Multimodal Machine Learning-Based Ductal Carcinoma in situ Prediction from Breast Fibromatosis

Introduction

Breast cancer is the most frequent malignant tumor in females worldwide with over half of these cases occurring in developed countries.¹ In the United States and Europe, breast cancer is one of the leading causes of cancer-related mortality.² Besides, breast cancer is diagnosed in 12.2% of women in China.³ With medical advances, the calcifications and breast imaging reporting and data system (BI-RADS) lexicon have been identified which represent variable underlying histologic conditions related to ductal carcinoma in situ (DCIS) or invasive breast cancer.⁴ It was estimated that about 51,400 females would have DCIS in 2022 according to the American Cancer Society.⁵ At stage 0 of breast cancer, DCIS is a precursor to invasive ductal carcinoma and features a neoplastic proliferation of epithelial cells surrounded by myoepithelial cells. Myoepithelial cells are confined by an intact basement membrane that separates these cells from the breast stroma and prevents tumor cells from metastasizing.⁶ Early detection can prevent DCIS from progressing into a more invasive cancer.⁶ Breast fibromatosis is a locally invasive lesion with no metastatic potential caused by fibroblasts or myofibroblasts that may occur within the breast parenchyma or originate from the thoracic fascia and extend into the breast.⁷ Therefore, it is necessary to investigate a novel method to identify DCIS or breast fibromatosis and give corresponding treatment to improve the diagnosis and prognosis of patients.

Ultrasound features including infiltrative margins, microcalcifications, and hypoechoic have been reported to be related to higher suspicion of malignancy.⁸ To achieve a more quantitative prediction method, radiomics has been introduced into many studies. Radiomics is involved in converting digital images into high-dimensional mineable features using high-throughput extraction analysis to explore the underlying tissue information of the images that are hard to recognize by the human eye.⁹ Radiomics performs more satisfactorily in predicting multi-parameter imaging features than traditional imaging indicators. These extracted features can be used to build efficient models by incorporating with genomic, histologic, laboratory, and clinical data for prediction, histological classification or classification of benign and malignant lesions, lymph node metastasis detection, and diagnosis by using machine learning (ML) methods.^10–13 In addition, the predictive value of radiomics for treatment selection or treatment response and clinical prognosis has been demonstrated, which is closely associated with proteomic, transcriptomic, and genomic characteristics.^14,15 However, few studies developed ML models to distinguish DCIS from breast fibromatosis.

ML is an emerging artificial intelligence tool, critical in improving prediction accuracy of the diagnosis and prognosis.¹⁶ Improved medical database management yielded a potential use of ML in medicine.¹⁷ Compared to other statistical methods, ML algorithms allow interactions between variables, identify key predictors, find optimal algorithms between the study outcomes and potential predictors by learning from dataset patterns, and show greater accuracy in clinical settings.¹⁸ Our study incorporated the clinical data, ultrasonic characteristics, and radiomics features to build related models. The predictive ability and clinical efficacy of the models were compared to select an optimal model for predicting DCIS.

Materials and Methods

Patients and Data Collection

Patients diagnosed with DCIS (n=198) via surgery and those diagnosed with breast fibromatosis (n=108) via B-mode ultrasound from Ningbo Medical Centre Lihuili Hospital between January 2016 and June 2023 were enrolled in the study. The 306 participants were randomly divided into the development cohort and internal validation cohort at a ratio of 6:4 (184 in the development cohort and 122 in the internal validation cohort).

Inclusion criteria: (1) Patients were confirmed as DCIS by postoperative pathology, and immunohistochemistry test results were obtained or those diagnosed with breast fibromatosis by B-mode ultrasound; (2) Breast ultrasound was performed within 1 week prior to surgery and the image quality was good; (3) There were no other primary tumors.

Exclusion criteria: (1) Poor image quality; (2) Patients who have previously undergone breast puncture, surgery, radiation, chemotherapy, or hormone therapy.

The clinical and B-ultrasonic features of patients were collected: age, tumor size, tumor location, morphology, calcification, blood flow, BI-RADS grade, orientation, margin, echo mode, and rear echo signature. The multivariate logistic regression analysis was utilized to select the independent clinical and B-ultrasonic features as a clinical-ultrasonic model to predict DCIS. The value of the clinical-ultrasonic model in predicting DCIS was analyzed by receiver operating characteristic (ROC) curve analysis in the training cohort, test cohort, and validation cohort.

Under the Declaration of Helsinki, this study involving human participants was reviewed and approved by the Ethics Committee of Ningbo Medical Centre Lihuili Hospital (Approval NO.KY2022PJ139). Written informed consent for participation was not required for this study following the national legislation and the institutional requirements.

Image Acquisition and Segmentation

The ultrasound doctor performed a routine two-dimensional ultrasound examination on both mammary glands of the patient, identified the breast lesions, recorded the location, shape boundary, and blood flow, marked the lesions on the body surface, and scanned the bilateral axillary lymph nodes. The diagnostic report was obtained by referring to the BI-RADS classification standard (5th edition): PHILIPS IU22, Q5, MINDRAY DC-8, APLIO5000, probe frequency 3–12 MHz.

The sonographer selected the largest plane of each breast lesion without knowing the histopathological information and delineated a region of interest (ROI) covering the entire lesion with the ITK-SNAP software (open-source software; http://www.itk-snap.org).

Radiomics Feature Extraction and Evaluation

The ultrasound radiomics features including first-order, gray-level co-occurrence matrix (glcm), gray-level dependence matrix (gldm), gray-level run length matrix (glrlm), gray-level size zone matrix (glszm), and neighbourhood gray-tone difference matrix (ngtdm) features were extracted using the “pyradimoics” package in the 3D slicer software (version 5.4.0) by the intraobserver and interobserver. The interclass correlation coefficient (ICC) was employed to evaluate the consistency of the extracted radiomics features. Features with ICC >0.70 demonstrated a good consistency of these characteristics and were used for subsequent feature selection.

Radiomics Feature Selection

All the radiomics features were performed with z-score normalization before screening features. Then, the least absolute shrinkage and selection operator (LASSO) method was adopted to select the important radiomics features. Next, collinearity analysis was conducted and features with a variance inflation factor (VIF) value greater than 6 indicating high collinearity were removed. Subsequently, the top five features were identified using XGBoost and random forest (RF) methods, respectively. The intersection of the two methods was considered as the vital features and used for the following radiomics model construction.

Radiomics Model Development and Performance

In the development cohort, 147 (80%) patients were randomly selected to train models and tune hyperparameters, and 37 (20%) patients to test the models. We used six machine learning algorithms to develop the models: XGBoost, logistic regression, RF, support vector machines (SVM), K Nearest Neighbors (KNN), and multilayer perceptron (MLP). Resampling was utilized to address class imbalance. The area under the ROC curve (AUC), the area under the precision-recall curve (AP), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 scores, and Kappa value served as evaluation metrics to compare the performance of six algorithms in the train and test sets. AUC was used to obtain the optimal model and to assess the consistency of the models in the training and test sets. The classifier with the best performance in the two sets was regarded as the radiomics model and the predicted probability was considered as Radscore (RD). The ROC was used to explore the predictive performance of the model in distinguishing DCIS from fibroma patients in the training cohort, test cohort, and validation cohort.

Clinical-RD Model Development

Factors in the clinical-ultrasonic model as well as the RD were enrolled into multivariate logistic regression analysis to select the independent predictors for clinical-RD construction. Then, the ROC was used to compare the predictive value of the clinical-ultrasonic model, the RD model, and the clinical-RD model for DCIS. Besides, decision curve analysis (DCA) was employed to evaluate the clinical usefulness of the three models in the development cohort and validation cohort, respectively.

Statistical Analysis

All statistical analyses were performed using SPSS software (version 23.0) and R software (version 4.4.2). Categorical variables were expressed as count (percent) and the two-group differences were compared by Chi-square test or Fisher’s exact test. After performing the Shapiro–Wilk test, all the continuous variables did not meet the normal distribution and were presented as median (quartiles). The group differences were analyzed by the Mann–Whitney-U test. P <0.05 indicated a significance level.

Results

Patient Characteristics and Clinical-Ultrasonic Model Construction

Patient characteristics in the development and validation cohorts are summarized in Table 1. All the variables were not significant between the two cohorts. Table 2 demonstrates the clinical characteristics and B-ultrasonic features in the development cohort. Patients had a median age of 37 and 52 years in the fibroma and DCIS groups, respectively (P <0.05). In addition, tumor size is significantly larger in the DCIS group (P <0.05). The lesions in the DCIS group tended to exhibit irregularity, dotted calcification, higher blood flow grade, and BI-RADS grade (P <0.05). The edges were more likely to be angular and fuzzy in the DCIS group (P <0.05). The uneven echo pattern took a proportion of 10.45% and 23.93% in the fibroma and DCIS groups, respectively (P <0.05). The other factors including tumor location, orientation, and rear echo signature had no remarkably different distributions between the two groups (P >0.05).

Table 1 Baseline Characteristics of Study Cohorts

Table 2 Clinical Characteristics and B-Ultrasonic Features Between Fibroma and Carcinoma in situ Groups

Following this, the identified significant factors were included in multivariate logistic regression analysis. Morphology-quasi-circle, morphology-irregularity, calcification-thick, and diffuse, edge-leaflet, edge-angular and fuzzy, edge-cyst, and tumor size were not significantly related to DCIS (P >0.05), while calcification-dotted, BI-RADS grade, and age were independent factors for predicting DCIS (P <0.05) (Table 3). Thus, the three significant predictors were used for constructing a clinical-ultrasonic model. The AUCs for the clinical-ultrasonic model in predicting DCIS were 0.943, 0.938, and 0.907 respectively in the training, test, and validation cohorts (Figure 1).

Table 3 Association of Clinical-Ultrasonic Features with Carcinoma in situ

Figure 1 The value of clinical-ultrasonic model in predicting carcinoma in situ. Abbreviations: ROC, receiver operating characteristic; AUC, the area under the curve.

Selection of Radiomics Features and Radiomics Model Construction

After LASSO screening, totally 14 features with non-zero coefficients were obtained (Figure 2A). The detailed information of these features is shown in Table 4. Further collinearity analysis showed that the VIF values of 10Percentile and interquartileRange were greater than 6, which were deleted for the following analysis (Table 5). Among the remaining 12 features, XGBoost and RF algorithm were respectively used to screen the top five important features (Figure 2B). Complexity and HighGrayLevelZoneEmphasis were the overlapped features of the two algorithms, which were used for constructing the radiomics model and its predicted probability called the RD.

Table 4 Detailed Information of Selected Features with Non-Zero Coefficients

Table 5 Collinearity Analysis

Figure 2 Selection of radiomics features. (A) The radiomics features screened by LASSO. (B) The top five radiomics features were ranked by XGBoost and random forest algorithm.

The AUCs for XGBoost, logistic regression, RF, SVM, KNN, and MLP were 1.000, 0.976, 1.000, 0.989, 0.993, and 0.838, respectively in the training set and their AUCs were 0.961, 0.991, 0.986, 0.952, 0.974, and 0.867, respectively in the test set (Figure 3A). The APs for XGBoost, logistic regression, RF, SVM, KNN, and MLP were 1.000, 0.986, 1.000, 0.993, 0.994, and 0.860, respectively in the training set. The APs were 0.974, 0.997, 0.993, 0.971, 0.979, and 0.931, respectively in the test set (Figure 3B). The Results of other evaluation indexes for six machine learning methods are exhibited in Table 6. XGBoost was most likely to be susceptible to overfitting, whereas logistic regression was likely to be relatively stable Thus, logistic regression was chosen as the final optimal model, which still owned satisfactory performance in the validation cohort (AUC=0.947) (Figure S1).

Table 6 Predictive Performance of Six Machining Learning Classifiers in the Training and Test Sets

Figure 3 Predictive performance of six machine learning models. (A) Receiver operating characteristic curve analysis of the six classifiers in the training and test sets. (B) Precision curve analysis of the six classifiers in the training and test sets. Abbreviations: ROC, receiver operating characteristic; AUC, the area under the curve; PR, precision curve; AP, area under the precision recall curve.

Clinical-RD Model Generation and Evaluation

To obtain a more personalized prediction model, the variables in the clinical-ultrasonic and RD were included in multivariate logistic regression analysis. The results showed that RD, age, and BI-RADS grade were associated with DCIS with statistical significance and applied for clinical-RD generation (P <0.05) (Table 7). Finally, the clinical value of the three models was compared by the ROC and DCA in the development and validation cohorts. As shown in Figure 4A, clinical-RD model (AUC=0.996) had a superior ability to clinical-ultrasonic model (AUC=0.947) and radiomics model (AUC=0.980) in the development cohort with statistical difference (P <0.05) (Table 8). The same results were observed in the validation cohort with the AUCs of 0.983, 0.945, and 0.907 for the three models (Figure 4B and Table 8). The DCA results also showed that the clinical-RD model had the highest clinical efficacy compared to the radiomics model and the clinical-ultrasonic model in both development and validation cohorts (Figure 4C and D).

Table 7 Logistic Regression Analysis Results

Table 8 The DeLong Test for Comparing the Area Under the Curves of Three Models

Figure 4 Predictive value and clinical efficacy of the three models in the development and validation cohorts. Receiver operating characteristic curve analysis of three models in the (A) development cohort and (B) validation cohort. Decision curve analysis of the three models in the (C) development cohort and (D) validation cohort. Model 1: clinical-ultrasonic model; model 2: radiomics model; model 3: Clinical-radiomics model. Abbreviation: AUC, area under the curve.

Discussion

In this study, the clinical and ultrasonic features of patients with DCIS and breast fibromatosis were collected and analyzed. The factors that could independently predict the DCIS were enrolled in the clinical-ultrasonic model construction. Besides, the ultrasound images were collected to extract a large number of radiomics features. Based on the LASSO method, collinearity analysis, and machine learning algorithm, significant radiomics features were selected for the radiomics model generation. Finally, the predictive value and clinical efficacy of the clinical-ultrasonic model, radiomics model, and the combined model were compared, among which the combined model exhibited superior ability.

Routine ultrasound is a basic imaging technique for detecting and diagnosing breast lesions.¹⁹ Fine linear branching and fine pleomorphic calcifications are closely associated with malignancy of DCIS as well as an elevated risk of recurrence than other calcification morphologies.²⁰ In the BI-RADS atlas (second edition), breast lesions are ultimately assigned a grade after analyzing ultrasound characteristics. Grade 0: a diagnosis that requires a combination of other imaging tests. Grade 1: negative results or no lesions. Grade 2: benign lesion, no suspicious features. Grade 3: benign lesions with a 2% probability of malignancy. Grade 4: suspicious lesion with 2–95% probability of malignancy, biopsy is recommended. Due to the wide range of the probability, Grade 4 is further divided into 4A, 4B, and 4C with a malignant probability of 2–10%, 10–50%, and 50–95%, respectively. Grade 5: suspicious lesion with over 95% probability of malignancy. Grade 6: malignancy proved by biopsy.²¹ Cha et al, reported that calcification was associated with pathological and biological markers.²² Besides, Sharma et al, showed the importance of age in DCIS.²³ Herein, calcification-dotted, BI-RADS grade ≥4A and older age were independently related to a higher probability of developing DCIS, and the three factors were used for building a clinical-ultrasonic model with satisfactory performance in distinguishing DCIS from breast fibromatosis.

As a method of quantitatively analyzing the gray value of medical images, radiomics reflects the texture and morphological characteristics of tumors.²⁴ Many quantitative features from medical images can be extracted by computer algorithms.²⁵ The degree of tumor heterogeneity is a prognostic biomarker related to cancer progression revealed by genomic analyses.^25,26 Notably, radiomics features characterize the heterogeneity throughout the entire tumor volume; however, biopsies only represent a small tumor portion or even a single site. Thus, radiomics features are promising markers of cancer aggressiveness.¹⁵ In addition, radiomics features represent an essential role in predicting treatment response and patient survival.¹⁴ After LASSO screening, and collinearity analysis, 12 radiomics features were identified and their feature classes included firstorder, glszm, ngtdm, gldm, and glcm. The First-order features including gray-level mean, maximum, minimum, variance, and percentiles are based on the global gray-level histogram.⁹ GLSZM counts the number of zones of linked voxels with the same gray level from the basis of the matrix.²⁷ NGTDM (Neighborhood Gray-Tone Difference Matrix) quantifies the difference between a gray level and the average gray level of its neighbors within a predefined distance.²⁸ GLDM (Gray Level Dependence Matrix) reflects the gray homogeneity of the local focus.²⁹ GLCM refers to the spatial relationship of pairs of pixels or voxels with predefined grayscale intensities and distances in different directions.²⁷ These traits that are impossible to detect by visual inspection can be exploited by radiomics, providing critical data for the prediction and diagnosis of diseases.³⁰ Interestingly, the logistic regression-based radiomics model outperformed the other models with good and stable prediction accuracy. Logistic regression is employed to assess the relationship between the binary dependent variables and independent variables, which can estimate the probability of the target variable. RF is an ensemble learning algorithm that uses multiple decision trees to provide a class prediction. SVM is a supervised model that identifies the optimal hyperplane, dividing the data into different classes. KNN is also a non-parametric learning algorithm in which output objects are classified using their local neighborhoods to formulate predictions. As a supervised learning algorithm, MLP can learn non-linear models. XG Boost uses an ensemble of weak prediction models for classification.^31,32

Radiomics as a non-invasive method has been widely used in various diseases. Jin et al, reported that the ultrasound-based radiomics model incorporating clinical information and RD can effectively predict central lymph node metastasis in papillary thyroid carcinoma patients with Hashimoto's thyroiditis.³³ The machine learning model based on ultrasound radiomics features performed well in discriminating primary and metastatic liver cancer.³⁴ Besides, the clinical parameters combined with the radiomics can independently predict the microvascular invasion in hepatocellular carcinoma.³⁵ The MRI-based radiomics model can predict the overall survival and tumor-infiltrating macrophages in gliomas.³⁶ This study revealed the significance of the model combined with the independent clinical-ultrasonic features and the RD obtained from the radiomics in distinguishing DCIS from breast fibromatosis.

For strengths, our study adopted the multimodal machine learning method to construct different models and selected the optimal classifier for distinguishing DCIS from breast fibromatosis. Then, the predictive value and clinical efficacy of the three models were explored in the development and validation cohorts using ROC and DCA. Despite some promising findings, only the ultrasound images were collected. The magnetic resonance imaging and mammography will be conducted to extract relevant radiomics features and the clinical efficacy of different radiomics models based on different examination methods should be compared in the future.

In conclusion, our study employed machine learning methods to develop a model effectively predicting the DCIS. The combined model achieved a satisfactory performance in distinguishing DCIS from breast fibromatosis.