Introduction

Extensive ductal carcinoma in situ (DCIS) or intraductal component (IC) in invasive breast cancer (IBC) (IBC-EIC) is a commonly mentioned risk factor for incomplete resection in breast-conserving surgery (BCS)1. The rates of positive margins2, re-operation3, and local–regional recurrence4 in patients with IBC-EIC were reported to be significantly higher than those observed in IBC without EIC patients. Meanwhile, the concept of optional surgical margins has evolved. Following the recommendation of achieving “no ink on tumor” for women with IBC5 and a 2 mm margin for women with pure DCIS in BCS6, breast surgeons are adjusting their resection margins increasingly close to the presumed border between healthy tissue and the known tumor (index tumor). Therefore, preoperative prediction of IBC-EIC may help surgeons tailor surgical resection margins to avoid re-excision or re-operation for positive margins in BCS.

Compared to the conventional imaging methods (mammography and breast ultrasound), it is well-established that breast MRI enhances the accurate delineation of the true extent of the known cancer7, particularly in cases of DCIS8 and IBC-EIC9. However, routine preoperative MRI for a known breast cancer is still a controversial indication for BCS patients10,11. As indicated by published results12,13,14, the use of preoperative MRI has shown a controversial impact on surgical outcomes, particularly concerning reoperation rates and mastectomy rates. One possible reason is that the lack of accurate imaging biomarkers for preoperatively identifying IBC-EIC results in a consistently high rate of additional surgery4. Hence, it is necessary to develop innovative MRI sequences capable of accurately predicting IBC-EIC preoperatively14.

Ultrafast DCE-MRI is a novel approach designed to capture kinetic information of tumor with high temporal resolution while maintaining reasonable spatial resolution15. Further, a quantitative analysis of the time–intensity (kinetic) curve derived from Ultrafast DCE-MRI can elucidate the tumor’s vascularity and perfusion status, facilitating the decoding of tumoral heterogeneity in vivo. Previous research has substantiated the efficacy of this approach in offering valuable imaging insights for breast cancer screening16, diagnosis17,18, molecular subtype classification19,20, and therapeutic evaluation21,22. However, the feasibility and capability of intra- and peri-tumor heterogeneity, decoded by quantitative analysis in Ultrafast DCE-MRI, for the prediction of IBC-EIC remains unknown.

In our previous radiomic study, the combined radiomic model, incorporating intra-tumoral and peritumoral features from an early phase of high-resolution Ultrafast DCE-MRI, demonstrated a moderate ability to predict IBC-IC preoperatively, achieving an AUC of 0.8223. In this recent study, our objective is to characterize kinetic intra-tumor heterogeneity (ITHkinetic) and peritumor heterogeneity (PTHkinetic) through quantitative analysis of clustering voxels exhibiting similar enhancement patterns in high-spatial resolution ultrafast DCE-MRI, using all dynamic phase information. Our hypothesis posits that the ITHkinetic and PTHkinetic can serve as promising imaging biomarkers, enhancing the ability to preoperatively discern IBC-EIC.

Materials and methods

We confirm that the Medical Ethics Committee of the Sichuan Cancer Hospital & Institute has granted approval for this study, with the informed consent requirement waived (approval number: SCCHEC2015029). All methods in this study followed approved protocols by The Medical Ethics Committee of the Sichuan Cancer Hospital & Institute in compliance with relevant guidelines and regulations. This study complies with the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD)24 and standards for reporting diagnostic accuracy (STARD)25 statements. The checklists are provided in Supplementary Materials.

Patients

Between September 2017 and December 2019, this study retrospectively enrolled 654 consecutive female patients. The inclusion criteria comprised patients with a breast mass who underwent surgical treatment and received high spatial resolution ultrafast DCE-MRI within 1 week prior to surgery. MRI was conducted for preoperative consultation and surgical planning for these patients. Exclusion criteria included patients with benign lesions (n = 82), only ductal carcinoma in situ (DCIS) (n = 78), prior breast surgery (n = 64), previous neoadjuvant or induction chemotherapy (n = 48), occult breast cancer (n = 2), lymphoma (n = 3), and those lacking qualified image data (n = 9). Finally, 368 patients were retained for further analysis (Fig. 1). Of these, 183 patients overlapped with the sample from our prior study, which focused on radiomic research for a different research aim23.

Figure 1
figure 1

Flow diagram of patient with inclusion and exclusion criteria. DCIS ductal carcinoma in situ, IBC invasive breast cancer, EIC extensive intraductal component.

Full size image

Pathologic analysis

Histopathology and immunohistochemistry data were directly extracted from the surgical pathology report. Two dedicated breast pathologists meticulously reviewed the existence and status of the intraductal component (IC). All cases underwent thorough evaluation to identify the presence of IC in hematoxylin–eosin staining sections. Additional immunohistochemical staining (cytokeratin 5/6 and p63) was performed as necessary. The status of the IC, including fraction and nuclear grade, was quantitatively measured and documented as its ratio to the entire lesion9. The nuclear grade of IC was categorized according to the modified Black nuclear grade system26. The histological grade was categorized according to the Nottingham grading system27. Estrogen receptor (ER) and progesterone receptor (PR) positivity was defined as the presence of 1% or more positively stained nuclei in 10 high-power fields. Human epidermal growth factor receptor 2 (HER2) positivity was defined as immunohistochemistry score of 3+ or 2+ with in situ hybridization amplification28.

In this study, IBC-EIC is defined as the presence of prominent IC within the confines of the invasive tumor, occupying at least 25% of the tumor, or the presence of IC in the grossly normal adjacent breast tissue. Additionally, it includes lesions predominantly composed of DCIS with one or more foci of invasive carcinoma1,4. In contrast, IBC without EIC is defined as invasive cancer lacking an EIC component.

MRI protocol

All eligible patients underwent MRI on the 3.0 T scanner (Skyra, Siemens Healthcare) with 16-channel breast coil. Axial T2WI images were acquired using turbo-inversion recovery-magnitude sequences. Gadodiamide (0.1 mmol/kg; Omniscan, GE Healthcare) was intravenously administered via a power injector at 2.5 mL/s, followed by a saline flush of 20 mL at the same rate. Axial T1-weighted DCE-MRI images were acquired through high-resolution ultrafast protocols using the CAIPIRINHA-Dixon-TWIST-VIBE sequence (Siemens Healthcare)23. The acquisition consisted of 26 phases and lasted for at least 5 min. The specific parameters of these sequences can be referenced in Supplementary Materials.

MRI features interpretation

Two radiologists, L.H.B. (15 years of experience) and C.Z. (5 years of experience), assessed DCE-MRI and T2WI features according to the breast imaging reporting and data system MRI lexicon, second edition, for inter-observer agreement. Sixty patients were randomly selected for intra-observer analysis, re-evaluated by radiologist 1 after a 1 month interval. The early DCE-MRI phase was employed to aid index tumor localization on T2WI. Both radiologists were blinded to pathological information. The detailed MRI features assessed in this study can be found in Table S1, and L.H.B.’s assessments were used for subsequent statistical analyses.

MRI model building

MRI model was constructed through univariate and multivariate analyses of MRI features. Significant predictors (p < 0.05), identified via univariate logistic regression, were integrated into a multivariate logistic regression model with backward stepwise selection, applying the likelihood ratio test with Akaike’s information criterion as the stopping rule.

Kinetic intra- and peri-tumor heterogeneity characterization

Kinetic intra- and peri-tumor heterogeneity (ITHkinetic and PTHkinetic) characterization was performed by Z.S.X. using MATLAB (R2018b, https://www.mathworks.com) following the process outlined below (Fig. 2):

Figure 2
figure 2

illustrates the process of kinetic intra and peri-tumor heterogeneity characterization. (A) Axial last-phase ultrafast T1-enhanced DCE MRI scan. (B) Tumor 3D segmentation display. (C) Intra-tumoral heterogeneity subregion. (D) High-resolution ultrafast DCE-MRI based time-signal intensity curve in the intra-tumor. (E) Peri-tumoral heterogeneity subregion. (F) High-resolution ultrafast DCE-MRI based time-signal intensity curve in the peri-tumor.

Full size image

Segmentation

The intra-tumoral volume region of interest (VOI) covering the entire MRI-visible tumor (i.e., index tumor) was manually delineated slice by slice in the axial plane using ITK-SNAP (version 3.8.0, http://www.itksnap.org). The largest tumor was selected as the index tumor in cases with multiple lesions. Segmentations were performed by one radiologist (C.Z., with 5 years of experience in breast imaging). To assess the intraclass correlation coefficient, 60 randomly selected cases were segmented again by another radiologist (L.H.B., with 15 years of experience in breast imaging). The peritumoral VOI was obtained by equidistant 3-dimensional dilation of the intra-tumoral regions by 4 mm using MATLAB.

Quantitative kinetic analysis

All dynamic phases of the entire acquisition process were utilized to quantitative kinetic analysis. The signal intensity of all post-contrast phases was normalized to a percentage change relative to the pre-contrast signal intensity. Quantitative kinetic analysis was conducted using a percentage ratio to display changes throughout the entire post-contrast acquisition process in DCE-MRI, relative to the signal intensity before contrast injection. Seven quantitative parameters were derived from the tissue uptake kinetics based on the average voxels of the intra- and peri-tumoral VOIs: time to enhancement (TTE), maximum slope (MS), peak of enhancement (SImax), time to peak intensity (TTP), wash-in slope (WIS), area under the signal intensity curve (AUC), and washout rate (WR)17.

Kinetic heterogeneity characterization

We employed TTP as a criterion for homogeneous voxel clusters29, analyzing the diverse patterns of contrast agent uptake dynamics observed within the VOI. Specifically, we defined the top-K as the set of clustered voxels achieving the fastest peak signal intensity within the first K%, while the opposite was defined as the bottom-K. For both intra and peri-tumor regions, we generated 10 groups, comprising top-10, top-20, …, top-50, and bottom-10, bottom-20, …, bottom-50, respectively. Each group underwent the aforementioned quantitative analysis from intra-tumoral and peritumoral VOIs, to derive ITHkinetic and PTHkinetic parameters.

Personalized ITHkinetic and PTHkinetic score construction

ITHkinetic and PTHkinetic parameters with intraclass correlation coefficient values < 0.75 and Pearson correlation coefficient > 0.80 were excluded. The remaining parameters were standardized using z-scores and normalized via min–max scaling. The most predictive ITHkinetic parameters associated with IBC-EIC were selected using the least absolute shrinkage and selection operator (LASSO) with tenfold cross-validation. Subsequently, a logistic regression model was constructed, and individualized ITHkinetic scores were calculated for each patient based on the model coefficients. A parallel process was applied to PTHkinetic parameters, involving LASSO selection, logistic regression modeling, and computation of PTHkinetic scores for individual patients.

Imaging combined model development

The Imaging Combined model was built by integrating significant MRI features identified in univariable analysis, along with the ITHkinetic and PTHkinetic scores. Initially, our dataset was randomly split into training and validation sets in a 7:3 ratio. We employed various modeling algorithms, including logistic regression, SVM, Random Forest, KNN, Decision Tree, Neural Network, and XGBoost, LightGBM, to construct combined models. The average accuracy of each model on the training set was assessed using five-fold cross-validation. Finally, the modeling algorithm, with highest cross-validation accuracy, was employed to construct the imaging combined model in this study. Variance inflation factor (VIF) analyses were conducted to assess correlation and multicollinearity among the final predictors.

Model evaluation

Discriminatory performance of model was assessed using receiver operating characteristic (ROC) analysis, yielding metrics such as area under curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) based on the Youden index-derived cut-off value. Box plots were employed to visualize the predictive performance of the ITHkinetic and PTHkinetic scores for IBC-EIC prediction. AUC comparisons between the Imaging Combined model with and without ITHkinetic and PTHkinetic score were conducted using the DeLong test. The incremental value of ITHkinetic and PTHkinetic score was complemented by net reclassification improvement (NRI) and integrated discrimination improvement (IDI) metrics. Calibration plots, Hosmer–Lemeshow test, and decision curve analysis (DCA) were used to assess the calibration and clinical utility of the models with and without the inclusion of the ITH and PTH score. Nomogram was constructed for personalized IBC-EIC prediction.

Patient selection for propensity score matching (PSM)

Patients in the IBC without EIC group were matched with those in the IBC-EIC group based on 15 covariates of clinicopathological characteristics, including age, menstruation status, gestation history, production factors, family history of breast cancer, history of other tumors, lateral involvement, pathology type, cell grade, lymph node status, molecular type, hormone receptor status (ER, PR), HER2 status, and Ki-67 expression.

Statistical analyses

Statistical analyses were performed using R software (version 4.2.3, http://www.r-project.org), Python (version 3.12.3, https://www.python.org), and the scikit-learn library (version 1.5.0). The R packages used in this study are listed in Supplementary Materials. Propensity score matching (PSM) at a 1:1 ratio was employed to address baseline and potential confounding clinicopathological differences between the IBC-EIC and IBC without EIC groups. The standardized mean difference was utilized to assess covariate balance before and after matching. The Mann–Whitney U-test or chi-square test was used to compare the differences in clinicopathological data between two groups. Cohen’s Kappa coefficient was utilized to evaluate inter- and intra-observer agreement of MRI features. A two-tailed p value less than 0.05 was considered statistically significant. The sample size was determined by the availability of patients during the retrospective study period.

Result

Clinicopathological characteristics

Finally, a total of 368 patients (mean age: 50 ± 10 [SD] years) remained, including 26.4% (97/368) IBC-EIC. Table 1 presents the clinicopathological characteristics of two groups before and after PSM. IBC-EIC displayed notable distinctions in baseline clinicopathological features, including a higher rate of family history of breast cancer (p = 0.03), a higher ratio of no special type IBC (p = 0.008), a lower ratio of high cell grade (p = 0.03), distinct molecular types (p < 0.001) and a higher rate of HER2 positivity (p < 0.001). After PSM, 97 matched pairs were successfully established, resulting in balanced clinicopathological covariates, with all standardized mean differences below 0.2, except for cell grade, which was only affected by incomplete pathological data.

Table 1 Clinicopathologic characteristics.
Full size table

MRI features differences

In univariate analysis, IBC-EIC showed higher odds of Non-Mass Enhancement (NME) (OR 15.483, CI 6.189–38.734, p < 0.001), peritumoral early NME presence (OR 4.759, CI 2.44–9.28, p < 0.001), non-circumscribed mass margins (OR 2.28, CI 1.268–4.098, p = 0.006), and increased long diameters (OR 2.08, CI 1.459–2.966, p < 0.001). Multivariate analysis identified NME (OR 9.074, CI 3.446–23.899, p < 0.001), the presence of peri-tumoral early NME (OR 3.074, CI 1.449–6.521, p = 0.003), and increased long diameters (OR 1.514, CI 1.001–2.292, p = 0.049) as significant predictors (Table 2). The Inter- and intra-observer agreement of MRI features were provided in Table S2.

Table 2 MRI features differences in the univariate and multivariate analysis.
Full size table

Personalized ITHkinetic and PTHkinetic score for IBC-EIC prediction

After PSM, the ITHkinetic score demonstrated a higher mean value in the IBC-EIC group (0.68 ± 0.23) compared to the IBC without EIC group (0.32 ± 0.25; p < 0.001) (Youden-cutoff value: 0.56). Similarly, the PTHkinetic score was higher in the IBC-EIC group (mean, 0.58 ± 0.19) compared to the IBC without EIC group (mean, 0.42 ± 0.18; p < 0.001) (Youden-cutoff value: 0.55) (Figs. 3, 4).

Figure 3
figure 3

Box plots depicting intra and peri-tumor heterogeneity (ITHkinetic and PTHkinetic) scores after propensity score matching (PSM), stratified by the true extensive intraductal component in invasive breast cancer (IBC-EIC) status. Mean ITHkinetic and PTHkinetic scores were significantly higher in IBC-EIC tumors (red) compared to IBC without EIC tumors (blue) after PSM (p < 0.001 for both comparisons). P values were calculated using the Mann–Whitney test. Horizontal lines in the boxes represent medians, with whiskers indicating minimum and maximum values. Filled circles denote outliers.

Full size image
Figure 4
figure 4

Axial fat-suppressed DCE MRI and color mapping illustrating kinetic intra- and peri-tumor heterogeneity (ITHkinetic and PTHkinetic) in extensive intraductal component in invasive breast cancer (IBC-EIC) (A–C) and pure invasive breast cancer (D–F). (A–C) In a 28 year-old woman, invasive breast cancer of no special type (HER2-enriched molecular type, HER2 positive) with 50% ductal carcinoma in situ (extensive intraductal component) was confirmed by surgical pathology. (A) Axial fat-suppressed DCE-MRI revealed non-mass enhancement (NME) with a tumor measuring 5.5 cm in length. The intra-tumor (B) and peri-tumor (C) regions were segmented and characterized using color mapping in DCE-MRI. Her personalized ITHkinetic score was 0.89 (Youden-cutoff value: 0.56), and her PTHkinetic score was 0.62 (Youden-cutoff value: 0.55), accurately classifying the tumor. (D–F) In another case, a 68 year-old woman was diagnosed with invasive breast cancer, no special type (luminal B molecular type, HER2 negative), without ductal carcinoma in situ. (D) Axial fat-suppressed DCE MRI revealed a mass measuring 1.7 cm in length in the left breast. The intra-tumor (E) and peri-tumor (F) regions were segmented and characterized using color mapping in the axial DCE MRI. Her personalized ITHkinetic score was 0.12, and her PTHkinetic score was 0.37.

Full size image

Before PSM, higher ITHkinetic and PTHkinetic scores were also observed in the IBC-EIC group compared to the IBC without EIC group (ITHkinetic: mean, 0.71 ± 0.20 vs. 0.49 ± 0.28, p < 0.001, Youden-cutoff value = 0.62; PTHkinetic: mean, 0.61 ± 0.18 vs. 0.50 ± 0.20, p < 0.001, Youden-cutoff value = 0.52) (Fig. S1).

Model development and evaluation

Logistic regression, with highest cross-validation accuracy (Table S3), was employed to construct the image combined model. The performance of image combined model in the training set and test set was visualized in Fig. S2. Finally, four predictive model (MRI model, ITHkinetic score, PTHkinetic score, and image combined model) were developed for IBC-EIC prediction. The performance of these models was presented in Table 3, with visualizations in Fig. 5 after PSM and Fig. S3 before PSM. The image combined models demonstrated good predictive performance for IBC-EIC prediction, achieving AUC values of 0.91 (95% CI 0.86–0.95) after PSM and 0.85 (95% CI 0.81–0.90) before PSM.

Table 3 Diagnostic performance of models for IBC-EIC prediction in this study.
Full size table
Figure 5
figure 5

Models’ performance for preoperative prediction of extensive intraductal component in invasive breast cancer (IBC-EIC) status after propensity score matching (PSM). The ROC curves of image combined models (red) yielded an AUC of 0.91 (95% CI 0.86–0.95) for preoperative IBC-EIC prediction. The AUCs for the MRI model (blue), kinetic intra-tumor heterogeneity score (green), and kinetic peri-tumor heterogeneity score (light green) were 0.80 (95% CI 0.74–0.86), 0.85 (95% CI 0.79–0.90), and 0.73 (95% CI 0.66–0.80), respectively.

Full size image

Calibration and clinical use

To facilitate clinical use, nomograms were created using predictors from the image combined model (Fig. 6, Fig. S4). The calibration curves of the nomograms yielded non-significant results (χ2 = 5.25, p = 0.73 after PSM; χ2 = 12.65, p = 0.12 before PSM) (Fig. S5). Decision curve analysis demonstrated that the combined models with ITHkinetic and PTHkinetic score provided greater net benefit in predicting IBC-EIC (Fig. S6).

Figure 6
figure 6

Nomogram based on the predictors in the image combined models after propensity score matching (PSM) for preoperative prediction of extensive intraductal component in invasive breast cancer (IBC-EIC). Each variable is assigned a specific point value indicated at the top of the graph, and the total points obtained by summing up the individual variable points are represented by a line connecting the total points to the bottom line. This line provides a preoperative estimate of the probability of IBC-EIC status.

Full size image

The adding value of ITHkinetic and PTHkinetic score for IBC-EIC prediction

The inclusion of the ITHkinetic and PTHkinetic score significantly improved IBC-EIC prediction, as evidenced by the Delong test (p < 0.001, 95% CI 0.05–0.16), Categorical NRI (0.39, 95% CI 0.24–0.55; p < 0.001), and IDI (0.21, 95% CI 0.15–0.27; p < 0.001) after PSM. These findings were consistent before PSM (Delong test: p = 0.003, 95% CI 0.02–0.08, Categorical NRI: 0.16, 95% CI 0.04–0.28; p < 0.001, and IDI: 0.09, 95% CI 0.06–0.12; p < 0.001).

Discussion

IBC-EIC is a common occurrence in clinical practice, with reported incidence rates ranging from 14.7% to 35%1,30. Accurate preoperative prediction of IBC-EIC is crucial for tailoring surgical resection margins and mitigating the risk of incomplete resection in breast cancer patients undergoing breast-conserving surgery (BCS). Our study characterized the kinetic intra-tumor and peri-tumor heterogeneity (ITHkinetic and PTHkinetic) of breast cancer, through quantitative analysis of clustering voxels exhibiting similar enhancement patterns in high-resolution ultrafast DCE-MRI. To enable a more accurate comparison of ITHkinetic and PTHkinetic differences, we utilized propensity score matching (PSM) to mitigate baseline and potential confounding clinicopathological differences between IBC-EIC and IBC without EIC groups. We found that IBC-EIC demonstrated a distinctive ITHkinetic and PTHkinetic profile, characterized by higher personalized ITHkinetic and PTHkinetic scores both before and after PSM compared to IBC without EIC. Additionally, disparities in MRI features between two groups are also identified. Leveraging these image predictors, our image combined model exhibited good predictive performance, achieving AUC values of 0.91 (after PSM) and 0.85 (before PSM). Importantly, the incorporation of the ITHkinetic and PTHkinetic scores significantly improved predictive performance, as evidenced by improved AUC based on the Delong test (p < 0.05), and further confirmed by statistical significance in categorical NRI (p < 0.001) and IDI (p < 0.001) analyses.

Our findings align with prior studies that have demonstrated the detectability of IBC-EIC through contrast-enhanced MRI1,9,23,30. The preliminary study identified early and overall peritumoral enhancement, the amount of fibroglandular tissue (FGT) around the MRI-visible tumor, and HER2 positivity as predictors of IBC-EIC. The predictive model based on these factors achieved an AUC of 0.791. In our earlier radiomic study, integrating intra-tumoral and peritumoral radiomic features from an early phase of ultrafast DCE-MRI demonstrated a moderate preoperative predictive ability for IBC-IC, with an AUC of 0.8223. Our recent model outperformed the aforementioned studies, highlighting the enhanced value of intra and peri-tumor kinetic heterogeneity, which was consistent with our hypothesis. Notably, our recent predictive model, solely based on preoperative imaging features, may offer superior clinical applicability. Given that the routine utilization of preoperative MRI still shows controversial impacts on the surgical outcomes of BCS, combining these preoperative image models with intra-operative margin assessment might offer an optimal strategy to guide BCS1, potentially enhancing its surgical outcomes in the future.

Our study demonstrated significant variations in ITHkinetic and PTHkinetic between the two groups, unraveling the intricate tumor heterogeneity as potential biomarkers for predicting IBC-EIC. Notably, after PSM, ITHkinetic (AUC 0.85) and PTHkinetic (AUC 0.73) scores demonstrated good diagnostic capabilities for IBC-EIC prediction, while there was a slight decrease in performance before PSM (ITHkinetic 0.73, PTHkinetic 0.64). This might be attributed to the confounding factors, specifically the differences in molecular subtypes and HER2 status before PSM, which may influence the characteristics of tumor heterogeneity. The heterogeneity among different molecular subtypes of breast cancer is widely acknowledged31. Recent studies also revealed significant HER2 heterogeneity in breast cancer32,33. The utilization of PSM allowed us to balance potential confounding clinicopathological covariates, facilitating a more rigorous comparison and identification of imaging disparities between the two groups.

In contrast, the MRI model exhibited robust performance before and after PSM (AUC 0.80 in both instances). The MRI features, such as NME, the presence of peri-tumoral early NME, and increased long diameters, were identified as independent predictors for IBC-EIC prediction in this study. Consistent with our findings, Van Goethem et al. also observed that IBC-EIC exhibited features like ductal or linear enhancement, long spicules, a regional enhancing area, or nodules adjacent to a mass30. Additionally, they also investigated the correlation of NME surrounding index tumor on MRI with pathological examination findings. The results showed that most NME surrounding a carcinoma corresponded to in situ or invasive extension of the carcinoma34. However, relying solely on conventional morphological features on MRI poses limitations in the era of individualized medicine. Visual evaluation of these MRI features introduces potential subjective variability and may be influence by the radiologist’s experience23. The integration of conventional MRI features with personalized tumor heterogeneity scores enhances the comprehensiveness and individualized MRI-based approach for preoperative IBC-EIC prediction.

This study had some limitations. Firstly, as a retrospective study, inherent variations and selection biases may be present. Therefore, a well-designed prospective study is necessary to validate these findings. Secondly, visual evaluation of MRI features may introduce potential variability. We addressed this concern by conducting assessments of both inter- and intra-observer agreement, revealing satisfactory agreement. Lastly, the models were developed using novelty high-resolution ultrafast DCE-MRI protocols on a single MRI scanner within a single center. Consequently, the impact of different MRI scanners on the variance was not investigated. To enhance the generalizability of our findings, further studies involving diverse datasets from various scanners and centers are essential.

In conclusion, our ITHkinetic and PTHkinetic-based model offer a non-invasive and more accurate preoperative prediction of IBC-EIC, providing personalized guidance for patients undergoing BCS.