Category: 3. Business

Introduction

Breast cancer (BC) has become the most commonly diagnosed cancer and remains a leading cause of cancer-related mortality among women worldwide.¹ Ki-67, a nuclear protein associated with cellular proliferation, serves as an essential marker for evaluating tumor growth dynamics.² Elevated Ki-67 expression is associated with an increased risk of tumor invasion and recurrence.³ Moreover, it correlates with the pathological complete response rate to neoadjuvant therapy (NAT) in BC patients.⁴ A study by Chen et al found that higher pre-treatment Ki-67 levels were associated with a better clinical response to neoadjuvant chemotherapy in luminal BC subtypes. Specifically, a Ki-67 cutoff value of 25.5% was identified as a predictor of treatment response, indicating that Ki-67 could serve as a valuable biomarker for guiding individualized treatment strategies.⁵ Furthermore, the POETIC trial demonstrated that changes in Ki-67 levels during preoperative endocrine therapy were predictive of long-term outcomes, underscoring the critical role of Ki-67 in informing treatment strategies and optimizing patient management.⁶ Currently, Ki-67 expression is assessed via immunohistochemical (IHC) analysis, which typically requires an invasive core needle biopsy (CNB) prior to surgery. However, due to intratumoral heterogeneity and the limited sampling of CNB, discrepancies between CNB and postoperative specimens are frequently observed, with reported inconsistency rates ranging from 10% to 40%.⁷ Therefore, there is a pressing need for an accurate, comprehensive, and non-invasive method to predict preoperative Ki-67 expression, which is essential for clinical decision-making.

Ultrasound (US) is widely used for the diagnosis of BC due to its simplicity, low cost, and non-invasive nature.⁸ Compared to conventional US, ABVS enhances reproducibility through automated scanning,⁹ representing a significant technological advancement in US imaging.¹⁰ Its standardized coronal imaging offers detailed lesion information, making it particularly suitable for radiomics analysis.¹¹ Radiomics enables the extraction of high-throughput quantitative features from medical images, allowing for a comprehensive characterization of tumor phenotype.¹² It has been increasingly applied in tumor diagnosis, treatment planning, and prognosis prediction.¹³ The significant potential of radiomics has been fully demonstrated across numerous oncology applications, particularly in improving the predictive performance of ABVS in BC management.^11,14–17 This progress marks a major step toward non-invasive tumor biological profiling and further integrates medical imaging with personalized medicine.¹⁸ However, tumor heterogeneity, arising from variations in cell composition and spatial distribution, poses a challenge to accurate characterization.¹⁹ To better visualize and quantify this heterogeneity, voxel clusters with similar tumor biological characteristics can be grouped into sub-regions.²⁰ Habitat radiomics, an emerging approach, segments tumors into biologically similar sub-regions and extracts features from these areas to enhance the assessment of tumor heterogeneity. A recent study,²¹ involving multi-modal logistic regression models based on magnetic resonance imaging (MRI), US, and mammography revealed that incorporating peripheral tumor features (within 5 mm) yielded the best performance in distinguishing benign from malignant breast nodules, with an AUC of 0.905 (95% CI: 0.805–1). This highlights the increasing value of multimodal radiomics approaches. Previous US-based radiomics studies have shown promise in predicting Ki-67 expression in BC.^16,17 However, these studies were limited by small sample sizes and single-modality imaging, which may not fully capture the complex characteristics of tumors. Additionally, habitat radiomics has not been widely explored in this context. As tumor heterogeneity plays a crucial role in understanding tumor biology, habitat radiomics may enhance predictive accuracy. To address these limitations, our study integrates multimodal radiomics derived from ABVS and 2D US, along with habitat radiomics, to provide a more comprehensive and accurate prediction of Ki-67 expression in BC.

Motivated by these insights, we aim to evaluate the predictive value of radiomics and habitat radiomics features, captured from entire tumors and their sub-regions, using a machine learning (ML) model. Our goal is to establish a robust, non-invasive model for predicting Ki-67 expression in BC, thereby supporting personalized treatment strategies and improving prognostic assessment.

Materials and Methods

Study Population

This retrospective study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of the First Affiliated Hospital of Anhui Medical University (Approval No. PJ2023-07-11). Given the retrospective nature of the study, which involved the use of previously collected cases and medical records without any new clinical interventions, the requirement for informed consent was waived in accordance with relevant ethical guidelines. To ensure data reliability, strict inclusion and exclusion criteria were applied. The inclusion criteria were as follows: (1) Patients pathologically diagnosed with BC who underwent both ABVS and conventional US examinations within two weeks prior to surgery; (2) No NAT administered before surgery; (3) Availability of complete clinical and pathological information. The exclusion criteria were: (1) Incomplete clinical or pathological data; (2) History of other malignant tumors; (3) Receipt of preoperative NAT; (4) Maximum tumor diameter > 50 mm as measured by US; (5) Presence of bilateral BC. A total of 426 patients with BC met the eligibility criteria. The cases were randomly divided into a training set (n = 297) and a validation set (n = 127), following a 7:3 ratio (Figure 1).

Clinical and Histopathological Data Acquisition and Assessment

Details regarding image acquisition, evaluation procedures, and US equipment specifications are provided in Supplementary Material S1. Breast lesion classification was performed according to the fifth edition of the American College of Radiology (ACR) Breast Imaging Reporting and Data System (BI-RADS).²² The Ki-67 proliferation index was calculated based on the percentage of malignant cells showing positive nuclear staining for Ki-67. A Ki-67 score ≥ 20% was defined as high expression, while a score < 20% was considered low expression.^23,24

Segmentation of Regions of Interest (ROIs) and Generation of Habitat Sub-Regions

To segment the lesion regions on ABVS coronal images and 2D US images, the ITK-SNAP software (version 3.8, website: www.itk-snap.org) was used for tumor segmentation. Details of the tumor ROI delineation process are provided in Supplementary Material S2. Before generating habitat sub-regions, the tumors were first localized on both ABVS and 2D images. Each image was paired with a tumor mask of identical dimensions. The bounding box of the tumor region was extracted from the corresponding mask file, and this region was used to define the area for further analysis. Radiomic features were then extracted from these sub-regions. For each pixel within the tumor, a 5 × 5×5 sliding window was applied, expanding outward by two pixels in each direction. Radiomic features within this window were extracted using the pyradiomics library. To enable subsequent clustering analysis, all features were normalized to a range of 0 to 1. This standardization mitigates the influence of features with larger numerical scales and enhances the accuracy and robustness of clustering.²⁵

Although larger windows and a greater number of features may improve robustness to noise, they also significantly increase computational complexity, especially when extracting features at the pixel level. Therefore, in this study, the number of radiomic features was limited to five, specifically those derived from the Gray-Level Co-occurrence Matrix (GLCM). The GLCM effectively captures subtle textural variations in the image, reflecting microscopic irregularity and complexity. It has been widely used to reveal the potential relationship between tumor tissue structure and biological behavior, making it an important tool in the study of tumor heterogeneity^25,26 Subsequently, to further quantify image data, each pixel’s local histological features were transformed into a five-dimensional feature vector. This vector integrates multiple aspects of local feature information, such as texture, contrast, and uniformity, facilitating subsequent quantitative analysis and model construction.

A Gaussian mixture model (GMM) clustering algorithm was employed to identify tumor sub-regions composed of biologically similar pixels. Clustering was performed at the cohort level, rather than the individual patient level, to ensure consistent cluster assignment across patients and to allow the propagation of cluster centers from the training set to the test set, ensuring consistent clustering during model application. To determine the optimal number of clusters (ie, habitats), the Silhouette coefficient was used to evaluate clustering performance across a range of k values from 2 to 10. Following clustering, each cluster was assigned a unique color label to generate a cluster label map, which reflected the global distribution of internal regions within the tumor.

Radiomics Feature Extraction

Radiomics features were extracted from both tumor regions and sub-regions within the ABVS and 2D US images. Prior to feature extraction, image preprocessing was performed to ensure consistency across datasets. First, the signal intensity of the original images was normalized and standardized to a range of 0–100 to minimize intensity variations between images. Next, spatial resampling was conducted to achieve a uniform in-plane pixel resolution of 2 mm in the XY direction. All feature extraction was confined to the two-dimensional plane. Image gray levels were discretized based on a predefined bin width (eg, grouping every 5 intensity values), enabling standardized texture quantification. All processing was applied exclusively to regions defined by the specified labels in the tumor mask. A wavelet filter was used for feature enhancement prior to extraction. Radiomics features were extracted using the open-source Python package PyRadiomics (https://pyradiomics.readthedocs.io/en/latest/index.html, version 3.0.1), developed by the Computational Imaging Bioinformatics Laboratory at Harvard Medical School. Both unfiltered (from original images) and filtered features were included in the analysis. A total of 464 features were extracted and categorized into the following classes: 90 first-order features, 9 shape features, 110 Gray-Level Co-occurrence Matrix (GLCM) features, 80 gray-level size zone matrix (GLSZM) features, 80 gray-level run length matrix (GLRLM) features, 25 neighboring gray-tone difference matrix (NGTDM) features, 70 gray-level dependence matrix (GLDM) features. PyRadiomics adheres to the standards of the Imaging Biomarker Standardization Initiative (IBSI), ensuring consistent definitions and methodologies for radiomic feature extraction.

Radiomics Feature Selection

The selection of predictive features associated with Ki-67 expression was performed through a multi-step process. This included screening radiomics features, sub-regional radiomics features, and clinical US features. The detailed feature selection workflow is presented in Supplementary Material S3.

Construction and Validation of ML Models

After feature selection and fusion, five predictive models were developed, namely the clinical model, the radiomics model (Rad _{ABVS + 2D}), the habitat radiomics model (Hab _{ABVS + 2D}), the combined radiomics model (Rad-Hab _{ABVS + 2D}), and the Clinical–Radiomics–Habitat _ABVS+2D Combined Model(CM _{Clinical + Rad-Hab}). The entire model construction workflow is shown in Figure 2. All models were established using radiomics features derived from ABVS and 2D images. Four ML classifiers were employed: logistic regression (LR), ExtraTree (ET), EXtreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM). To reduce overfitting, 5-fold cross-validation was applied within the training cohort to optimize hyperparameters for each classifier. Model performance was evaluated by plotting receiver operating characteristic (ROC) curves and calculating the area under the curve (AUC). The DeLong test was applied to statistically compare ROC performance across different models, while the Hosmer-Lemeshow test assessed the models’ goodness-of-fit. Clinical utility was further evaluated using the DCA to estimate the net benefit of each model in guiding clinical decision-making.

Figure 2 Workflow of radiomics analysis. This figure illustrates the segmentation, feature extraction, and feature selection process for ABVS and 2D images in breast cancer. Abbreviations: ABVS, automated breast volume scanner; ICC, intraclass correlation coefficient; LASSO, least absolute shrinkage and selection operator; PCC, Pearson correlation coefficient; RadABVS+2D, Radiomics ABVS and 2D model; HadABVS+2D, Habitat radiomics ABVS and 2D model; Rad-HabABVS+2D, Radiomics and Habitat radiomics ABVS and 2D model; CMClinical + Rad-Hab, Clinical–Radiomics–Habitat ABVS+2D Combined Model.

Statistical Analysis

All statistical analyses and data visualizations were performed using R software (version 4.4.2) and JD_DCPM (V6.03, Jingding Medical Technology Co., Ltd.) and Python (version 3.8; https://www.python.org). Continuous variables were presented as mean ± standard deviation, while categorical variables were expressed as counts (n) and percentages (%). For quantitative data following a normal distribution, Student’s t-test was used. Levene’s test was employed to assess the homogeneity of variance. The Kruskal–Wallis test was used for non-normally distributed data. The Chi-square test was applied to compare categorical data. The DeLong test was used to compare the ROC performance among different models. All statistical tests were two-sided, and statistical significance was set at P < 0.05.

Results

Comparison of Baseline Data

A total of 426 eligible BC patients were included in this study, with 297 patients assigned to the training set and 129 to the validation set. A summary of baseline clinical and US characteristics is presented in Table 1. Among these variables, multivariate logistic regression analysis identified T-stage and US-ALNs as independent predictors. These two factors were therefore incorporated into the clinical model (Table 2). Using LR, the clinical model achieved an AUC of 0.720 (95% CI: 0.662–0.775) in the training set and 0.648 (95% CI: 0.557–0.734) in the validation set.

Table 1 Baseline Clinical Ultrasound Characteristics in the Training and Validation Sets

Table 2 Univariate and Multivariate Logistic Regression Analyses of Clinical Ultrasound Characteristics in the Training Set

Screening of Radiomics Features

Rad _{ABVS + 2D} Feature Selection

In the training set, the tumor ROI were delineated on both ABVS and 2D US images, and a total of 464 radiomics features were extracted from each image. After standardization, features with an intraclass correlation coefficient (ICC) > 0.75 were retained for subsequent analysis, resulting in 898 features. Subsequently, univariate t-tests and Pearson correlation coefficient (PCC) analyses were conducted to evaluate the relationships among these features. Finally, the least absolute shrinkage and selection operator (LASSO) algorithm was used to select the most predictive features based on the optimal λ value. The Rad _{ABVS + 2D} model identified 15 radiomics features, 9 from ABVS images and 6 from 2D US images, that were significantly correlated with Ki-67 expression (λ=0.037, Figure S1).

Hab _{ABVS + 2D} Feature Selection

In this study, the Silhouette Coefficient was used to evaluate clustering performance and determine the optimal number of clusters (Figure 3A). The analysis revealed that when the number of clusters was set to 3, the silhouette coefficient reached its highest value, indicating the best clustering performance. Accordingly, the tumor ROI was divided into three habitat sub-regions for subsequent feature extraction and model construction. For each sub-region, 464 radiomics features were extracted, resulting in a total of 2694 features (898 features × 3 sub-regions), following intraclass correlation coefficient (ICC) filtering. Feature selection was then conducted, and the final Hab _{ABVS + 2D} model identified 13 ABVS and 12 2D radiomics features that were significantly correlated with Ki-67 expression (λ=0.018, Figure S2). The habitat feature maps and corresponding sub-region segmentations are shown in Figure 3B and C.

Figure 3 Determination of optimal clusters and visualization of habitat clusters with corresponding feature maps. (A) Silhouette coefficient plot was used to determine the optimal number of clusters (k), indicating that k=3 is optimal. (B) Distribution of habitats under different clustering numbers, with different colors representing different clusters. (C) Feature maps for the following parameters: original_glcm_DifferenceVariance, original_glcm_Idm, original_glcm_InverseVariance, original_glcm_JointAverage, original_glcm_JointEnergy, original_glcm_MaximumProbability.

Rad-Hab _{ABVS + 2D} Feature Selection

According to the methodology described above, the Rad-Hab _{ABVS + 2D} model initially included a total of 3592 features (898 + 898×3). After feature selection and dimensionality reduction, the final model identified 16 ABVS features and 8 2D radiomics features that were significantly correlated with Ki-67 expression (λ=0.012, Figure S3).

Construction and Validation of Each Models

The optimal model for the traditional Rad _{ABVS + 2D} configuration was LR, as detailed in Table S1 and Figure S4A and B. In the training set, the model achieved an AUC of 0.755 (95% CI: 0.688–0.813), with accuracy, sensitivity, specificity, and F1-score of 0.670, 0.824, 0.602, and 0.605, respectively. In the internal validation set, the AUC was 0.603 (95% CI: 0.515–0.690), with corresponding values of 0.682 for accuracy, 0.442 for sensitivity, 0.802 for specificity, and 0.481 for F1-score.

Figure 4 Comparison of the performance of different radiomics models in the training and validation sets. This figure shows the distribution of radiomics model outputs between the high- and low-expression Ki-67 groups. (A and B) RadABVS+2D; (C and D) HabABVS+2D; (E and F) Rad-HabABVS+2D; (G and H) CMClinical + Rad-Hab. Abbrteviations: RadABVS+2D, Radiomics ABVS and 2D model; HadABVS+2D, Habitat radiomics ABVS and 2D model; Rad-HabABVS+2D, Radiomics and Habitat radiomics ABVS and 2D model; CMClinical + Rad-Hab, Clinical–Radiomics–Habitat ABVS+2D Combined Model; ABVS, automated breast volume scanner.

For the Hab _{ABVS + 2D} model, the best-performing algorithm was ExtraTree, as detailed in Table S2 and Figure S4C and D). In the training set, the model achieved an AUC of 0.779 (95% CI: 0.712 −0.825), with accuracy, sensitivity, specificity, and F1-score of 0.781, 0.505, 0.903, and 0.586. In the validation set, the AUC was 0.664 (95% CI: 0.579–0.755), and accuracy, sensitivity, specificity, and F1-score were 0.605, 0.721, 0.547, and 0.549, respectively.

The Rad-Hab _{ABVS + 2D} model performed best when using the XGBoost, as detailed in Table S3 and Figure S4E and F. In the training set, the model achieved an AUC of 0.935 (95% CI: 0.910–0.962), with accuracy, sensitivity, specificity, and F1-score of 0.869, 0.868, 0.869, and 0.802, respectively. In the validation set, the AUC was 0.850 (95% CI: 0.789–0.918), with values of 0.806 for accuracy, 0.744 for sensitivity, 0.837 for specificity, and 0.719 for F1-score.

After combining Rad-Hab_{ABVS + 2D} features with clinical features, the ML model achieving the best performance was based on LightGBM, as detailed in Table S4 and Figure S4G and H. This resulting model was designated CM _{Clinical + Rad-Hab}. In the training set, the model achieved an AUC of 0.951 (95% CI: 0.928–0.973), with accuracy, sensitivity, specificity, and F1-score of 0.886, 0.890, 0.883, and 0.827, respectively. In the validation set, the AUC was 0.884 (95% CI: 0.831–0.949), with corresponding values of 0.783 for accuracy, 0.860 for sensitivity, 0.744 for specificity, and 0.725 for F1-score. Across the radiomics-based models, statistically significant differences in Ki-67 expression were observed between the high- and low-expression groups, with the exception of the Rad _{ABVS + 2D} in the validation set (P <0.01, Figure 4). The differences in radiomics features from specific habitat sub-regions included in the CM _{Clinical + Rad-Hab} model are shown in Figure 5.

Figure 5 Box plot showing differences in ABVS and 2D image habitat radiomics features between the low- and high-expression groups in the CMClinical+Rad_Hab model. The P value indicates statistical significance and is displayed in each feature chart. Black dots represent individual data points; the horizontal line within each box indicates the median (50th percentile), and the upper and lower edges represent the 25th and 75th percentiles, respectively. Abbreviation: ABVS, automated breast volume scanner.

Comparison of Model Performance

In the training set, the DeLong test results indicated that the AUC differences between the CM _{Clinical + Rad-Hab} model and the Clinical, Rad _{ABVS + 2D}, Hab _{ABVS + 2D}, and Rad-Hab _{ABVS + 2D} models were all statistically significant (Z = 7.979, P < 0.001; Z = 7.162, P < 0.001; Z = 6.017, P < 0.001; Z = 2.669, P = 0.007). Similar results were observed in the validation set (Z = 4.829, P < 0.001; Z = 5.665, P < 0.001; Z = 4.885, P < 0.001; Z = 2.662, P = 0.009). These findings indicate that the CM _{Clinical + Rad-Hab} model significantly outperformed the other models in predicting Ki-67 expression, as presented in Figure 6A and B, Table 3. The calibration curves of the CM _{Clinical + Rad-Hab} model exhibited strong agreement between expected and observed outcomes in both the high and low Ki-67 expression groups, surpassing the calibration performance of the other models, as illustrated in Figure 6C and D. The Hosmer-Lemeshow test further confirmed good model calibration for the CM _{Clinical + Rad-Hab} model, with P = 0.645 and 0.587 for the two sets.

Table 3 Comparison of Radiomics and Sub-Region Features Across Different Models Using ABVS and 2D Imaging

Figure 6 ROC curves (A and B), calibration curves (C and D), and DCA curves (E and F) for different models in the training and validation sets. Abbreviations: DCA, Decision Curve Analysis; ROC, Receiver Operating Characteristic; ABVS, automated breast volume scanner; RadABVS+2D, Radiomics ABVS and 2D model; HadABVS+2D, Habitat radiomics ABVS and 2D model; Rad-HabABVS+2D, Radiomics and Habitat radiomics ABVS and 2D model; CM Clinical + Rad-Hab, Clinical–Radiomics–Habitat ABVS+2D Combined Model.

The DCA demonstrated that CM _{Clinical + Rad-Hab} achieved superior net clinical benefit compared to all- or no-treatment strategies, as shown in Figure 6E and F. Furthermore, incorporating the Rad-Hab _{ABVS + 2D} model with clinical risk factors (T-stage, US-ALNs) significantly improved the predictive performance of the CM _{Clinical + Rad-Hab}. This improvement was confirmed by significant increases in both the net reclassification improvement (NRI) and integrated discrimination improvement (IDI) indicators in the training and validation sets (Table 4). NRI and IDI values for the other models can be found in Table S5.

Table 4 Evaluation of CM Clinical + Rad-Hab and Clinical Models Using NRI and IDI

Discussion

In this study, we proposed a new approach that integrates ABVS- and 2D-based lesion imaging with radiomics and habitat radiomics to predict the expression of Ki-67 in BC. Our results demonstrated that the Rad-Hab _ABVS+2D model could accurately predict Ki-67 expression, and that incorporating clinical factors into this model (CM _{Clinical + Rad-Hab}) further enhanced predictive performance. This approach provides a new strategy for constructing Ki-67 prediction models in BC. More importantly, by leveraging ML to integrate radiomics features, our approach addresses tumor heterogeneity and enables complex nonlinear feature interpretation, thereby improving predictive accuracy.

A previous study utilized six ML algorithms to integrate US radiomics with postoperative pathological features to predict Ki-67 expression in breast malignancies, reporting that the LR achieved the best average predictive performance (training AUC: 0.793, validation AUC: 0.798).¹⁷ Another similar study also attempted to predict Ki-67 expression, but the predictive performance was only moderate.²⁷ In our study, the Rad-Hab_{ABVS + 2D} model achieved AUC values of 0.935 in the training cohort and 0.850 in the validation cohort, demonstrating strong predictive capability. Furthermore, by integrating US features (US-ALNs and T-stage), the LightGBM-based CM_{Clinical + Rad-Hab} model achieved even higher performance (training AUC: 0.951, validation AUC: 0.884). The calibration curve demonstrated excellent predictive accuracy, while DCA showed good clinical benefits. These findings highlight the potential of ML-based multimodal radiomics as a non-invasive tool for personalized clinical diagnosis and treatment planning.

Radiomics has emerged as a specialized field within medical imaging, enabling the extraction of high-throughput quantitative features from medical images. This approach captures the intrinsic characteristics of breast tumors, and to some extent, provides additional insights into Ki-67 expression in BC.²⁸ By applying various ML or deep learning algorithms to different imaging modalities,^2,11,27 key information about Ki-67 expression in BC can be obtained, supplementing conventional pathological assessments. In our study, dimensionality reduction of the Rad-Hab_{ABVS + 2D} model yielded 24 key radiomics features related to the Ki-67 status. These features, combined with clinical variables, were evaluated using four ML algorithms (LR, ET, XGBoost, and LightGBM). The results showed that LightGBM demonstrated the best predictive performance. As a gradient boosting decision tree algorithm, LightGBM enhances performance by increasing the number of boosting trees and offers efficiency and flexibility in modeling nonlinear relationships. Its ability to handle large datasets with high-dimensional features makes it particularly effective for predictive modeling.²⁹ Numerous studies have confirmed the robustness and reliability of LightGBM in classification and regression tasks,^29–31 further proving its robustness and reliability as a powerful classification tool.

The results of this study indicated that, within our developed CM _{Clinical + Rad-Hab} model, the Rad-Hab _{ABVS + 2D} component was the most influential predictive factor. Tumors represent complex ecosystems, and intratumoral heterogeneity is often distributed unevenly throughout the lesion.³² However, regional heterogeneity within the tumor is frequently overlooked. Recent studies have shown that voxels at different spatial locations within an image may share similar imaging features, and these sub-regions may exhibit comparable biological characteristics.³³ Therefore, to better study and quantify such regional heterogeneity, habitat radiomics divides tumors into sub-regions consisting of voxel clusters with similar characteristics for unsupervised analysis.³⁴ Recently, this approach has achieved promising results in the assessment of parotid gland tumors, liver cancer, and BC, among others.^3,32,35 In our CM_{Clinical+Rad-Hab} model, it is noteworthy that most of the 24 selected radiomics features were derived from wavelet features (16 features). These wavelet features effectively capture heterogeneity at multiple spatial scales (Figure S3),³⁶ patterns that are typically imperceptible through visual inspection but can be extracted using radiomics and mathematically correlated with the Ki-67 status. Although the Rad_Hab_ABVS+2D model alone demonstrated strong predictive performance, incorporating clinical parameters (T-stage, OR = 3.078; US-ALNs, OR = 4.759) significantly improved the overall performance of the CM_{Clinical+ABVS+2D} model (training set: Z = 2.669, P = 0.007; validation set: Z = 2.662, P = 0.009). This observation result is consistent with the studies,^9,11,37 emphasizing that only by fully leveraging multi-dimensional data can the radiomics features reach their full potential in predicting Ki-67 expression (Table 3 and Figure 6).

Although previous studies have used either ABVS or 2D single-modality images to predict Ki-67 expression, to our knowledge, this study is the first to integrate radiomics features from both ABVS and 2D US images, including sub-regional (habitat) features, for Ki-67 prediction. However, several limitations should be noted. First, as a single-center retrospective study, potential selection bias may limit the generalizability of the findings; future multicenter, prospective studies are needed to address this issue. Second, while this study is the first to employ multimodal sub-region radiomics analysis for Ki-67 prediction, model interpretability remains challenging due to the complexity of habitat radiomics, and incorporating explainable AI techniques could improve clinical interpretability. Third, our study was restricted to ABVS and 2D US images, excluding other modalities such as mammography and MRI, which could provide complementary diagnostic information; integrating these modalities in future studies may further enhance model performance. Lastly, radiomics features were extracted only from ABVS coronal images and the maximum 2D tumor section; extending analyses to 3D imaging or multiple planes may provide a more comprehensive characterization of tumor heterogeneity. In the future, we plan to explore correlations between radiomics or deep learning features and biomarkers such as Ki-67 and HER-2 expression, as well as treatment responses to NAT, using multimodal imaging, to provide deeper insights for the precise treatment of BC.

Conclusion

In conclusion, this study demonstrates that integrating ML with ABVS and 2D US tumor- and sub-regional-based radiomics features can effectively predict Ki-67 expression in BC. The developed CM_{Clinical + Rad-Hab} model, which combines US indicators with radiomics features, achieves excellent classification performance and shows substantial clinical value. This approach holds significant potential for improving preoperative diagnostic accuracy and facilitating therapeutic efficacy assessment of BC biomarkers.

Ethics Statement

This retrospective study was approved by the Ethics Committee of the First Affiliated Hospital of Anhui Medical University, with a waiver of informed consent. All research data were de-identified and processed in strict accordance with relevant privacy protection regulations to ensure the confidentiality of participant information.

Acknowledgments

We sincerely acknowledge all the staff involved in implementing the intervention and assessing the research components. We also acknowledge the PixelMed AI platform and its developers for their valuable assistance with the code used in this revised manuscript.

Funding

This study was supported by Anhui Provincial Natural Science Foundation (Grant number: 2308085MH278), Health Research Program of Anhui (Grant number: AHWJ2023A10017), Anhui Provincial Health Commission Scientific Research Project (Grant number: AHWJ2024Aa30096), and Scientific Research Foundation for High-level Talents of First Affiliated Hospital of Wannan Medical College (Grant number: YR202436).

Disclosure

The author declares no potential conflicts of interest with respect to the research, authorship, and publication of this article.

References

1. Sung H, Ferlay J, Siegel RL, et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2021;71(3):209–249. doi:10.3322/caac.21660

2. Li HE, Cheng C. Development and assessment of a predictive model for Ki-67 expression using ultrasound indicators and non-morphological magnetic resonance imaging parameters before breast cancer therapy. Ultrason Imaging. 2024;46(6):332–341. doi:10.1177/01617346241271107

3. Yerushalmi R, Woods R, Ravdin PM, Hayes MM, Gelmon KA. Ki67 in breast cancer: prognostic and predictive potential. Lancet Oncol. 2010;11(2):174–183. doi:10.1016/S1470-2045(09)70262-1

4. Mukai H, Yamaguchi T, Takahashi M, et al. Ki-67 response-guided preoperative chemotherapy for HER2-positive breast cancer: results of a randomised Phase 2 study. Br J Cancer. 2020;122(12):1747–1753. doi:10.1038/s41416-020-0815-9

5. Chen R, Ye Y, Yang C, et al. Assessment of the predictive role of pretreatment Ki-67 and Ki-67 changes in breast cancer patients receiving neoadjuvant chemotherapy according to the molecular classification: a retrospective study of 1010 patients. Breast Cancer Res Treat. 2018;170(1):35–43. doi:10.1007/s10549-018-4730-1

6. Smith I, Robertson J, Kilburn L, et al. Long-term outcome and prognostic value of Ki67 after perioperative endocrine therapy in postmenopausal women with hormone-sensitive early breast cancer (POETIC): an open-label, multicentre, parallel-group, randomised, Phase 3 trial. Lancet Oncol. 2020;21(11):1443–1454. doi:10.1016/S1470-2045(20)30458-7

7. Rossi C, Fraticelli S, Fanizza M, et al. Concordance of immunohistochemistry for predictive and prognostic factors in breast cancer between biopsy and surgical excision: a single-centre experience and review of the literature. Breast Cancer Res Treat. 2023;198(3):573–582. doi:10.1007/s10549-023-06872-9

8. Fang J, Zhao W, Li Q, Zhang B, Pu C, Wang H. Correlation analysis of conventional ultrasound characteristics and strain elastography with Ki-67 status in breast cancer. Ultrasound Med Biol. 2020;46(11):2972–2978. doi:10.1016/j.ultrasmedbio.2020.06.024

9. Li F, Zhu TW, Lin M, et al. Enhancing Ki-67 prediction in breast cancer: integrating intratumoral and peritumoral radiomics from automated breast ultrasound via machine learning. Acad Radiol. 2024;31(7):2663–2673. doi:10.1016/j.acra.2023.12.036

10. Vourtsis A. Three-dimensional automated breast ultrasound: technical aspects and first results. Diagn Interv Imaging. 2019;100(10):579–592. doi:10.1016/j.diii.2019.03.012

11. Wu Y, Ma Q, Fan L, Wu S, Wang J. An automated breast volume scanner-based intra- and peritumoral radiomics nomogram for the preoperative prediction of expression of Ki-67 in breast malignancy. Acad Radiol. 2024;31(1):93–103. doi:10.1016/j.acra.2023.07.004

12. Lambin P, Rios-Velazquez E, Leijenaar R, et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer. 2012;48(4):441–446. doi:10.1016/j.ejca.2011.11.036

13. Mayerhoefer ME, Materka A, Langs G, et al. Introduction to radiomics. J Nucl Med. 2020;61(4):488–495. doi:10.2967/jnumed.118.222893

14. Jiang W, Deng X, Zhu T, Fang J, Li J. ABVS-based radiomics for early predicting the efficacy of neoadjuvant chemotherapy in patients with breast cancers. Breast Cancer. 2023;15:625–636. doi:10.2147/BCTT.S418376

15. Ma Q, Shen C, Gao Y, et al. Radiomics analysis of breast lesions in combination with coronal plane of ABVS and strain elastography. Breast Cancer. 2023;15:381–390. doi:10.2147/BCTT.S410356

16. Wang SJ, Liu HQ, Yang T, et al. Automated Breast Volume Scanner (ABVS)-based radiomic nomogram: a potential tool for reducing unnecessary biopsies of BI-RADS 4 lesions. Diagnostics. 2022;12(1). doi:10.3390/diagnostics12010172

17. Wu J, Fang Q, Yao J, et al. Integration of ultrasound radiomics features and clinical factors: a nomogram model for identifying the Ki-67 status in patients with breast carcinoma. Front Oncol. 2022;12:979358. doi:10.3389/fonc.2022.979358

18. Chen Y, Xie Y, Li B, et al. Automated Breast Ultrasound (ABUS)-based radiomics nomogram: an individualized tool for predicting axillary lymph node tumor burden in patients with early breast cancer. BMC Cancer. 2023;23(1):340. doi:10.1186/s12885-023-10743-3

19. Li J, Qiu Z, Zhang C, et al. ITHscore: comprehensive quantification of intra-tumor heterogeneity in NSCLC by multi-scale radiomic features. Eur Radiol. 2023;33(2):893–903. doi:10.1007/s00330-022-09055-0

20. Jardim-Perassi BV, Huang S, Dominguez-Viqueira W, et al. Multiparametric MRI and coregistered histology identify tumor habitats in breast cancer mouse models. Cancer Res. 2019;79(15):3952–3964. doi:10.1158/0008-5472.CAN-19-0213

21. Wu J, Li Y, Gong W, Li Q, Han X, Zhang T. Multi-modality radiomics diagnosis of breast cancer based on MRI, ultrasound and mammography. BMC Med Imaging. 2025;25(1):265. doi:10.1186/s12880-025-01767-1

22. D’Orsi C, Morris E, Mendelson E. ACR BI-RADS® Atlas, Breast Imaging Reporting and Data System. 2013:121–140.

23. Jiang M, Zhang D, Tang SC, et al. Deep learning with convolutional neural network in the assessment of breast cancer molecular subtypes based on US images: a multicenter retrospective study. Eur Radiol. 2021;31(6):3673–3682. doi:10.1007/s00330-020-07544-8

24. de Gregorio A, Friedl TWP, Hering E, et al. Ki67 as proliferative marker in patients with early breast cancer and its association with clinicopathological factors. Oncology. 2021;99(12):780–789. doi:10.1159/000517490

25. Zhang Y, Ma H, Lei P, Li Z, Yan Z, Wang X. Prediction of early postoperative recurrence of hepatocellular carcinoma by habitat analysis based on different sequence of contrast-enhanced CT. Front Oncol. 2024;14:1522501. doi:10.3389/fonc.2024.1522501

26. Zhang J, Wang X, Zhang L, et al. Radiomics predict postoperative survival of patients with primary liver cancer with different pathological types. Ann Transl Med. 2020;8(13):820. doi:10.21037/atm-19-4668

27. Zhu Y, Dou Y, Qin L, Wang H, Wen Z. Prediction of Ki-67 of invasive ductal breast cancer based on ultrasound radiomics nomogram. J Ultrasound Med. 2023;42(3):649–664. doi:10.1002/jum.16061

28. Fan M, Liu Z, Xu M, et al. Generative adversarial network-based super-resolution of diffusion-weighted imaging: application to tumour radiomics in breast cancer. NMR Biomed. 2020;33(8):e4345. doi:10.1002/nbm.4345

29. Basha SM, Rajput DS, Vandhan V. Impact of gradient ascent and boosting algorithm in classification. Int J Intell Eng Syst. 2018;11(1):41–49. doi:10.22266/ijies2018.0228.05

30. Zhang J, Mucs D, Norinder U, Svensson F. LightGBM: an effective and scalable algorithm for prediction of chemical toxicity-application to the Tox21 and mutagenicity data sets. J Chem Inf Model. 2019;59(10):4150–4158. doi:10.1021/acs.jcim.9b00633

31. Yanagawa R, Iwadoh K, Akabane M, et al. LightGBM outperforms other machine learning techniques in predicting graft failure after liver transplantation: creation of a predictive model through large-scale analysis. Clin Transplant. 2024;38(4):e15316. doi:10.1111/ctr.15316

32. Ma Q, Wang J, Tu Z, et al. Prediction model of axillary lymph node status using an automated breast volume ultrasound radiomics nomogram in early breast cancer with negative axillary ultrasound. Front Immunol. 2025;16:1460673. doi:10.3389/fimmu.2025.1460673

33. O’Connor JP, Rose CJ, Waterton JC, Carano RA, Parker GJ, Jackson A. Imaging intratumor heterogeneity: role in therapy response, resistance, and clinical outcome. Clin Cancer Res. 2015;21(2):249–257. doi:10.1158/1078-0432.CCR-14-0990

34. Kim J, Ryu SY, Lee SH, Lee HY, Park H. Clustering approach to identify intratumour heterogeneity combining FDG PET and diffusion-weighted MRI in lung adenocarcinoma. Eur Radiol. 2019;29(1):468–475. doi:10.1007/s00330-018-5590-0

35. Xiao Y, Huang P, Zhang Y, et al. Component prediction in combined hepatocellular carcinoma-cholangiocarcinoma: habitat imaging and its biologic underpinnings. Abdom Radiol. 2024;49(4):1063–1073. doi:10.1007/s00261-023-04174-8

36. Xie C, Yang P, Zhang X, et al. Sub-region based radiomics analysis for survival prediction in oesophageal tumours treated by definitive concurrent chemoradiotherapy. EBioMedicine. 2019;44:289–297. doi:10.1016/j.ebiom.2019.05.023

37. Liu J, Wang X, Hu M, et al. Development of an ultrasound-based radiomics nomogram to preoperatively predict Ki-67 expression level in patients with breast cancer. Front Oncol. 2022;12:963925. doi:10.3389/fonc.2022.963925

People living in a village were left without water for five hours after a pipe burst.

The main failed at about 06:40 BST, impacting about 150 households in The Street, Pump Lane, Mill Hill and Howe Green Road, in Purleigh, near Chelmsford and Maldon, Essex.

Essex and Suffolk Water installed a temporary overland pipe “to get water flowing again” while repairs were carried out, before “most supplies” were restored around midday.

A spokesperson said: “It may take a little time for the pressure to build up in our network. If your water is discoloured, please run your cold kitchen tap for up to 40 mins or until clear. Thanks for your patience.”

Prior to carrying out work in the village, the water firm’s engineers informed residents about the issue and installed a road closure at Lodge Lane.

At the time they apologised for “any inconvenience” caused and assured homeowners they were “doing everything” they could “to get things back to normal” as quickly as possible.

Unlock the Editor’s Digest for free

Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.

MKS Instruments, a supplier to Taiwan Semiconductor Manufacturing Company, is selling a $1bn speciality chemicals division in a bid to focus its operations on supplying chipmakers, according to people familiar with the matter.

The Massachusetts-based technology group, which specialises in advanced manufacturing equipment crucial to the semiconductor supply chain, is working with advisers to divest the division, which it acquired as part of its $5.1bn takeover of Atotech in 2021.

The unit, which generates about $100mn in adjusted earnings a year, focuses on supplying technology used to apply coatings and finishes to automobiles and industrial equipment. MKS will hold on to the remainder of the division that provides equipment used to produce semiconductors and circuit boards.

MKS is trying to sell investors on how the boom in artificial intelligence and other technologies will drive a surge in demand for its manufacturing instruments, which it says are essential to the next wave of innovation. The company supplies semiconductor giants such as TSMC, Applied Materials and Lam Research.

Both its semiconductor and electronics divisions delivered revenue growth above analyst projections in the most recent quarter.

John Lee, MKS chief executive, said on an earnings call in August that the double-digit growth in its electronics and packaging arm was “validating MKS’s position in a market where complex electronics applications like AI are driving growth”.

MKS declined to comment. Shares in MKS stood at $121.3 each at Friday’s close, up 14.4 per cent this year, giving it a market value of $8.3bn.

A wide array of strategic buyers and private equity groups had been approached as part of the auction, the people said. The sale process was at an advanced stage, but there were no guarantees that a deal will be clinched, they added.

Private equity groups have jumped on similar carve-outs in recent weeks. This month, Carlyle struck a €7.7bn deal to take control of BASF’s coating unit, as part of which the German chemicals giant will retain a minority stake.

Before MKS struck a deal to buy Atotech, Carlyle owned 79 per cent of the outstanding shares.

Intel’s future has never seemed so uncertain. But most of the company’s roller-coaster ride of a year has been a lead-up to its next-gen CPU launch, announced this week. The chips will be known as Intel Core Ultra Series 3, codenamed Panther Lake, and they’re being manufactured in its new Arizona-based fabrication plant.

Intel claims the first configurations will ship before the end of the year and then more broadly starting in January 2026. We don’t have a complete lineup yet, but Panther Lake will include up to 16-core CPUs with a “more than 50 percent faster CPU” performance over the previous generation. Intel claims that the new integrated GPU with have up to 12 GPU cores that are also 50 percent faster than the prior generation, boosted by a new architecture.

Intel is fighting back against the stiff competition. Qualcomm dramatically entered the Windows laptop race in 2024 with its Arm-based, highly-efficient Snapdragon X chips, doubling the battery life of current Intel-powered laptops in some cases. While Intel was able to respond to the battery-life competition with its Core Ultra Series 2 V-series chips in late 2024, performance took a hit on these laptops, and the efficiency only applied to flagship, thin, and light laptops. Budget-level and high-performance laptops used a different architecture and therefore didn’t get that same bump in efficiency.

That made shopping for a laptop in 2025 even more head-scratching than normal. These next chips will attempt to fix this problem, with the company promising “Lunar Lake–level power efficiency” and “Arrow Lake–class performance.” Intel really needs to achieve that promise, because with Qualcomm’s Snapdragon X2 Elite having just been previewed and the Apple M5 on the way, the stakes keep rising. —Luke Larsen

Apple’s Next Hardware Launch Is Coming Soon

If you’re thinking, didn’t Apple just have an event? Yes, the company debuted new iPhones, Apple Watches, and AirPods just last month. But rumors are heating up that the company will announce more products this month, focused on iPads and MacBooks. That’s not unusual, as the company has held October events for the past few years, usually for the tablet and Mac lineups. It’s unclear whether this will be an actual event or a silent launch via press release. The company has done both in the past.

So what can you expect? The marquee announcement will revolve around the anticipated M5 chipset, which may debut inside a new MacBook Pro and the iPad Pro. The flagship tablet likely won’t look or feel too different from the prior M4 version. MacBooks are a little more up in the air on launch timing; it could be at this event or early in 2026. If they are announced, it’ll be a new 14- and 16-inch MacBook Pro with an M5, M5 Pro, and M5 Max chip. Apple has also reportedly been gearing up for a budget MacBook launch powered by an iPhone processor, but this may arrive early in 2026 instead.

Other hardware that may debut at this October event includes a new Vision Pro powered by an M4 or M5 chip with a comfier head strap, though it’s otherwise the same as the original headset. There may be a new Apple TV with a faster chipset, the new version of Siri (though this won’t come until 2026), and Wi-Fi 7 support. And we may finally see a second-gen AirTag, with a longer range.

The PlayStation 6 May Arrive in a ‘Few Years’

A year after tech firm OpenAI roiled Hollywood with the release of its Sora AI video tool, Chief Executive Sam Altman was back — with a potentially groundbreaking update.

Unlike the generic images Sora could initially create, the new program allows users to upload videos of real people and put them into AI-generated environments, complete with sound effects and dialogue.

In one video, a synthetic Michael Jackson takes a selfie video with an image of “Breaking Bad” star Bryan Cranston. In another, a likeness of SpongeBob SquarePants speaks out from behind the White House’s Oval Office desk.

“Excited to launch Sora 2!” Altman wrote on social media platform X on Sept. 30. “Video models have come a long way; this is a tremendous research achievement.”

But the enthusiasm wasn’t shared in Hollywood, where the new AI tools have created a swift backlash. At the core of the dispute is who controls the copyrighted images and likenesses of actors and licensed characters — and how much they should be compensated for their use in AI models.

The Motion Picture Assn. trade group didn’t mince words.

“OpenAI needs to take immediate and decisive action to address this issue,” Chairman Charles Rivkin said in a statement Monday. “Well-established copyright law safeguards the rights of creators and applies here.”

By the end of the week, multiple agencies and unions, including SAG-AFTRA, chimed in with similar statements, marking a rare moment of consensus in Hollywood and putting OpenAI on the defensive.

“We’re engaging directly with studios and rightsholders, listening to feedback, and learning from how people are using Sora 2,” Varun Shetty, OpenAI’s vice president of media partnerships, said in a statement. “Many are creating original videos and excited about interacting with their favorite characters, which we see as an opportunity for rightsholders to connect with fans and share in that creativity.”

For now, the skirmish between well-capitalized OpenAI and the major Hollywood studios and agencies appears to be only just the beginning of a bruising legal fight that could shape the future of AI use in the entertainment business.

“The question is less about if the studios will try to assert themselves, but when and how,” said Anthony Glukhov, senior associate at law firm Ramo, of the clash between Silicon Valley and Hollywood over AI. “They can posture all they want; but at the end of the day, there’s going to be two titans battling it out.”

Before it became the focus of ire in the creative community, OpenAI quietly tried to make inroads into the film and TV business.

The company’s executives went on a charm offensive last year. They reached out to key players in the entertainment industry — including Walt Disney Co. — about potential areas for collaboration and trying to assuage concerns about its technology.

This year, the San Francisco-based AI startup took a more assertive approach.

Before unveiling Sora 2 to the general public, OpenAI executives had conversations with some studios and talent agencies, putting them on notice that they need to explicitly declare which pieces of intellectual property — including licensed characters — were being opted-out of having their likeness depicted on the AI platform, according to two sources familiar with the matter who were not authorized to comment. Actors would be included in Sora 2 unless they opted out, the people said.

OpenAI disputes the claim and says that it was always the company’s intent to give actors and other public figures control over how their likeness is used.

The response was immediate.

Beverly Hills talent agency WME, which represents stars such as Michael B. Jordan and Oprah Winfrey, told OpenAI its actions were unacceptable, and that all of its clients would be opting out.

Creative Artists Agency and United Talent Agency also argued that their clients had the right to control and be compensated for their likenesses.

Studios, including Warner Bros., echoed the point.

“Decades of enforceable copyright law establishes that content owners do not need to ‘opt out’ to prevent infringing uses of their protected IP,” Warner Bros. Discovery said in a statement. “As technology progresses and platforms advance, the traditional principles of copyright protection do not change.”

Unions, including SAG-AFTRA — whose members were already alarmed over the recent appearance of a fake, AI-generated composite named Tilly Norwood — also expressed alarm.

“OpenAI’s decision to honor copyright only through an ‘opt-out’ model threatens the economic foundation of our entire industry and underscores the stakes in the litigation currently working through the courts,” newly elected President Sean Astin and National Executive Director Duncan Crabtree-Ireland said in a statement.

The dispute underscores a clash of two very different cultures. On one side is the brash, Silicon Valley “move fast and break things” ethos, where asking for forgiveness is seen as preferable to asking for permission. On the other is Hollywood’s eternal wariness over the effect of new technology, and its desire to retain control over increasingly valuable intellectual property rights.

“The difficulty, as we’ve seen, is balancing the capabilities with the prior rights owned by other people,” said Rob Rosenberg, a partner with law firm Moses and Singer LLP and a former Showtime Networks general counsel. “That’s what was driving the entire entertainment industry bonkers.”

Amid the outcry, Sam Altman posted on his blog days after the Sora 2 launch that the company would be giving more granular controls to rights holders and is working on a way to compensate them for video generation.

OpenAI said it has guardrails to block the generation of well-known characters and a team of reviewers who are taking down material that doesn’t follow its updated policy. Rights holders can also request removal of content.

The strong pushback from the creative community could be a strategy to force OpenAI into entering licensing agreements for the content they need, legal experts said.

Existing law is clear — a copyright holder has full control over their copyrighted material, said Ray Seilie, entertainment litigator at law firm Kinsella Holley Iser Kump Steinsapir.

“It’s not your job to go around and tell other people to stop using it,” he said. “If they use it, they use it at their own risk.”

Disney, Universal and Warner Bros. Discovery have previously sued AI firms MiniMax and Midjourney, accusing them of copyright infringement.

One challenge is figuring out a way that fairly compensates talent and rights holders. Several people who work within the entertainment industry ecosystem said they don’t believe a flat fee works.

“Bring monetization that is not a one size fits all,” said Dan Neely, chief executive of Chicago-based Vermillio, which works with Hollywood talent and studios and protects how their likenesses and characters are used in AI. “That’s what will move the needle for talent and studios.”

Visiting journalist Nilesh Christopher contributed to this report.

The numbers are nothing short of staggering. Take Sam Altman, Open AI’s CEO. He reportedly wants 250 gigawatts of new electricity—equal to about half of Europe’s all-time peak load—to run gigantic new data centers in the U.S. and elsewhere worldwide by 2033.

Gold prices increase in both international and local markets.

In the international bullion market, the price of gold rises by $21 per ounce, reaching $4,016.

In the local market, the price of gold per tola increases by Rs 2,100 to reach Rs 422,700.

Similarly, the price per 10 grams rises by Rs 1,800, closing at Rs 362,397.

The upward trend reflects ongoing fluctuations in global demand and market conditions.

Read: Gold prices hit record high, cross Rs425,000 mark

Earlier, Spot gold fell nearly 2% to $3,959.48 per ounce by 01:53 p.m. ET (17:53 GMT). U.S. gold futures for December delivery fell 2.4% to settle at $3,972.6.

Wu, however, immediately outlined a clear road map for Alibaba’s AI development, with a goal towards so-called artificial superintelligence (ASI) – when the firm’s Qwen open-source models and cloud services would serve as the software and computing infrastructure of the future.

In essence, Alibaba aimed to become the “world’s leading full-stack AI service provider”, he said. Alibaba owns the Post.

Do you have questions about the biggest topics and trends from around the world? Get the answers with SCMP Knowledge, our new platform of curated content with explainers, FAQs, analyses and infographics brought to you by our award-winning team.

The blueprint laid out in Wu’s 23-minute speech signified not just a strategic upgrade for Alibaba, but also highlighted the competition between Chinese and US tech giants for the future of artificial intelligence – a field that has drawn some of the largest investments in history, with profound economic, social and geopolitical implications.

As he spoke, Alibaba’s shares surged to a four-year high in Hong Kong, leading several banks to raise their price targets for the stock.

Alibaba CEO Eddie Wu Yongming. Photo: Weibo

A day later, US chipmaker Nvidia’s co-founder and CEO Jensen Huang referenced Wu’s remarks during a podcast with tech investors Brad Gerstner and Bill Gurley, in which he underscored the importance of spending big on AI.

The AI arena has now shifted from just large language models to include upstream hardware and downstream applications, according to Kyle Chan, a postdoctoral researcher at Princeton University.

China was engaged in a “different AI race” from the US, and it was no longer enough to have the strongest foundational model: one must also possess the best chips, algorithms and applications across the entire AI stack to stand out in a crowded field, Chan said.

“Only in a pure ‘race to AGI’ world would the US be miles ahead, but that is probably not the world we live in,” he said, referring to artificial general intelligence – a hypothetical AI system capable of matching human performance in economically valuable tasks.

Some estimates suggested that US and Chinese tech giants would collectively spend more than US$400 billion on AI infrastructure this year – roughly equivalent to the gross domestic product of Romania, the world’s 39th-largest economy according to the International Monetary Fund.

That prompted some analysts to argue that the AI competition between China and the US was now being waged by “hyperscalers” – the world’s largest tech companies with major capabilities across the entire AI stack.

Both Washington and Beijing have voiced support for their respective AI industries. The Trump administration’s AI Action Plan, released in July, aimed to promote the export of “American AI” technology globally, led by Nvidia and OpenAI – the world’s most valuable company and start-up, respectively.

Open AI CEO Sam Altman. Photo: Getty Images/TNS

As part of their partnership, Nvidia is helping OpenAI establish its own “self-hosted” data centres, which the start-up previously relied on Microsoft to provide. The move could also allow OpenAI to catch up with Tesla founder Elon Musk’s xAI, which is building its own Colossus data centres in Memphis, Tennessee.

Alongside its recent deals with Advanced Micro Devices (AMD) and Samsung Electronics, as well as the US$500 billion in pledged funding for the Stargate Project – OpenAI’s joint venture with SoftBank Group and Oracle – the start-up’s computing deals amounted to at least US$1 trillion this year.

More partnerships could be announced “in the coming months”, OpenAI CEO Sam Altman said on a podcast on Thursday.

“To make the bet at this scale, we kind of need the whole industry, or a big chunk of the industry, to support it,” he said. “And this is from the level of electrons to model distribution and all the stuff in between, which is a lot.”

China, too, has its share of hyperscalers, but their size lags behind their US counterparts. The big three American players – Amazon Web Services, Microsoft Azure and Google – command about 63 per cent of the US$900 billion global cloud computing market, according to Synergy Research Group.

In China, Alibaba’s AI and cloud computing arm Alibaba Cloud holds a clear lead with 36 per cent of the market, according to research firm Omdia.

At last month’s conference, Wu announced additional AI infrastructure spending beyond the initial US$53 billion commitment unveiled earlier this year. The company hinted that these extra funds would support the company’s largest overseas data centre expansion to date, including its first hubs in Brazil, France and the Netherlands. Wu said demand overseas “far exceeded” domestic growth.

Nvidia CEO Jensen Huang. Photo: AFP

Despite their advances, Chinese companies remained significantly behind their US peers in terms of investment. Alibaba’s three-year spending pledge is less than what any one of the US big three hyperscalers spends in a single year.

OpenAI is currently valued at US$500 billion, while US AI model developer Anthropic saw its valuation nearly triple to US$183 billion following a funding round in September. In contrast, China’s leading start-ups, such as Moonshot AI and Z.ai, are valued at US$3.3 billion and US$5.6 billion, respectively.

That did not necessarily mean China was falling behind in AI, Princeton’s Chan said. In the US, Silicon Valley executives – including Altman – stressed the urgency of beating China to achieve AGI.

The US preoccupation with achieving AGI before China had led to an excessive focus on scaling computing resources and restricting Chinese access to advanced semiconductors, at the expense of developing the full US stack, Chan said.

“Chinese policymakers are not ‘AGI-pilled’,” he said. “I think they see AI as something like the internet that can turbocharge, if not transform, existing industries, where the focus is on diffusing the technology broadly and increasing adoption,” said Chan, adding that he did not believe AGI was imminent.

Alibaba chairman Joe Tsai, meanwhile, has stressed the importance of adoption. At an event hosted by the US podcast All-In last month, he said the winner in AI should not be defined by “who comes up with the strongest AI model”, but on “who can adopt it faster”.

“I’m not saying China technologically is winning the model war,” he said. “But in terms of the actual application and also people benefiting from AI, it has made a lot of developments.”

The Chinese government is betting on the integration of AI with the country’s formidable industrial and manufacturing sectors to win the tech race, a strategy known as “AI plus”.

A Unitree robot takes part in an obstacle race at the World Humanoid Robot Games in Beijing. Photo: Reuters

China now leads the world in industrial robot installations, with a record 2.027 million active robots, according to the International Federation of Robotics.

The country’s humanoid robot market has also seen rapid growth, with prominent start-ups like Shanghai-based AgiBot and Hangzhou-based Unitree Robotics landing orders from state-owned firms.

In March, for the first time, Beijing designated “embodied intelligence” – AI integrated into physical machines – as a key future industry. Authorities later outlined plans to promote robotics adoption across various sectors, including manufacturing, aerospace and logistics.

Government support has filtered down to the entrepreneurial level, with nearly half of AI fundraising this year directed towards embodied intelligence start-ups, according to consultancy IT Juzi.

“China is running away with the hard-power part of AI – robotics,” Martin Casado and Anne Neuberger, a general partner and senior adviser, respectively, at Silicon Valley venture capital firm Andreessen Horowitz, said in a recent post.

“We start seeing intelligence embedded in the physical world – culminating in generalist robots that perform a wide variety of tasks across applications, from manufacturing to services to defence,” they wrote. “The country betting on that future is China, not the US.”

Signs indicate that the US increasingly recognises the importance of AI applications in hard technology. OpenAI is reportedly ramping up hires for its robotics team and has partnered with autonomous driving start-up Applied Intuition.

However, none of the world’s “big four” industrial robotics firms – ABB Robotics, Fanuc, Kuka and Yaskawa Electric – are based in the US.

Huawei’s computing cluster on display at the World Artificial Intelligence Conference in Shanghai. Photo: NurPhoto via Getty Images

The spending disparity between Silicon Valley and Chinese firms may not be critical, as Chinese hyperscalers do not always compete directly with their US counterparts, according to Poe Zhao, a Beijing-based tech analyst and founder of the Hello China Tech newsletter.

“At least in the AI field, the market has become completely parallelised, with each playing its own game,” he said. “I think many people in the English-speaking world do not understand just how big the Chinese cloud market really is, with many demands from different segments, from large state-owned enterprises to small and medium-sized enterprises.”

“It is impossible for any company to be like Amazon – to be a one-stop shop that meets everyone’s needs, which gives Alibaba, Huawei, Baidu and ByteDance different opportunities.”

It also remained unclear just how far ahead US foundational models were compared to their Chinese rivals, according to Tilly Zhang, a Beijing-based tech analyst at Gavekal Dragonomics.

Chinese models consistently top popular global AI leader boards, particularly in image and video generation, often delivering comparable performance at a fraction of the training costs of US competing products.

The US government acknowledged the potential of China’s open-source ecosystem in driving global adoption.

Meanwhile, partners at Andreessen Horowitz pointed out that US start-ups and universities were heavily reliant on Chinese models.

The AI Action Plan emphasised the need for the US to develop leading open-source models, as the country’s previous open-source leader, Facebook owner Meta Platforms, has signalled it is no longer interested in open-sourcing its Llama models.

OpenAI swiftly responded to the government’s call in August with its first open model in six years, but the gap with China’s well-established ecosystem – similar to that in robotics – may already be too wide to bridge, according to open-source AI expert Nathan Lambert.

“Qwen alone is roughly matching the entire American open model ecosystem today”, Lambert said at a recent industry conference.

He highlighted the depth of China’s open-source ecosystem, which spans from Big Tech giants such as Huawei Technologies and ByteDance to unexpected developers like food delivery giant Meituan and Alibaba’s fintech affiliate Ant Group, which open-sourced a 1 trillion-parameter model on Thursday.

Just as OpenAI has allied itself with Nvidia and AMD, a self-sufficient AI ecosystem is emerging in China through a collaboration between Huawei and DeepSeek.

In the latest example, when DeepSeek introduced a new programming language called TileLang as part of its new foundational model, Hygon Information Technology and Cambricon Technologies quickly announced “day zero” chip support for the new model, while Huawei said it was developing core operators for TileLang.

“This synchronicity suggests a strategic alignment,” Hello China Tech’s Zhao said. “It is the second phase of a deliberate campaign to build a self-sufficient AI stack, free from Nvidia’s influence.”

The jury is still out on whether Chinese AI players can achieve ASI with local hardware, although Huawei touted that its clustering solution could address computing power needs.

Meanwhile, American lawmakers have called for broader chip export controls, believing access to US technologies remains crucial for China’s AI ambitions.

At the Apsara conference, hundreds of developers and customers listened intently to the presentations, many using a Qwen-powered translation and transcription tool. Alibaba appeared undeterred, as it stressed its commitment to cultivating a vibrant AI ecosystem.

There would only be “five or six hyperscalers globally” in the future, Wu said, implying that Alibaba would be one of them.

This article originally appeared in the South China Morning Post (SCMP), the most authoritative voice reporting on China and Asia for more than a century. For more SCMP stories, please explore the SCMP app or visit the SCMP’s Facebook and Twitter pages. Copyright © 2025 South China Morning Post Publishers Ltd. All rights reserved.

Copyright (c) 2025. South China Morning Post Publishers Ltd. All rights reserved.