for each experimental group. The top panel illustrates the performance curve of the model utilizing the optimal feature set, which was selected through recursive feature elimination with cross-validation. The bottom panel features a plot of feature importance, highlighting the most influential predictors of neoadjuvant chemotherapy efficacy. This visualization reveals key features that balance model complexity with predictive accuracy. Notable predictors include wavelet-HH first-order Median and original glcm Small Dependence Low Gray Level Emphasis, both of which consistently demonstrated significance across pre-treatment and post-treatment models. These features reflect the model’s ability to capture tumor heterogeneity. Furthermore, wavelet-transformed and gray-level co-occurrence matrix-derived features, such as wavelet-HL glcm MCC and original shape Elongation, were crucial for the model’s predictive performance, underscoring the importance of texture and shape features in radiomics. The bottom panel of Figs. 7 and 8 emphasizes how these features contribute to the model’s predictions. By selecting these key features, the model is optimized, reducing dimensionality while retaining critical information, thereby enhancing both interpretability and generalizability. This approach ensures the model’s clinical applicability without unnecessary complexity.

Fig. 7.

Model performance curves derived from RFECV experiment 1-3. Model performance curves for the three RFECV experiments based on different feature sets: (a) pre-treatment features, (b) post-treatment features, and (c) combined pre- and post-treatment features

Fig. 8.

Feature importance derived from RFECV experiment 1-3. Feature importance rankings from RFECV for the three feature sets. Bar plots represent the relative importance scores of selected features, identifying key imaging biomarkers related to treatment response prediction in HER2-low breast cancer

Feature reproducibility analysis

The test-retest reliability analysis (TRRA) showed a reduction in the number of stable radiomics features after neoadjuvant chemotherapy. On the basis of the thresholds for mean and standard deviation differences, 299 out of the 488 features were stable in the pre-treatment dataset, whereas 121 features were stable in the post-treatment dataset. This decrease suggests that the tumor’s response to NAC may contribute to increased variability in radiomics features, possibly due to chemotherapy-induced morphological and textural changes. Specific features are listed in Supplementary Material 9 and Supplementary Material 10.

The coefficient of variation analysis revealed that 470 features exhibited a CV below 10% in the pre-treatment dataset, indicating high reproducibility. After treatment, only 100 features maintained a CV below the 10% threshold. This reduction in stable features after NAC suggests that chemotherapy-induced changes in the tumor could affect the reproducibility of radiomics features, leading to greater variability. Specific features are listed in Supplementary Material 11 and Supplementary Material 12.

The density plot of the mean differences indicated that the majority of features had minor deviations between the two delineation sets, indicating high stability (Fig. 9). Similarly, the density plot of the SD differences confirmed that most features exhibited consistent variability, though some features displayed significant discrepancies, particularly in the post-treatment dataset, reflecting the influence of NAC (Fig. 10).

Fig. 9.

Density plot of mean differences. Density plots of mean differences in radiomic features between two independent delineations. Panel (a) shows pre-treatment, and panel (b) shows post-treatment datasets. Narrow distributions with peaks near zero suggest higher feature reproducibility

Fig. 10.

Density plot of standard deviation differences. Density plots of SD differences in radiomic features between independent delineations, used to evaluate consistency. Panel (a) represents pre-treatment, and panel (b) post-treatment datasets. Features with SD values concentrated near zero are considered more stable

The box plot comparing the mean and SD differences provided a visual representation of the feature variability (Fig. 11). Most features remained within an acceptable range, but a subset, especially in the post-treatment dataset, exhibited noticeable variability. This trend was further reflected in the bar plot of the CV values, where the top 30 features with the lowest CV demonstrated minimal variation between the delineation sets, highlighting their consistency under test-retest conditions (Fig. 12).

Fig. 11.

Box plot of mean vs. Standard deviation differences. Box plots of mean and SD differences across delineations. Panel (a) pre-treatment; panel (b) post-treatment. Features with smaller interquartile ranges show better reproducibility across segmentations

Fig. 12.

Top 30 Features with CV. Bar plots of CV values for the top 30 most stable radiomic features. Panel (a) pre-treatment; panel (b) post-treatment. CV is calculated as the ratio of standard deviation to the mean; features with lower CV are considered more stable and robust for modeling

Discussion

This study presents an innovative approach for predicting the efficacy of neoadjuvant chemotherapy in patients with HER2-low breast cancer via ultrasound radiomics. The results underscore the significant potential of both post-treatment ultrasound features and the combination of pre- and post-treatment features for predicting chemotherapy response. Ultrasound radiomics provides a non-invasive, cost-effective alternative to traditional biomarkers, which typically rely on invasive procedures. This method shows potential in predicting treatment response and patient prognosis, thereby supporting personalized therapeutic strategies and informing subsequent clinical decision-making.

A key strength of this study lies in the application of RFECV. This technique optimizes the feature selection process by systematically removing the least informative features, thereby enhancing model predictive power. Through rigorous feature selection, RFECV ensures that only the most relevant features are retained, leading to improved model accuracy, robustness, and generalizability [36, 37, 51, 52]. Moreover, RFECV reduces the risk of overfitting, which is a common challenge in radiomics research, thereby making the model more applicable to external datasets and diverse clinical settings [53].

Additionally, the feature reproducibility analysis performed in this study adds to its strength. By evaluating the stability of radiomics features pre and post-treatment, this analysis demonstrates that certain features maintain consistent reproducibility, even after chemotherapy. This finding suggests that ultrasound radiomics, especially when guided by reproducible features, can be a reliable tool for assessing treatment efficacy. Importantly, the analysis also indicates that NAC influences feature variability, highlighting the need for further research to better understand the impact of chemotherapy on feature stability. Such studies will ensure the continued relevance of ultrasound radiomics in clinical practice.

While post-treatment ultrasound features initially showed the highest AUC before feature selection, a key observation from this study is the performance improvement of the pre-treatment model after RFECV, with the AUC increasing from 0.716 to 0.746. This suggests that pre-treatment features—when effectively filtered for relevance—can provide meaningful predictive information for assessing chemotherapy response prior to treatment initiation. In contrast, the decrease in the AUC for the post-treatment model after RFECV may reflect a reduction in available and stable features due to tumor shrinkage or disappearance after chemotherapy. These findings highlight the practical value of pre-treatment imaging in early decision-making and emphasize the importance of robust feature selection in enhancing model performance. Although post-treatment features appeared promising in the initial model, differences among the models were not statistically significant, and thus should be interpreted with caution.

The external validation conducted in this study further reinforces the robustness and clinical applicability of the proposed models. The validation process, which involved ultrasound data from multiple centers, confirmed the ability of the model to maintain predictive accuracy across diverse patient populations and imaging protocols. This cross-center evaluation strengthens the argument that ultrasound radiomics can be effectively employed in clinical practice, enhancing its potential as a reliable, non-invasive tool for predicting chemotherapy outcomes.

Despite these promising findings, this study has several limitations. Firstly, the retrospective nature of the study introduces potential selection bias [19], as the data used were not prospectively collected. Secondly, variability in imaging protocols across the four participating centers may have affected the consistency of ultrasound data acquisition, potentially leading to inter-center variability in feature extraction and model performance. Another limitation is the relatively small cohort size [54]. Although adequate for a proof-of-concept study, validation in larger and more diverse populations is needed to confirm model robustness and applicability. Ongoing efforts in cross-institutional harmonization and standardization of ultrasound imaging protocols are essential to mitigate such variability and enhance model robustness across centers. These initiatives, combined with advanced preprocessing techniques, will facilitate broader adoption of radiomics in clinical practice. Additionally, While ultrasound radiomics has shown promising results, integrating multimodal imaging data, such as MRI or CT, along with genomic and molecular data, may further improve the predictive accuracy of the models [55, 56]. Moreover, although we evaluated short-term treatment response, it remains unclear whether these outcomes correlate with long-term survival benefits—underscoring the need for studies with extended follow-up. Finally, the process of manual delineation and feature extraction is time-consuming and requires considerable expertise. Automated approaches using machine learning could streamline this process, reduce operator variability, and improve scalability in clinical practice [57, 58].

In conclusion, this study highlights the utility of ultrasound radiomics as a non-invasive tool for predicting the efficacy of neoadjuvant chemotherapy in HER2-low breast cancer. Feature selection using RFECV improved model performance, particularly for pre-treatment features, underscoring the value of baseline imaging. Although the integration of pre- and post-treatment features did not significantly improve predictive accuracy, its consistent performance indicates a potential role for longitudinal imaging in monitoring phenotypic changes over the course of therapy. Future research with larger, prospective cohorts is needed to validate these findings and explore their clinical applicability.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

NAC

Neoadjuvant chemotherapy

MRI

Magnetic Resonance Imaging

RFECV

Recursive Feature Elimination with Cross-Validation

AUC

Area Under the Receiver Operating Characteristic Curve

ORR

Overall Response Rate

SVM

Support Vector Machine

CR

Complete Response

PR

Partial Response

PD

Progressive Disease

SD

Stable Disease

ROI

Regions Of Interest

SDD

Standard Deviation Difference

CV

Coefficient Of Variation

AUC-PR

Precision-Recall AUC

TRRA

Test-Retest Reliability Analysis

PFS

Progression-free survival

OS

Overall survival

Author contributions

Qing Peng contributed to the image processing, feature extraction design, the implementation of the machine learning models, the reproducibility analysis of radiomics features and data analysis. She also assisted in drafting the manuscript.Ziyao Ji was responsible for the data collection, data analysis and contributed to the manuscript revision.Nan Xu contributed to the image processing, data analysis and contributed to the manuscript revision.Zixian Dong conducted the data collection and data analysis.Tian Zhang contributed to the statistical analysis and interpretation of the results.Mufei Ding, Le Qu, Yimo Liu, Jun Xie and Feng Jin participated in the image annotation and region of interest (ROI) delineation in collaboration with the radiologists.Bo Chen was responsible for the critical review and final approval of the manuscript.Jiangdian Song provided overall guidance and supervision throughout the research and was responsible for the critical review.Ang Zheng provided valuable input into the study design and critically reviewed the manuscript.All authors read and approved the final manuscript.

Funding

This study was funded by the National Natural Science Foundation of China (82203873) and Liaoning Provincial Social Science Planning Fund (L22CGL021).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

This retrospective multicenter study was conducted in accordance with the ethical standards of the Declaration of Helsinki and was approved by the institutional review boards of all participating hospitals. Given the retrospective nature of the study and the use of anonymized clinical and imaging data, the requirement for informed consent was waived by the respective ethics committees.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.