Introduction

Breast cancer (BC) has emerged as the most prevalent form of cancer affecting women and is a significant contributor to mortality rates among female carcinomas worldwide1,2. BC is a heterogeneous disease characterized by varying clinical features, behaviors, and treatment response profiles3. The classification of breast cancer subtypes aims to accurately predict prognosis and inform treatment strategies. BC molecular subtypes are classified into luminal A, luminal B, human epidermal growth factor receptor 2 (HER2)-enriched, and triple-negative, based on the expression levels of four key receptors: estrogen receptor (ER), progesterone receptor (PR), HER2, and Ki-67 4. Generally, luminal subtypes constitute the majority, accounting for approximately 70% of invasive breast cancers, and they typically respond favorably to endocrine therapy6,7. The selection of targeted antibody therapy is a priority for HER2-enriched BC8. In contrast, triple-negative breast cancer, which is negative for both hormone receptors and HER-2, lacks the opportunity for endocrine and targeted therapies, limiting treatment options to chemotherapy and radiotherapy, and resulting in the poorest prognosis9. The luminal subtype demonstrates a better prognosis and a favorable response to hormone receptor-targeted therapies, whereas the non-luminal subtype is more aggressive, associated with poorer outcomes than the luminal subtype, and exhibits a higher rate of locoregional recurrence along with a lower survival rate following distant metastasis10,11. Given that the prognosis of BC varies according to molecular subtypes, guidelines recommend the immunohistochemistry (IHC) assessment of these subtypes during the initial diagnosis12.

Currently, the preoperative assessment of the molecular subtypes of BC commonly relies on IHC results obtained after core needle biopsy. Acquiring a tissue specimen is an invasive procedure typically performed using true-cut biopsy. This technique may yield inconclusive results due to the potential undersampling of tumor tissue or inadequate biopsy quality. Furthermore, the limited amount of biopsy tissue complicates a comprehensive evaluation of tumor heterogeneity. Thus, there is an urgent need to identify an efficient method for differentiating luminal and non-luminal molecular subtypes in BC patients before surgery.

Radiomics is an innovative tool that extracts a substantial number of quantitative features from medical images using mathematical algorithms. These extracted features can represent the shape, intensity, and texture of tumors. Radiomic features have been employed to noninvasively evaluate intratumoral heterogeneity and are widely used in disease identification, differential diagnosis, prognosis prediction, and treatment response evaluation13,14. The ability of radiomics features generated from computed tomography, ultrasound, magnetic resonance imaging (MRI), and mammography for the prediction of response to neoadjuvant chemotherapy15,16, identifying the molecular subtype17, determining sentinel lymph node metastasis18, and distinguishing between benign and malignant breast tumor19 in BC has been demonstrated. Dynamic contrast-enhanced MRI (DCE-MRI) is a satisfactory imaging modality that can provide temporal information about contrast agent dynamics in suspicious lesions at acceptable spatial resolution. It is a sensitive sequence for detecting breast cancer lesions, especially dense breast lesions. DCE-MRI can indirectly reflect abnormal tumor vascular proliferation through lesion hemodynamic characteristics. Several studies have demonstrated that radiomics features based on DCE-MRI have the ability to discriminate the luminal from non-luminal BC. Although this study only included 91 patients, Leithner et al.20 demonstrated that DCE-MRI radiomic signature had the potential to differentiate between luminal and triple-negative BC. In addition, previous work evaluating the four intrinsic molecular subtypes has focused primarily on intratumoral regions and has not adequately considered the regions surrounding breast tumors. However, growing evidence suggests that surrounding breast tissue may also contain important biological information related to lymphatics, blood vessels, and immune responses21. Niu et al.17 demonstrated that the radiomic signature derived from combined intratumoral and peritumoral regions improved prediction AUCs compared with those from intratumoral or peritumoral regions alone. Peritumoral regions can provide complementary information to intratumoral regions for the prediction of molecular subtypes. However, to our knowledge, the role of the peritumoral region in the preoperative differentiation of luminal and non-luminal molecular subtypes in BC patients has not been studied.

In this multicenter study, we studied whether intratumoral and peritumoral radiomics features derived from DCE-MRI could be adopted as a predictive biomarker for discriminating the luminal from non-luminal subtypes in patients with BC, and we assessed their diagnostic performance.

Results

Patient characteristics

The clinic-pathological characteristics are detailed in Table 1. The numbers of luminal and non-luminal in the training set, internal validation set, and external validation set were 99 and 63, 40 and 25, and 55 and 23, respectively. No significant differences were observed in age, menopausal status, lesion type, lesion internal enhancement, mass margin, histological type, and extensive intraductal component between luminal and non-luminal groups in the three sets (all P > 0.05). There was, however, significant differences in mass shape between the luminal and non-luminal groups in the internal validation set (P = 0.027) but not in the training (P = 0.658) and external validation sets (P = 0.674).

Table 1 Baseline characteristics of the study population.
Full size table

Radiomic feature selection

208 original radiomic features were extracted. Further details of radiomic features are provided in Supplementary Tables 1–2. After calculating the intraclass correlation coefficient (ICC), the numbers of reproducible features (ICCs > 0.75) from the intratumoral and peritumoral region were 89 and 55 respectively. Significant differences between luminal and non-luminal molecular subtypes by Mann-Whitney U test (P < 0.05) and Spearman correlation score (r > 0.9) were entered into the least absolute shrinkage and selection operator logistic regression algorithm (LASSO) to determine the most valuable radiomics features. The process of LASSO analysis was presented in Supplementary Figure. 1.

Development and evaluation of the models

Ultimately, 2, 3, and 5 features were selected to establish intratumoral, peritumoral, and combined intratumoral and peritumoral radscore, respectively. (Supplementary Table 3). The mean and standard deviation of these features are provided in Supplementary Tables 4 to conduct z-score normalization before calculation. According to univariate logistic regression analysis, intratumoral, peritumoral, and combined intratumoral and peritumoral radscore were independent predictors of luminal and non-luminal molecular subtypes in BC patients (Table 2). Therefore, intratumoral, peritumoral, and combined intratumoral and peritumoral radiomic signature were established. The radiomic signature of intratumoral was constructed using the following formula: radscore = 6.217–26.2220 × GLRLM_ShortRunEmphasis − 0. 0188 × GLDM_LargeDependenceLowGrayLevelEmphasis. The radiomic signature of peritumoral was constructed using the following formula: radscore = 13.140 + 1.1907e-09 × GLSZM_LargeAreaLowGrayLevelEmphasis + 189.1940 × GLDM_ DependenceNonUniformityNormalized − 0.0647 × GLDM_ LargeDependenceLowGrayLevelEmphasis. The radiomic signature of combined intratumoral and peritumoral was constructed using the following formula: radscore = -48.267 + 3.45e-06 × intratumoral_firstorder_Energy + 59.0963 × intratumoral_GLRLM_ShortRunEmphasis − 5.83e-10 × peritumoral_GLSZM_LargeAreaLowGrayLevelEmphasis + 125.3061 × peritumoral_GLDM_DependenceNonUniformityNormalized + 0.0513 × peritumoral_GLDM_LargeDependenceLowGrayLevelEmphasis. The distributions of these radiomic signature in the training set, internal validation set, and external validation set are shown in the Supplementary Figure. 2.

Table 2 Univariate logistic regression analysis for the risk factors in the training set.
Full size table

Multivariate logistic regression analysis for the risk factors in the training set are shown in Supplementary Table 5. A clinical-radiomic model was constructed by adding intratumoral radscore to the clinical factor [odds ratio (OR) = 0.231 for intratumoral radscore, P < 0.001, 95% confidence interval (CI): 0.144–0.370; OR = 2.329 for menopausal status, P = 0.045, 95% CI: 1.017–5.333,]. The area under the receiver operating characteristic curve (ROC-AUC) for the clinical-radiomic model was 0.858 [95% CI:, 0.795–0.908], 0.833 [95% CI: 0.720–0.914], and 0.864 [95% CI: 0.768–0.931] in the training set, internal validation set, and external validation set, respectively. The AUC, accuracy, sensitivity, and specificity of each radiomic signature and clinical-radiomic model are shown in Table 3. The ROC curves of radiomic signatures and clinical-radiomic model are shown in Fig. 3. The Hosmer–Lemeshow test yielded non-significant results for the combined intratumoral and peritumoral radiomic signature in the training set (P = 0.979), internal validation set (P = 0.187 ), and external validation set (P = 0.162), indicating no departure from the perfect fit.

Table 3 Diagnostic performance of the radiomic signatures and clinical-radiomic model.
Full size table

Comparison of models

The DCA showed that the combined intratumoral and peritumoral radiomic signature had more net benefit than the other two radiomic signatures in predicting molecular subtypes when the threshold probability is 0-90% (Fig. 4). In comparison with the intratumoral radiomic signature, the combined intratumoral and peritumoral radiomic signature’s net reclassifcation improvement (NRI) was 0.300 (95% CI: 0.117–0.482), P = 0.001 in internal validation set; NRI, 0.224 (95% CI: 0.038–0.410), P = 0.018 in external validation set. The AUC of the clinical-radiomic model was not improved compared with the intratumoral radiomic signature in the internal validation set (0.833 vs. 0.838) and external validation set (0.864 vs. 0.883).

Discussion

In this multicenter study, we developed and validated radiomic signature that incorporates both intratumoral and peritumoral radiomic features for the noninvasive, individualized prediction of luminal and non-luminal molecular subtypes in BC patients. The results demonstrated that the combined intratumoral and peritumoral radiomic signature had good discrimination, with AUC values of 0.956, 0.945, and 0.896 in the training set, internal validation set, and external validation set, respectively, and DCA supported its clinical usefulness.

Accurately differentiating between luminal and non-luminal molecular subtypes in patients with BC is essential for selecting the most appropriate treatment11. However, preoperative biopsies may result in errors in the discrimination of molecular subtypes due to the sampling of only small lesion areas. Additionally, interobserver variability in tumor subtyping can occur, even among experienced breast pathologists22. Furthermore, inaccurate preoperative subtyping may result in inadequate treatment, necessitating further surgery and increasing morbidity. It is well established that luminal tumors exhibit a more favorable prognosis and demonstrate a positive response to endocrine therapy23. Conversely, non-luminal tumors tend to be more aggressive, typically requiring systemic chemotherapy and resulting in a poorer prognosis4. In our study, we included a total of 305 patients, of which 194 were classified as luminal and 111 as non-luminal. The luminal and non-luminal were 63.6% and 36.3%. The findings of our study are consistent with those reported in other series24,25,26.

With the development of biomedicine, researchers have found that peritumoral tissue is also closely related to tumor progression and prognosis, which has led to the rapid expansion of peritumoral radiomics research16,21. These studies have predominantly examined the MRI radiomic features of the peritumoral region, demonstrating that peritumoral MRI radiomics holds substantial predictive value for recurrence, chemotherapy response, and survival time in BC patients. The research conducted by Ding et al.27 systematically investigates the impact of peritumoral region sizes on the predictive performance of radiomic analysis. This study demonstrated that: (1) utilizing a peritumoral region thickness of 4 mm for sentinel lymph node status prediction resulted in the highest AUC in both the training set (AUC = 0.942) and the validation set (AUC = 0.841); (2) for radiologist’s ROIs, 4 mm yielded higher accuracy compared to other sizes; and (3) the prediction performance for molecular subtypes and histological grades could be improved after incorporating peritumoral features (4 mm). In this study, we chose 4 mm around the tumor boundary as the peritumoral region according to previous studies. Li et al.28 reported that MRI-based radiomics could distinguish between HER2 + versus HER2−, triple-negative versus others, with an AUC of only 0.65 and 0.67 respectively. Huang et al.24 also revealed that radiomic signature could distinguish between luminal and non-luminal, with an AUC of 0.80 in the validation set, but only 135 patients were included in this study, making it more difficult to accurately assess the feasibility of radiomics. In another study, Feng et al.25 demonstrated that radiomics based on the features of the intratumoral subregion could distinguish between luminal and non-luminal, however, the logistic regression classifier exhibited low performance, with AUC ranging from 0.571 to 0.718 in the validation set. However, these studies mainly focused on features from the intratumoral region, without considering the utility of the peritumoral microenvironment and lacking independent external validation. In this study, we found the combined intratumoral and peritumoral radiomic signature had an AUC of 0.945 and 0.896 in the internal and external validation set, respectively, much higher than using the intratumoral radiomic signature alone, which was 0.838 and 0.883, respectively. This may be due to the fact that some molecular subtypes of BC are difficult to identify based on intratumoral radiomic features alone, but can be discerned through peritumoral radiomic characteristics. Therefore, peritumoral radiomic features can provide valuable assistance in situations where intratumoral radiomic features alone fall short of enabling accurate differentiation. Besides, the present study included data from different MRI scanners and various acquisition parameters (matrix, image thickness, etc.), which could affect the stability and generalization of the radiomic features. Nonetheless, great efforts have been made to minimize the impacts (trilinear interpolation algorithm, Gaussian filter, etc.). Our results indicated that 4 radiomic models showed good diagnostic performance in the external validation set (all AUC > 0.850). Thus, the proposed prediction models based on intratumoral and peritumoral radiomic features in this study are more convincing compared to existing models. This is due to the satisfactory performance and generalization of the combined model across multicenter and cross-field strength validations, indicating its capability to accurately discriminate between luminal and non-luminal BC.

We hypothesized that genetic and structural differences, along with intratumoral heterogeneity between luminal and non-luminal breast cancers, would result in greater grayscale variation and distinct textural properties in DCE-MRI images. This variation could be analyzed through quantitative radiomics. Our study demonstrates that luminal and non-luminal tumors exhibit different radiomic signatures on DCE-MRI images. Energy is a measure of the magnitude of voxel values in an image. A larger value implies a greater sum of the squares of these values. We found a lower absolute value for the Energy from the intratumoral region of the luminal subtype, indicating that luminal subtypes are more complex in the image. This result is in concordance with previous radiomics studies29,30. The four selected texture features for the radiomic signature of combined intratumoral and peritumoral can be used to comprehensively quantify tumor heterogeneity across different spatial scales at different directional orientations. The significant gray level run length matrix feature included ShortRunEmphasis from the intratumoral region. ShortRunEmphasis is a measure of the distribution of short run lengths, with a greater value indicative of shorter run lengths and more fine textural textures. The luminal had a higher ShortRunEmphasis, indicating that luminal have a finer texture and smaller tumor heterogeneity in the intratumoral region. The significant gray-level size zone matrix feature included LargeAreaLowGrayLevelEmphasis from the peritumoral region, which measures the proportion in the image of the joint distribution of larger size zones with lower gray-level values. The non-luminal subtype had a higher LargeAreaLowGrayLevelEmphasis, indicating that non-luminal have greater tumor heterogeneity in the peritumoral region. Besides, the remaining two texture features (DependenceNonUniformityNormalized, LargeDependenceLowGrayLevelEmphasis) of gray level dependence matrix from the peritumoral region were also selected to develop the radiomic signature of combined intratumoral and peritumoral. All of these features serve as indicators of image heterogeneity or homogeneity. For instance, DependenceNonUniformityNormalized measures the similarity of dependence throughout the image, with a lower value indicating more homogeneity among dependencies in the image. As expected, we found higher values for the features that reflect image heterogeneity in non-luminal tumors.

Deep learning (DL) techniques have been thoroughly investigated for the analysis of medical images, showcasing exceptional performance in different areas. The ability of DL for the prediction of response to neoadjuvant chemotherapy31, identifying the molecular subtype32, and determining axillary lymph node metastasis33 has been demonstrated. However, the key properties and dependencies of DL techniques must be considered before commencing work on DL modeling for real-world applications: (1) DL typically relies on a substantial amount of data to construct a data-driven model for a specific problem. This dependence arises from the fact that when the volume of data is limited, DL algorithms tend to exhibit suboptimal performance34. In this study, only 162 patients were used to construct model. Thus, DL was not utilized; (2) DL relies more on high-performance machines with GPUs than radiomics methods, which limits its generalizability; (3) In general, training a DL algorithm is time-consuming due to the extensive number of parameters involved; thus, the model training process takes longer, whereas training with radiomics takes relatively little time; and (4) It’s difficult to explain how a DL result was obtained, i.e., “black-box”. Radiomics provide a standard process, and the extracted features are easily interpretable.

Our study has several limitations. First, the retrospective design may have introduced selection bias. Secondly, this multicenter study has a relatively small sample size for model development and validation, especially in the validation sets. This limits the statistical power and generalizability of the results. Therefore, further studies involving larger datasets are needed to ensure the generalizability of our findings. Thirdly, the study relies on manual segmentation of tumor regions, which is time-consuming and subject to inter-observer variability. Automated or semi-automated segmentation methods (e.g. DL algorithm) in future studies is urgently required, which could improve the efficiency and reproducibility of radiomics analysis, especially in larger datasets. Additionally, we also expect that the predictive performance can be enhanced by incorporating more advanced radiomic features and clinicopathologic characteristics.

Conclusions

In conclusion, this study proposed a radiomic signature combining intratumoral and peritumoral features obtained from DCE-MRI, which performed well in distinguishing luminal and non-luminal breast cancer. Further validation of this result is necessary before the clinical application of the model.

Methods

Patients

The Ethics Committees of Sichuan Cancer Hospital & Institute, Sichuan Provincial Maternity and Child Health Care Hospital, and Chengdu Women’s and Children’s Central Hospital approved this retrospective study and waived the need for informed consent. All methods in this study adhered to relevant guidelines and regulations, and the protocols received approval from the Ethics Committees of Sichuan Cancer Hospital & Institute, Sichuan Provincial Maternity and Child Health Care Hospital, and Chengdu Women’s and Children’s Central Hospital. We conducted this study in accordance with the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) statement35, categorizing it as type 3. The checklist was presented in Supplementary Table 6. From September 2017 to November 2021, 362 female patients from three hospitals were consecutively recruited in this study. Inclusion criteria were as follows: (1) patients diagnosed with breast cancer by surgical or biopsy pathology; (2) DCE-MRI examinations were performed within two weeks before surgery; (3) no anti-tumor treatment before baseline DCE-MRI examination. The exclusion criteria were as follows: (1) unmeasurable breast lesion (< 5 mm) in the baseline MRI (n = 10); (2) significant artifacts on MRI images (n = 9); (3) chemotherapy or biopsy before MRI (n = 18); (4) patients with incomplete clinical characteristics (n = 20). 305 patients with breast cancer were finally enrolled. These patients were divided into three groups, namely the training set (n = 162) from hospital 1, internal validation set (n = 65) from hospital 1, and an external validation set (n = 78) from hospital 2 and 3. The flowchart of the patient selection process is shown in Fig. 1.

Fig. 1
figure 1

Flowchart of patient inclusion and exclusion criteria. DCE-MRI, dynamic contrast-enhanced MRI.

Full size image

Molecular classification

Histopathological and IHC data were obtained directly from surgical pathology reports. In this study, the critical value of Ki-67 status was set at 20% 36. ER and PR positivity was defined as > 1% staining36,38. If the IHC staining result is 3+, HER2 expression was considered positive. HER2 2 + results were considered borderline and required confirmation by fluorescence in situ hybridization39. Therefore, molecular subtypes of breast cancer were calculated based on the status of ER, PR, Ki-67, and HER2.

MRI acquisition

Patients in hospital 1 underwent DCE-MRI using a 16-channel breast coil on a 3.0-T MRI scanner (Skyra, Siemens Healthcare, Erlangen, Germany) for the training and internal validation set. The external validation set included patients from hospital 2 (GE Signa HDxt, dedicated 7- channel phased-array breast coil) and hospital 3 (Ingenia, Philips, Amsterdam, the Netherlands, dedicated 7- channel phased-array breast coil). Gadodiamide (0.1 mmol/kg; Omniscan, GE Healthcare) was intravenously administered using a power injector (rate, 2.5 mL/s) before DCE-MRI. Then, a 20-mL saline fush was injected (rate, 2.5 mL/s). To avoid magnetic stress and motion artifacts, patients were instructed to remove all metal-containing objects and lie prone in the scanner before scanning. All patients must wear earplugs and headphones before their MRI scan to reduce noise. Supplementary Tables 7–8 show the details of the acquisition and reconstruction parameters. The timing of the first post contrast phase image of training set, internal validation set, and external validation set were 19s, 83s, and 83s, respectively.

MRI image analysis

All MRI images were retrospectively analyzed by two radiologists in accordance with the American College of Radiology Breast Imaging Reporting and Data System. Both radiologists were blinded to the pathological outcomes and assessed the following MRI morphologic features of each case on the first post-contrast phase of DCE-MRI: shape (oval/round or irregular), lesion type (mass or non-mass enhancement), margin (circumscribed or irregular/spiculated) of masses, and lesion internal enhancement (homogeneous or heterogeneous). Discrepancies between the two radiologists were resolved by consulting a third radiologist with 23 years of experience in breast imaging.

Preprocessing of images

The fifth phase images from 26 dynamic phases in the training set, the fourth phase images from 20 dynamic phases in the internal validation set, and the first phase images from 6 dynamic phases in the external validation set were selected for processing and segmentation. These images correspond to the first post-contrast images captured during the early arterial phase40. These images were then preprocessed using Artificial Intelligence Toolkit software (A.K., version 3.2.0, GE Healthcare). All images were firstly resampled into isotropic voxels of size 1 × 1 × 1 mm3 using a trilinear interpolation algorithm. The position of the interpolation grid is centered, and the dimensions of the interpolation grid are rounded to the nearest integer. Secondly, Gaussian filtering and bias field correction were performed after reconstruction. Finally, in order to eliminate the influence of different gray value ranges, Z-score normalization was used to normalize the image.

Image segmentation, extraction/selection of radiomic features

Three-dimensional manual segmentation of intratumoral region of interest (ROIs) was delineated by a junior radiologist, with 8 years of breast cancer diagnosing experience, using the ITK-SNAP software (v. 3.8.0, http://www.itksnap.org). The ROIs were drawn to cover the whole tumour on each consecutive section of the preprocessed MRI images; cystic and necrosis areas were avoided when delineated ROIs. Multicentre and multifocal cases were selected for the largest lesions. Peritumoral regions (4 mm of extravasation) was automatically mapped using the A.K. software by equidistant 3-dimensional dilation of the intratumoral ROIs27. The peritumoral ROI was ultimately obtained by subtracting the intratumoral ROI from the dilated ROI. Figure. 2 shows a typical DCE-MRI image and its associated intratumoral and peritumoral ROIs. To assess interobserver reliability, 30 cases were selected randomly and re-delineated by another senior radiologist with 25 years of experience in diagnosing breast cancer. The same delineation method and software were employed to calculate the ICC. Both radiologists were blinded to the clinical and pathologic details of each subject when ROIs delineation. Radiomics features with an ICC greater than 0.75, indicating excellent stability, were deemed suitable for subsequent analysis.

Fig. 2
figure 2

Representative slice of the dynamic contrast-enhanced MRI images and the corresponding intratumoral and peritumoral regions of interest (ROIs). The intratumoral ROI (red regions) drawn by a junior radiologist with the peritumoral ROI (purple regions) generated by equidistant 3-dimensional dilation of the intratumoral regions with 4 mm. (a, b) A 36-year-old female with luminal breast cancer. (c, d) A 61-year-old female with non-luminal breast cancer.

Full size image

A total of 208 radiomic features from the intratumoral and peritumoral ROIs (104 features for each) were extracted using an open-source Python software (PyRadiomics, version 3.0, https://pyradiomics.readthedocs.io)41 that followed the image biomarker standardisation initiative standard42. The feature set was divided into 3 groups: (I) shape, (II) first order statistic, and (III) texture features. Additionally, image intensity was discretized using a fixed bin number of 32 43. All features were standardized using z-score normalization before feature selection. After that, to reduce the number of irrelevant and redundant radiomic features, we devised a three-step program for dimensional reduction and robust feature selection. First, we used the Mann-Whitney U test to select the features that were statistically different between patients with luminal and non-luminal. Features with P < 0.05 were retained as potential predictive features. Second, the correlation between features was calculated by the Spearman correlation coefficient. For features with a correlation coefficient > 0.9, one of the two features was retained. Third, the LASSO44, with penalty parameter tuning conducted by 10-fold cross-validation, was applied to select molecular subtypes-related features with nonzero coefficients from the training set (Figs. 3 and 4).

Fig. 3
figure 3

Receiver operating characteristic curves of radiomic signatures and clinical-radiomic model in distinguishing between luminal and non-luminal subtypes in the training set (a), internal validation set (b), and external validation set (c), respectively.

Full size image
Fig. 4
figure 4

The DCA for the radiomic signatures of intratumoral, peritumoral, and combined intratumoral and peritumoral in the training set. The DCA showed that the combined intratumoral and peritumoral radiomic signature had more net benefit than the other two radiomic signatures in predicting molecular subtypes when the threshold probability is 0-90%.

Full size image

Radiomic signature building

The radscore was calculated using a linear combination of the selected features, which were weighted according to their respective coefficients. Radiomic signature of intratumoral, peritumoral, and combined intratumoral and peritumoral were built using the radscore respectively in the training set.

Clinical-radiomic model construction

Univariate logistic regression analysis was employed to identify associations between variables and molecular subtypes in the training set. Variables included in multivariable logistic regression analysis for further selection were those with a statistical significance (P < 0.05) or with clinical importance. Subsequently, a clinical-radiomic model that integrated the radscore (intratumoral, peritumoral, and combined intratumoral and peritumoral) and the significant independent risk factors was developed.

Performance and validation of the established model

The predictive performance of the radiomic signatures and clinical-radiomic model was evaluated in the training set and subsequently validated in the validation set using the AUC. The threshold for classification was determined by employing the maximum Youden index (sensitivity + specificity − 1) on the training set, and this threshold was consistently applied to the validation set. Additionally, we calculated various performance metrics, including accuracy, sensitivity, and specificity. The 95% CIs were computed using the exact binomial method. The goodness-of-fit of the predictive models was evaluated using the Hosmer-Lemeshow test. The clinical utility of the radiomic signatures was evaluated by the DCA45. Net NRI measure recommended by Steyerberg et al.46 was used in this study to evaluate the performance of prediction models.

Statistical analysis

SPSS 23.0 (version 26.0, available at https://www.ibm.com) and MedCalc (version 18.2.1, available at https://www.medcalc.org) were employed for all statistical analyses. Feature selection and radiomic signature building were conducted using R software (version 4.0.0, http://www.r-project.org). Group differences were compared using the independent two sample t test for continuous variables and the chi-square test (or Fisher’s exact test) for categorical variables. A p-value of less than 0.05 was considered statistically significant (two-tailed).