A clinical radiomics nomogram preoperatively to predict ductal carcinoma in situ with microinvasion in women with biopsy-confirmed ductal carcinoma in situ: a preliminary study

Purpose To predict ductal carcinoma in situ with microinvasion (DCISMI) based on clinicopathologic, conventional breast magnetic resonance imaging (MRI), and dynamic contrast enhanced MRI (DCE-MRI) radiomics signatures in women with biopsy-confirmed ductal carcinoma in situ (DCIS). Methods Eighty-six women with eighty-seven biopsy-proven DCIS who underwent preoperative MRI and underwent surgery were retrospectively identified. Clinicopathologic, conventional MRI, DCE-MRI radiomics, combine (based on conventional MRI and DCE-MRI radiomics), traditional (based on clinicopathologic and conventional MRI) and mixed (based on clinicopathologic, conventional MRI and DCE-MRI radiomics) models were constructed by logistic regression (LR) with a 3-fold cross-validation, all evaluated using receiver operating characteristic (ROC) curve analysis. A clinical radiomics nomogram was then built by incorporating the Radiomics score, significant clinicopathologic and conventional MRI features of mixed model. Results The area under the curves (AUCs) of clinicopathologic, conventional MRI, DCE-MRI radiomics, traditional, combine, and mixed model were 0.76 (95% confidence interval [CI] 0.59–0.94), 0.77 (95%CI 0.59–0.95), 0.74 (95%CI 0.55–0.93), 0.87 (95%CI 0.73–1), 0.8 (95%CI 0.63–0.96), and 0.93 (95%CI 0.84–1) in the validation cohort, respectively. The clinical radiomics nomogram based on mixed model showed higher AUCs than both clinicopathologic and DCE-MRI radiomics models in training/test (all P < 0.05) set and showed the greatest overall net benefit for upstaging according to decision curve analysis (DCA). Conclusion A nomogram constructed by combining clinicopathologic, conventional MRI features and DCE-MRI radiomics signatures may be useful in predicting DCISMI from DICS preoperatively. Supplementary Information The online version contains supplementary material available at 10.1186/s12880-023-01092-5.


Background
Breast cancer is one of the most common malignant tumors in women worldwide [1].With medical advances, the number of patients with ductal cancer in situ (DCIS) and DCIS with microinvasion (DCISMI) is increasing [2].Malignant epithelial cell growth inside the mammary duct lumen but no penetration beyond the basement membrane is the histological hallmark of DCIS [3].DCISMI is thought to be the transitional stage between DCIS and invasive ductal cancer (IDC) [4].In this study, we used the AJCC's definition of "microinvasion" as one foci (DCISMI-one) or more foci (DCISMI-more) of invasive carcinoma ≤ 1 mm in diameter inside an area of DCIS [5].Numerous studies indicate that DCISMI's prognosis and natural history are quite similar to those of DCIS [6][7][8][9][10].However, a large-scale clinical study [11] revealed that the prognosis of DCISMI is more like that of small invasive carcinoma than DCIS.
In comparison to mammography (MG) and ultrasonography (US), dynamic contrast enhanced MRI (DCE-MRI), a common procedure for breast MRI, offers the best sensitivity for identifying DCIS or DCIS coupled with invasive cancer [12].It is critical to "identify MRI features that can be combined with clinical and biological characteristics to better stratify risk in patients with DCIS" [13].Clinically, the majority of DCIS and DCISMI exhibit a clinically comparable morphological appearance.There have been a number of publications on characteristics helping to predict microinvasion and invasion, including clinical findings and findings on conventional imaging but with inconsistent results.However, there have been only a few publications that have compared MRI findings between pure DCIS and DCISMI cases [14,15].To the best of our knowledge, no previous report has researched the predictors of DCISMI using radiomicsbased machine learning algorithms.A radiomics-based signature could provide a more thorough approach by combining both geographical and temporal data to define the tumor more fully.Thus, we speculate that radiomics characteristics obtained from DCE-MRI may represent cellular and molecular data and may be able to foretell upstaging in females with biopsy-proven DCIS.
In this study, we aimed to compare the performance of clinical-pathological characteristics, conventional breast MRI features, DCE-MRI radiomics signatures, and combined multiple features in predicting DCISMI and to construct a nomogram to better understand the risk factors.

Patients
The institutional ethics committee approved this retrospective study and granted a waiver of informed consent.
We included 246 women who underwent preoperative breast MRI and US guided core needle biopsy (US-CNB) after MRI preoperatively with primary breast DCIS from January 2015 to August 2022.US-CNB was performed with a 14-guage automated biopsy gun (Stericut; TSK Laboratory, Tochigi, Japan) with five samples obtained from each lesion.The exclusion criteria were ① those who had IDC or were associated with another disease (n = 59); ② those who had an inadequate MRI protocol or poor image quality (n = 9); ③ those who had received therapy prior to the MRI (n = 36); ④ those whose tumors had unclear images (n = 13); ⑤ those who had not undergone surgery in our hospital (n = 25); ⑥ those who lacked clinicopathologic data (n = 18).This implies that the lesions were all US and MRI visible which would bias the series.Finally, the research cohort consisted of 86 individuals (mean age 44.30 ± 9.30 years) with 87 lesions (Fig. 1).One patient had DCIS that was verified by biopsy in both breasts.

MRI Examinations
Two 3.0 T scanning systems (GE Discovery 750W and MAGNETOM Skyra, Siemens Healthcare), each with an eight-channel breast-specific coil, were used for all breast MRI scans.Women were in the prone position.The imaging protocol included a T2-weighted short tau inversion recovery turbo spin-echo pulse sequence (T2WI-FS) (repetition time [TR]/echo time [TE] = 4160/85 ms; matrix size = 512 × 512; field of view [FOV] = 350 × 350 mm 2 , section thickness = 5 mm for the GE 750W scanner; TR/TE = 3600/53 ms, matrix size = 320 × 320, FOV = 340 × 340 mm 2 , section thickness = 4 mm for the Simens Skyra scanner) and a DCE-MRI (TR/ TE = 4.32/2.10;matrix size = 512 × 512; FOV = 350 × 350 mm 2 , section thickness = 0.7 mm for the GE 750W scanner; TR/TE = 4.49/1.68ms, matrix size = 320 × 320, FOV = 340 × 340 mm 2 , section thickness = 1.2 mm for the Simens Skyra scanner).Gadolinium-DTPA (0.5 mmol/ mL, Bayer) was power-injected at a dose of 0.1 mmol/kg body weight and a flow rate of 2.6 mL/s, followed by a 15 mL saline flush.The two scanners employed the same contrast agent and scanning mode.To balance the sample sizes on the two MRI protocols and to avoid the effect of excessive sample size bias on the results, we randomly selected the image data of above 86 patients with the same DICS and DCISMI component ratio.

Data collection
Retrospective data retrieval from the hospital information system (HIS) database and the picture archiving and communication system (PACS) were performed on clinical-pathological and imaging data.Clinical-pathologic characteristics of the DCIS and DCISMI groups were shown in Table 1.The clinical-pathological including age, body mass intensity (BMI), DCIS grades of the biopsy specimen (low, intermediate, or high), the status of the estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2), Ki-67 proliferation, p63 and axillary lymph node metastasis.HER2 positivity was defined as a protein overexpression score of 3 + determined by immunohistochemistry or the presence of gene amplification (positive in situ hybridization) [16].Ki-67 scores ≥ 20% were considered high [9].Conventional MRI features were included in Table 2, including lesion side, lesion number, lesion size on DCE-MRI scan ( ≫ 2cm, < 2cm ), and morphologic feature at T2WI-FS and DCE-MRI according to the Breast Imaging Reporting and Data System MRI lexicon.Intra tumoral high signal intensity (SI) [17] and peritumoral edema [18] were evaluated on T2WI-FS in this study.Reviewers were blinded to the clinicopathologic results.
Medical picture registration was done using ITK-SNAP programmed software (version 3.4.0;http:// www.itksn ap.org).In the phase of the contrast-enhanced acquisition, which clearly displayed the lesion, two radiologists with five years of experience delineated the tumor region of interest (ROI) along the edge of the lesion in each layer, and then duplicated the ROIs for the remaining four acquisitions.The third radiologist with 10 years of experience reexamined and validated the final boundary after the discussion when there was a significant discrepancy between the two radiologists.This method produced accurate feature extraction and accurate tumor outlines.We chose the sets of all the lesions completed by the two radiologists (5 years of work experience) to assess the repeatability of radiomics features.The interobserver reproducibility of feature extraction was assessed by the intra-class correlation coefficient (ICC).ICC ≥ 0.75 indicated high consistency and selected for further investigation.

Establishment models
Firstly, selectkbest (P < 0.05) was used to select the optimal predictive features to establish the clinicopathologic and conventional MRI models by the logistic regression (LR) classifier based on clinicopathologic characteristics and conventional MRI features, respectively.
PyRadiomics (version 3.0.1;http:// github.com/ Radio mics/ pyrad iomics) was used to extract radiomics features of DCE-MRI.As the images were derived from two MRI scanners with different parameters, normalization was performed before features could be extracted from the ROIs of the DCE-MRI images.To facilitate consistent feature extraction, the image data were normalized and preprocessed in the following steps: spatial resampling to 1 × 1 × 1 mm 3 and intensity discretization to a fixed bin width of 25.Radiomics features can be calculated on the pre-processed images using the wavelet and Laplacian of Gaussian (LoG) flters with varying sigma (= 1.0, 2.0, 3.0, 4.0, 5.0).To reduce overfitting or selection bias, Select Percentile (P < 0.05 ) and least absolute shrinkage selec- tion operator (LASSO) were used to further select the optimal predictive features to establish the DCE-MRI radiomics model.Shape-based, first-order statistical, gray-level cooccurrence matrix (GLCM), gray-level region matrix (GLSZM), gray-level run-length matrix (GLRLM), and gray-level dependence matrix (GLDM) were extracted from original and filtered images for a total of 7046 features.The workflow is presented in Fig. 2. The clinicopathologic, conventional MRI, DCE-MRI radiomics, combine (including conventional MRI and DCE-MRI radiomics), traditional (including clinicopathologic and conventional MRI) and mixed (including clinicopathologic, conventional MRI and DCE-MRI radiomics) models were constructed by the LR classifier with a threefold cross-validation, to ensure that it was not affected by insufficient sample size.Meanwhile, we also analyzed whether the DCE-MRI radiomics model could distinguish DCISMI-more from DCISMI-one.

Statistical analysis
Clinicopathologic and conventional MRI features were compared between the DCIS and DCISMI groups using the Chi-square or Fisher's exact test for categorical variables and the analysis of variance (ANOVA) or Kruskal-Wallis test for continuous variables.
All statistical analyses were performed with SPSS software version 20.0 (SPSS, Chicago, IL, USA).A twosided P < 0.05 was considered statistically significant.To evaluate the predictive ability of different models, the area under the curve (AUC) of the receiver operating curve (ROC), sensitivity (SEN), specificity (SPE) with 95% confidence interval (CI), accuracy (ACC), positive predictive value (PPV), and negative predictive value (NPV) were calculated together by using a given cutoff of the predicted probability of DCISMI and distinguished DCISMI-more from DCISMI-one by the DCE-MRI radiomics model as well.AUCs between different models were compared using DeLong's test.Calibration curves were employed to assess the predictive performance of each model.To evaluate each model's clinical applicability, decision curve analysis (DCA) was used.

Clinicopathologic and conventional MRI features assessment
Among the 87 biopsy-confirmed DCIS lesions, 42 (48.28%)were found to be pure DCIS in the final surgical pathology, and 45 (51.72%) were upgraded to DCISMI.The mean age at diagnosis and BMI did not significantly differ between the two groups.Histologically, higher nuclear grade (P = 0.0006), negative p63 (P = 0.0004), negative ER (P = 0.0191), and negative PR (P = 0.0074) were more significantly common in DCISMI group.The DCISMI group tended to show higher positive HER2 than the DCIS group; however, this difference did not reach statistical significance (P = 0.0878).The clinicopathologic features of these samples are summarized in Table 1.Table 2 presents the conventional MRI features of the two groups.The DCISMI group tended to show more peritumoral edema (P = 0.0022) and heterogeneous enhancement (P < 0.0001) frequent than the DCIS group.There were no significant differences between groups aside from above the two features of conventional MRI.
Among all the clinicopathologic and conventional MRI features, variables were selected by the selectkbest to establish in subsequent machine learning by the LR classifier (Table 3).The results revealed that

Model comparisons and establishment of a preoperative nomogram
DeLong's test showed that the traditional model achieved higher AUCs than the clinicopathologic model in both the training and test sets (all P < 0.05).The mixed model showed better AUCs than both the clinicopathologic and DCE-MRI radiomics models in both the training and test sets as well (all P < 0.05).A comparison of the six models is shown as a heat map in Fig. 4a.
The calibration curves of the models are shown in Fig. 4b, which shows good calibration.DCA (Fig. 4c) illustrated that the mixed model showed the greatest overall net benefit for upstage and the second was the combine model within reasonable threshold probabilities.The DCE-MRI radiomics model, which showed all but the same net benefit as the traditional model, showed better than the conventional MRI model.To provide a visualized outcome measure, a preoperative nomogram figure was plotted based on training cohort by combining the p63, nuclear grade, peritumoral edema on T2WI, heterogeneous enhancement pattern and Radiomics score in Fig. 5.

Discussion
Our preliminary analysis showed that the preoperative clinicopathologic, conventional breast MRI and DCE-MRI radiomics features could predict the preoperative histological upstage of DCIS and that DCE-MRI radiomics signatures may be different for DCISMI-more from DCISMI-one.The mixed model showed excellent predictive performance.The DCE-MRI radiomics model, which could be replaced by the traditional model, showed better than the conventional MRI model.A nomogram constructed by combining clinicopathologic, conventional MRI features and DCE-MRI radiomics signatures may be useful in predicting DCISMI from DICS, preoperatively.DCISMI represents an intermediate state between DCIS and IDC [19].Although DCIS, DCISMI, and T1a IDC (invasive tumor size > 0.1 cm but ≤ 0.5 cm in great- est dimension was classified as T1a) all have generally excellent prognosis, some studies revealed that DCISMI more closely resembles small invasive carcinoma than pure DCIS, and many practitioners are treating it as such [11,20].In comparison to women with pure DCIS or DCISMI-one, DCISMI-more is linked to a higher probability of invasive local recurrence in women with DCIS treated with breast conserving surgery [21].The rate of axillary lymph node metastasis is approximately 0% to 20% in DCISMI [19].According to approved treatment guidelines, sentinel lymph node biopsy was used in conjunction with mastectomy because of these relatively low but clinically significant rates [22].A change in diagnosis from pure DCIS before surgery to DCISMI after surgery creates a great deal of patient anxiety and possibly a second surgery [23].Therefore, an accurate prediction of the histological DCISMI and even DICSMI-more could help with preoperative risk stratification and the best choice of patients who could benefit from more extensive surgery while avoiding overtreatment of patients at low risk.
In this study, higher nuclear grade as the independent factor associated with histologic upgrade in the clinicopathologic model and both the peritumoral edema and heterogeneous enhancement as significant independent factors in the conventional MRI model, which were partly consistent with those of previous studies for predicting invasive breast cancer [15,[24][25][26][27]. Several prior studies have described the MRI features [14,15,[28][29][30][31][32] that can help predict the invasive component of a biopsyproven DCIS.However, there have been few reports that have compared MRI findings between pure DCIS and DCISMI, rarely by the features using machine learning algorithms based on the clinicopathologic, conventional breast MRI, and DCE-MRI radiomics signatures.In this study, the mixed model established by the LR classifier showed that heterogeneous enhancement pattern and Radiomics score were independent predictors of upstage.DCISMI lesions showed more heterogeneous enhancement, which is partly consistent with previous studies [15,28,33].The main applications of breast radiomics investigations include the molecular categorization, lymph node metastases, and molecular markers of IDC and DCIS prediction [34].There is only one report based on US radiomics to predict the molecular biomarkers of DICS [35].Radiomics is a precision medical method for non-invasive diagnosis, evaluation of efficacy, and biological behavior [36].Contrary to our prior hypothesis, it was found that the DCE-MRI radiomics model's AUC (AUC = 0.74) was the lowest one below 0.8.This may be because the MRI radiomics mostly depend on a set of MRI sequences [35].In our study, only DCE-MRI radiomics features were included rather than diffusion weighted imaging (DWI) or apparent diffusion coefficient (ADC) radiomics features.DWI may provide more accurate tumor microenvironment monitoring [37], which has been widely explored to differentiate benign from malignant breast lesions [38,39], and showed its diagnostic ability in DCISMI [33] and Our research revealed that, with a histologic upgrade rate somewhat higher than that of other studies with 8.8%-51.5% of DCIS patients upstaged to invasive disease [15,41], 51.72% of lesions having a preoperative diagnosis of DCIS were histopathologically upgraded to DCISMI.There might be a few reasons.Firstly, the study's sample size was small.Secondly, patients with DCIS histopathologically upgraded to IDC were excluded from our analysis according to the study's objectives.Thirdly, the lesions were all US and MRI visible which would bias the series toward a higher rate of microinvasive cancer.To date, few features have consistently surfaced as strong predictors of upstaging on excision.
We discovered that the mixed model performed exceptionally well for preoperatively predicting the histological upstage of DCISMI.In our investigation, they were combined into a unique clinicopathologic This study had several limitations that should be noted.First, this was a retrospective study with a relatively small Fig. 5 The nomogram combining nuclear grade, p63, peritumoral edema, heterogeneous enhancement pattern and radiomics scores.The clinical radiomics nomogram was developed for the prediction of DCISMI with nuclear grade, p63, peritumoral edema on T2WI, heterogeneous enhancement pattern, and radiomics scores.Abbreviations: T2WI, T2-weighted image; DCISMI, ductal carcinoma in situ with microinvasive sample size at a single center.Second, two different breast MRI protocols (3.0 T Siemens and GE) were used at our hospital during the study period.At the same time, this may reflect the stability of the models in our study.Third, the DWI results were not examined, which would have revealed more data.Fourth, it was difficult to draw the margin of some DCIS lesions with non-mass enhancement.Finally, we did not perform an external validation test using an independent data set, although a 3-fold cross-validation was used.To verify our nomogram, additional external validation utilizing several independent data sets would be required.

Conclusion
Our preoperative nomogram model specifically for DCISMI patients with clinicopathologic, conventional MRI, and DCE-MRI radiomics signatures enabled a more accurate prediction of upstaging in women with biopsyproven DCIS.This could help to select women who were indicated for sentinel lymph node biopsy at initial surgery, thus avoiding unnecessary axillary surgery and preventing delayed secondary surgery.Although validation requires a larger sample size, DCE-MRI radiomics may discriminate between DCISMI-more and DCISMI-one.

Fig. 3
Fig. 3 The receiver operating characteristic curves (ROC) of the six models in the upstaging of DICS. a The ROC curves of clinicopathologic, conventional MRI, DCE-MRI radiomics, combine, traditional, and mixed models based on LR algorithm in the training cohort.b The ROC curves of clinicopathologic, conventional MRI, DCE-MRI radiomics, combine, traditional, and mixed models based on LR algorithm in the test cohort.c The ROC curve of the DCE-MRI radiomics model in distinguishing DCISMI-more from DCISMI-one in the training cohort.d The ROC curve of the DCE-MRI radiomics model in distinguishing DCISMI-more from DCISMI-one in the test cohort.Abbreviations: ROC, receiver operating characteristic curves; DCIS, ductal carcinoma in situ; MRI, magnetic resonance imaging; DCE-MRI, dynamic contrast enhanced MRI; LR, logistic regression; CI, confidence interval; AUC, area under the curve; DCISMI-more, DCIS with multifocal of microinvasive carcinoma; DCISMI-one, DCIS with one focus of microinvasive carcinoma

Fig. 4
Fig. 4 The heatmap, calibration curves and DCA of the six models.a Heatmap comparison of the clinicopathologic, conventional MRI, DCE-MRI radiomics, combine, traditional, and mixed models.b Calibration curves for the clinicopathologic, conventional MRI, DCE-MRI radiomics, combine, traditional, and mixed models based on the LR algorithm.It is the curve with the model-predicted probability of DCISMI as the X-axis and the actual rate acquired by the bootstrapping method as the Y-axis.The degree of agreement between the depicted calibration curve and the 45° straight line reflects the predictive performance of each model.c The DCA for the clinicopathologic, conventional MRI, DCE-MRI radiomics, combine, traditional, and mixed models based on the LR algorithm.The Y-axis represents the net benefit.DCA showed that in six models within reasonable threshold probabilities, the mixed model showed the greatest overall net benefit for upstage and the second was the combine model.The DCE-MRI radiomics model, which showed all but the same net benefit as the traditional model, showed better than the conventional MRI model.The combine model added more net benefit than the traditional model at the range of 0.4 ~ 1.0.The clinicopathologic model added more net benefit than the conventional MRI model and DCE-MRI radiomics model from 0.65 to 1.0 and from 0.7 to 1.0, respectively.Abbreviations: DCA, decision curve analysis; MRI, magnetic resonance imaging; DCE-MRI: dynamic contrast enhanced MRI; LR, logistic regression + conventional MRI + radiomics nomogram, which demonstrated sufficient prediction performance.The clinicopathologic prediction model reflected the role of clinical and pathological baseline information in upstage prediction.While the radiomics model based on DCE-MRI included quantification of pictures, the conventional MRI model based on T2WI-FS and DCE-MRI involved qualitative assessment of images.The mixed model nomogram has the potential to increase diagnostic effectiveness and net benefit over the whole spectrum of threshold probabilities in DCA, in addition to displaying and customizing the likelihood of DCISMI for each patient.

Table 1
Clinicopathologic characteristics of DCIS in comparison with DCISMI in the study cohort Data values indicate the number of patients (with percentages in parentheses), or the mean ± standard deviation Abbreviations: DCIS Ductal carcinoma in situ, DCISMI Ductal carcinoma in situ with microinvasive, BMI Body Mass Index, ER Estrogen receptor, PR Progesterone receptor, HER2 Human epidermal growth factor receptor 2

Table 2
Conventional MRI characteristics in patients with DCIS in comparison with DCISMI Data indicate the number of lesions (with percentages in parentheses) or the mean ± standard deviation Abbreviations: DCIS Ductal carcinoma in situ, DCISMI Ductal carcinoma in situ with microinvasive, SI Signal intensity, T2WI T2-weighted image, NME Non-mass enhancement, NAC Nipple-areolar complex a Heterogeneous enhancement pattern includes heterogeneous, clumped, and clustered ring pattern

Table 3
The features selection methods and their corresponding parameters in the models

Table 4
Predictive performances of the six models and DCE-MRI radiomics model in distinguishing DCISMI-more from DICSMI-one Abbreviations: DCISMI Ductal carcinoma in situ with microinvasive, DCISMI-more DCIS with multifocal of microinvasive carcinoma, DCISMI-one DCIS with one focus of microinvasive carcinoma, AUC Area under the curve, SEN Sensitivity, SPE Specificity, ACC accuracy, PPV Positive predictive value, NPV Negative predictive value, CI Confidence interval, DCE-MRI Dynamic-contrast enhanced MRI a Combine model was constructed based on conventional MRI and DCE-MRI radiomics features b Traditional model was constructed based on clinicopathologic and conventional MRI features c Mixed model was constructed based on clinicopathologic, conventional MRI and DCE-MRI radiomics features