Breast cancer (BC) is one of the most common malignancies in women worldwide, particularly affecting women aged 30 to 45 years (1). It is the second most deadly malignancy affecting women (2), representing a major health concern for female populations. According to the St. Gallen International Breast Cancer Expert Consensus (3), BC can be classified into four subtypes based on four immunohistochemical markers (4). Among these, Luminal A and Luminal B types are categorized as Luminal-type BC, accounting for approximately 60% to 70% of all cases (5), and typically respond well to endocrine therapy (6, 7). Meanwhile, HER2-OE and triple-negative types are classified as non-Luminal BC, often requiring targeted therapy or more intensive treatment regimens, and are associated with poorer prognosis (8–10). Since these two types show significant differences in treatment selection and prognosis (11), accurate preoperative differentiation of subtypes is crucial for developing individualized treatment plans. Currently, BC molecular subtyping primarily relies on histopathological detection methods such as immunohistochemistry and in situ hybridization, but these techniques are invasive and prone to sampling bias (12).

In recent years, radiomics has emerged as a research focus for preoperative BC subtyping due to its non-invasive nature and strong reproducibility. Numerous studies have attempted to develop predictive models based on imaging modalities such as magnetic resonance imaging (MRI) and ultrasound (US) (13–16). However, MRI examinations are costly, time-consuming, and susceptible to motion artifacts (17). More importantly, its multi-parameter scanning characteristics affect the stability of radiomic features, thereby compromising model generalizability (18). US demonstrates limited capability in discriminating small lesions (19), while its operator-dependent image acquisition leads to poor reproducibility and biological consistency of extracted radiomic features (20). Therefore, there is an urgent need for more efficient and precise radiomic approaches to optimize preoperative BC subtyping. As an emerging imaging technology, dual-layer detector spectral CT (DLCT) offers high resolution and provides multiple spectral images, thereby expanding radiomics research possibilities (21–23). Preliminary studies have demonstrated DLCT radiomics’ effectiveness in predicting malignant tumors (24, 25), yet its application in BC molecular subtyping remains unexplored.

Therefore, this study aims to innovatively integrate DLCT radiomic features with machine learning (ML) algorithms to construct a non-invasive discrimination model for Luminal versus non-Luminal BC, exploring the potential of DLCT radiomics for precise BC subtyping.

This study is the first to propose a ML model based on DLCT chest enhancement imaging radiomic features for distinguishing Luminal from non-Luminal BC. The results demonstrate that the GNB model combining 13 radiomic features and 3 clinical features exhibits good predictive performance (AUC = 0.778). These findings provide valuable reference for early diagnosis and precision treatment of Luminal and non-Luminal BC, offering new imaging evidence for future subtyping research.

Recent years have seen increasing research focus on developing non-invasive methods for early differentiation between Luminal and non-Luminal BC. Studies by Xu et al. (31), Feng et al. (32), and Wang et al. (33) developed models using MRI radiomic features and functional parameters, with AUC values of 0.830, 0.879, and 0.830 respectively. Liu et al. (34) developed a model using US radiomic features with an AUC of 0.752. In terms of model performance, MRI-based models outperformed US-based ones. Umutlu et al. (35) created a model using PET/MRI radiomic features with an AUC of 0.950, though without further test set validation and with higher PET/MRI costs. Liu et al. (36) developed a predictive model based on DLCT quantitative parameters with an AUC of 0.754, performing less well than our DLCT radiomics-based model. Our GNB model outperformed US radiomics-based and DLCT quantitative parameter-based models, though showed lower performance than MRI radiomics, functional parameter, and PET/MRI radiomics-based models. However, compared to MRI and PET/MRI examinations, DLCT offers higher examination efficiency and lower cost. Additionally, unlike the prone position typically used in MRI, DLCT’s supine position matches surgical positioning, minimizing potential image and lesion deviations caused by posture changes, while simultaneously evaluating skin, chest wall, internal mammary lymph nodes, bilateral axillary lymph nodes, and supraclavicular lymph nodes (37).

Among the seven ML models constructed based on the same DLCT radiomics features, Figures 4, 5; Table 2 reveal that the XGBoost, LightGBM, AdaBoost, and RF models exhibit outstanding predictive performance on both the training and validation sets (AUC = 1.000 on the training set, and AUC > 0.900 on the validation set). However, their performance significantly declines on the independent test set (AUC < 0.600). This marked performance discrepancy strongly suggests the presence of overfitting in these models. We attribute the occurrence of overfitting primarily to the high-dimensional, small-sample challenge faced in this study. A total of 1,037 features were extracted from each patient’s images, while the initial training set comprised only 89 samples, expanding to 130 after SMOTE balancing. In such a scenario where the feature dimension substantially exceeds the sample size, models are highly prone to capturing noise and spurious correlations in the training data, leading to diminished generalization capability. To mitigate overfitting, multiple safeguards were incorporated into the study pipeline, such as SMOTE, LASSO, cross-validation, and the strict establishment of an independent test set. Although overfitting was observed in some complex models, the GNB model we ultimately selected demonstrated the smallest performance gap across the training, validation, and test sets, with AUC values of 0.900, 0.869, and 0.778, respectively. Although the GNB model assumes strong independence among features, multiple previous radiomics modeling studies have reported that the GNB model demonstrated the best performance in their research (38, 39). We hypothesize that in high-dimensional, small-sample scenarios, the strong independence assumption of the GNB model can, to some extent, prevent the model from fitting noise and complex feature interactions in high-dimensional data, thereby enhancing its generalization capability. In contrast, more flexible models such as RF and XGBoost are more prone to overfitting. The robust generalization performance of the GNB model is further corroborated by the specific classification patterns revealed in its confusion matrix. As shown in Figure 7, the model correctly identified 25 cases of Luminal BC (true positives) and 7 cases of non-Luminal BC (true negatives) in the test set (n=39), achieving an overall accuracy of 84.6%. The model’s errors exhibit a clear asymmetry, with false negatives being the primary source of error, while false positives are relatively fewer. Future feature engineering efforts could focus on these misclassified cases.

Radiomics converts medical images into quantitative, objective features to non-invasively explore tumor heterogeneity and characteristics (16). Existing research has demonstrated radiomics’ potential for non-invasive BC subtyping, though most studies utilized MRI or US. Feng et al. (32) combined clinical factors with intratumoral subregion MRI radiomic features to develop a nomogram model (AUC = 0.830), while Wu et al. (40) created a nomogram based on ultrasound radiomic features (AUC = 0.767). Previous CT-based investigations such as the work of Wang et al. (41), who created a radiomic model distinguishing Luminal from non-Luminal BC (AUC = 0.757) using CT features. To our knowledge, no prior studies have utilized DLCT radiomic features to differentiate between Luminal type and non-Luminal type BC. Our DLCT radiomics-based GNB model achieved test set AUC, accuracy, sensitivity and specificity of 0.778 (95% CI: 0.582–0.974), 0.821, 0.833 and 0.778 respectively, demonstrating good performance. Among the 13 radiomic features ultimately selected for modeling, three were shape features and first-order statistics from the original images: ‘original_shape_MajorAxisLength’ representing the longest axis, ‘original_shape_Elongation’ indicating the ratio of the shortest to longest axis, and ‘original_firstorder_90Percentile’ denoting the 90th percentile value. The other ten features, accounting for the largest proportion, were all wavelet-transformed first-order statistical and texture features. Bian et al. (42) found wavelet features played important roles in their multiparametric MRI-based model predicting HER2-low BC; Yang et al. (43) demonstrated strong correlations between wavelet features and neoadjuvant chemotherapy response; Zhou et al. (44) showed wavelet features’ good performance in evaluating neoadjuvant chemoradiotherapy for BC patients. These findings suggest wavelet features may have predictive value in BC, consistent with our results. From Figure 9a, it can be observed that the top five most contributing radiomics features in the GNB model are waveletHHH_firstorder_Skewness, waveletHLL_firstorder_Kurtosis, waveletHHH_glszm_GrayLevelNonUniformityNormalized, waveletHLH_glcm_Imc2, and waveletHHL_firstorder_Median. The wavelet-HHH-firstorder-Skewness measures the asymmetry of the CT value distribution in the tumor region. Figure 9b shows that the lower values of this feature are concentrated in the right high SHAP value region. The lower the skewness value, the more the model tends to classify the tumor as the Luminal type. In imaging, lower skewness indicates a more symmetric CT value distribution. We speculate that this may reflect a relatively homogeneous microenvironment in this type of tumor, lacking significant microcalcifications or micro-necrotic areas. Conversely, high skewness may be associated with heterogeneous components in non-Luminal types. Wavelet-HLL-firstorder-Kurtosis describes the kurtosis of the CT value distribution. As observed in Figure 9b, the lower the kurtosis value, the more the model tends to classify the tumor as Luminal type. Low kurtosis indicates a broad, dispersed distribution of voxel CT values. In contrast, high kurtosis typically signifies that voxel values are highly concentrated within a narrow range, which may correspond to highly homogeneous areas on imaging, such as large necrotic or liquefied regions or abnormally uniform areas of marked enhancement. These patterns may be more closely associated with certain features of non-Luminal types (45). Wavelet-HHH-glszm-GrayLevelNonUniformityNormalized quantifies the spatial dominance of different density regions in high-frequency texture details. Figure 9b shows that the higher its feature value, the greater its contribution to predicting the Luminal type. This suggests that, at the high-frequency texture scale, the microstructure of Luminal BC may be spatially dominated by a few highly homogeneous tissue components. On imaging, this may manifest as dominant regions composed of large, relatively uniform glandular parenchyma or stromal components. Wavelet-HLH-glcm-Imc2 evaluates the complexity and regularity of pixel gray-level co-occurrence patterns. Figure 9b shows that the lower its feature value, the greater its contribution to predicting the Luminal type. We hypothesize that the relatively regular structure of Luminal BC results in a low Imc2 value at the algorithmic level. Conversely, a high Imc2 may correspond to extreme homogenization of large-scale structures, such as extensive necrosis or abnormally uniform areas of enhancement, which are common features of non-Luminal BC. It is noteworthy that this study found that higher values of the wavelet-HHL-firstorder_Median feature (indicative of higher enhancement) positively contribute to predicting the Luminal type. This appears to contradict some conventional imaging views that non-Luminal types have richer blood supply (46). We propose that this discrepancy may arise from the following reasons: this study evaluates the median CT value in a specific frequency subband (HHL) after wavelet transformation, rather than the average enhancement of the entire tumor on the original images. The enhancement patterns observed at this specific texture scale may differ biologically from overall enhancement. Radiomics features capture spatial distribution and texture patterns. A high Median value may more strongly reflect the concentration and consistency of enhancement distribution at the HHL scale, rather than merely the peak enhancement intensity. Luminal BC may exhibit more uniform and consistent overall enhancement, whereas non-Luminal BC may display heterogeneous, focal marked enhancement (47).

The GNB model based on DLCT radiomics developed in this study provides proof-of-concept for a rapid, objective, and non-invasive tool for preoperative BC subtyping. We envision the following clinical integration pathway: the model is designed to serve as an auxiliary diagnostic tool, integrated into the post-processing stage of BC imaging examinations. After patients undergo chest contrast-enhanced DLCT scanning, radiologists or technicians can invoke this model on a PACS workstation or a dedicated radiomics analysis platform. By inputting the 55 keV images and performing delineation, the model will automatically extract radiomic features from the ROI and execute the prediction algorithm. Within seconds, the system will generate a structured report containing predicted probability values, confidence intervals, and visualizations of key discriminative features, among other information. The results can serve as an auxiliary reference for radiologists, reducing subjective variability in diagnosis. Moreover, prior to the availability of pathological results, it can provide clinicians with preliminary insights into molecular subtype tendencies, facilitating earlier planning for subsequent treatment strategy discussions.

This study has several limitations. First, the relatively limited sample size may affect the statistical power and stability of the model. Second, all data were derived from a single center, where patient population characteristics, imaging acquisition equipment, and protocols are relatively uniform. This may limit the generalizability of the model developed in this study to other institutions, different equipment, or diverse patient populations. Following this study, we plan to design a prospective study to conduct real-time validation with consecutively enrolled patients at our institution in the future, and actively seek multi-center collaborations. Third, despite the use of the SMOTE algorithm for data balancing, the imbalance in the number of Luminal-type and non-Luminal-type patients may have affected model performance. Future studies should continue to collect cases with an emphasis on balancing patient ratios, and include comparative experiments with and without SMOTE as a key component of the research. Fourth, this study employed manual ROI segmentation. Although strict standardized protocols were followed to ensure quality, the absence of a multi-observer consistency test may somewhat affect the reproducibility of the method. In subsequent research, we will prioritize the adoption of semi-automatic segmentation algorithms or organize multi-center observer agreement studies to further enhance reproducibility. Fifth, this study did not systematically evaluate the impact of variations in imaging acquisition and reconstruction parameters on the stability of radiomic features. Future research must incorporate rigorous testing of feature stability to identify robust features that are insensitive to technical parameters.

Based on DLCT radiomic features, this study preliminarily explored and constructed seven ML models. Among these, the GNB model demonstrated relatively superior and stable predictive performance in internal testing. The findings suggest that DLCT radiomics may offer a potential auxiliary tool for distinguishing between Luminal and non-Luminal BC, thereby potentially aiding in early diagnosis and preliminary discussions on treatment strategies. This study provides preliminary evidence and hypotheses for this field, but its clinical translation prospects urgently require further validation and advancement through large-scale, prospective, multicenter studies in the future.