PREDICTION OF BREAST CANCER MOLECULAR SUBTYPES BASED ON MULTI-PARAMETER MRI

1. Introduction

The incidence of breast cancer has been increasing annually in recent years, making it the leading cancer threatening women’s health. Early detection and treatment can significantly reduce breast cancer mortality and improve the cure rate. The molecular subtyping of breast cancer is a key factor that influences treatment approaches and provides the necessary basis for personalized and precise treatment for breast cancer patients.^1,2,3 The concept of molecular subtypes of breast cancer was first proposed by the National Cancer Institute (NCI) in the United States in 1999. It refers to the classification of breast cancer based on the expression of breast cancer cell-related receptors, including estrogen receptor (ER), progesterone receptor (PR), HER-2, and tumor cell proliferation index (Ki-67). Breast cancer can be classified into four molecular subtypes: Luminal A (ER/PR positive, HER-2 negative, Ki-67 expression $<14$%), Luminal B (ER and/or PR positive, HER-2 negative, Ki-67 expression $\ge 14$%), HER-2 overexpression (ER and PR negative, HER-2 positive), and triple-negative (ER, PR, and HER-2 all negative). Patients with different molecular subtypes of breast cancer exhibit different clinical features, recurrence rates, and survival rates, and also respond differently to the same treatment regimens.⁴ Among them, Luminal A subtype has the highest survival rate and is more responsive to endocrine therapy, but less responsive to chemotherapy⁵; Luminal B subtype has a higher proliferation capacity and is commonly seen in older populations, but is also responsive to endocrine therapy. HER-2 overexpression subtype shows good effectiveness in neoadjuvant chemotherapy and targeted therapy, but has a worse prognosis compared to Luminal A and Luminal B subtypes, with many patients in advanced stages. Triple-negative breast cancer is the subtype with the worst prognosis among the four molecular subtypes, with the highest rates of relapse and mortality, and limited effectiveness of other neoadjuvant treatment approaches apart from chemotherapy. Currently, the main method to determine molecular subtypes is through immunohistochemical testing (IHC) using core needle biopsy, which often comes with significant side effects. Therefore, establishing a safe and effective computer-aided diagnostic system for breast cancer molecular subtyping holds high medical value.

As a high-resolution three-dimensional medical imaging technique, MRI can provide comprehensive and detailed information about tumor tissue, which is very important for predicting molecular subtypes of breast cancer. Moreover, studies have demonstrated that MRI is useful in predicting the molecular subtypes of breast cancer.⁶ Fan et al.⁷ extracted 88 imaging features and background parenchymal enhancement features from DCE. They also combined two clinical features and used an evolutionary algorithm to obtain an optimal subset of features. They then used logistic regression to build a multi-classifier for breast cancer molecular subtyping, achieving an overall area under the curve (AUC) value of 0.869. The AUC values for Luminal A, Luminal B, HER-2 overexpression, and triple-negative were 0.867, 0.786, 0.888, and 0.923, respectively. Grimm et al.⁸ extracted 56 imaging features from preoperative contrast-enhanced MRI (CE-MRI) of 275 breast cancer patients. They used multivariate analysis to determine the association between imaging features and Luminal A and Luminal B subtypes. Holli-Helenius et al.⁹ extracted imaging texture features from T1WI to build a predictive model to differentiate Luminal A and Luminal B subtypes. The AUC value was 0.828. They also concluded that texture features such as heterogeneity, variance, smoothness, and homogeneity may directly or indirectly reflect the underlying growth patterns of breast tumors. Zhang et al.¹⁰ utilized DCE and employed convolutional neural networks (CNN) and convolutional long short-term memory networks (CLSTM) for transfer learning to differentiate three molecular subtypes (HR+/HER2−, HER2+, and triple-negative). The accuracy of CNN increased from 0.47 before transfer to 0.78 after transfer, and the accuracy of CLSTM increased from 0.39 before transfer to 0.74 after transfer. Koo et al.¹¹ found that two pharmacokinetic parameters derived from DCE were higher in triple-negative breast cancer patients compared to Luminal A and Luminal B patients. Li et al.¹² proposed an improved region growing algorithm for lesion segmentation, which can accurately identify lesion boundaries in DCE. They used a recursive feature elimination algorithm based on multiple models for feature selection. The identification accuracies for Luminal A, Luminal B, HER-2 overexpression, and triple-negative subtypes were 0.91, 0.89, 0.83, and 0.87, respectively. Sun et al.^13,14 utilized three CNNs with weighted voting based on DCE to differentiate between Luminal and non-Luminal subtypes. The average AUC value from five-fold cross-validation was 0.910. In the same year, they incorporated an unsupervised pretraining model to further optimize the predictive performance of the model. Yang et al.¹⁵ extracted radiomics features from MRI with two types of parameters, CDT-VIBE and RS-EPI, and analyzed the relationship between these features and molecular subtypes. They built a binary classifier to distinguish between triple-negative and Luminal A subtypes, with the highest AUC value reaching 0.811. Ha et al.¹⁶ developed a 13-layer CNN based on pre-treatment MRI and used random affine transformation to augment the datasets with different lesion models. Through this method, they increased the data volume and achieved an accuracy of 70% and an AUC value of 0.871 on the test set. Sutton et al.¹⁷ collected data on three molecular subtypes of breast cancer (ERPR+, ERPR−/HER2+, TN) and used nine pathological and radiological features to build a predictive model. The overall prediction accuracy was 83.4%, with prediction accuracies of 69.9% for ERPR+, 62.9% for ERPR-/HER2+, and 81.0% for TN subtypes. Leithner et al.¹⁸ extracted imaging features from CE-MRI and used Fisher’s discriminant analysis, error probability, and average correlation for feature selection. They selected linear discriminant analysis and k-nearest neighbors algorithm to predict receptor status and molecular subtypes. The results showed a high correlation between CE-MRI imaging features and receptor status as well as molecular subtypes.

While previous studies have primarily focused on single-parameter MRI, there is a lack of research on the combined use of multiple MRI parameters for breast cancer subtyping. Most discussions only address the relationship between medical imaging features and molecular subtyping of breast cancer using statistical methods. There were also a few studies directly investigating the four-class classification of molecular subtyping in existing research on molecular subtyping prediction. This study addresses this gap by constructing molecular subtyping prediction models based on four single-parameter MRIs and seven multi-parameter MRIs using image quantification features. This study introduces a novel approach by integrating multiple MRI parameters, which has the potential to enhance the accuracy and reliability of breast cancer subtyping.

2. Materials and Methods

2.1. Dataset

Multi-parametric MRI data (DCE, DWI, T1WI, and T2WI) were collected from 441 female breast cancer patients aged between 23 and 75 years at Liaoning Cancer Hospital for a retrospective study. None of these patients had undergone any surgical treatment prior to MRI imaging and all had complete pathological examination records. Based on the expression levels of ER, PR, HER-2, and Ki-67 in the pathological examination results, these patients were classified into four different molecular subtypes: 80 cases of Luminal A type, 221 cases of Luminal B type, 65 cases of HER-2 overexpression type, and 75 cases of triple-negative type. After exclusion screening, a total of 318 patient cases that met the experimental criteria were obtained, including 57 cases of Luminal A type, 162 cases of Luminal B type, 46 cases of HER-2 overexpression type, and 53 cases of triple-negative type. The exclusion data included: 89 cases missing any one of the DCE, DWI, T1WI, or T2WI MRI parameter images; 19 cases with incomplete pathological data and missing clinical information; and 15 cases with poor image quality, difficult tumor location confirmation, or unclear internal details or margins of the tumor.

Due to the significantly larger number of Luminal B type samples compared to the other three molecular subtypes, in order to ensure a balanced dataset, 40 Luminal A type cases, 40 Luminal B type cases, 37 HER-2 overexpression cases (the number of cases in this type is the smallest, so to ensure there are enough cases for testing), and 40 triple-negative cases were randomly selected as the training set, while the remaining data were used as the test set. The division results of the MRI image dataset are shown in Table 1.

2.2. MRI imaging techniques

All MRI data were acquired using a 1.5T MRI scanner (HDX, GE Healthcare) with an eight-channel breast coil. The patients were positioned in a prone position during MRI imaging, with both breasts naturally hanging down. The image storage format was DICOM. The scanning parameters for the four types of sequences were as follows.

The DCE and T1WI sequences were obtained using the volume interpolated breath-hold examination with real-time adaptive nonlinear techniques (VIBRANT), featuring an acquisition time (TA) of 42s, a repetition time (TR) of 6.32ms, and an echo time (TE) of 3.024ms, a flip angle of $10^{\circ }$, a slice thickness of 3.2mm, and a field of view (FOV) of 36cm × 36cm. Each sequence contained 48 images, each with a resolution of 512 × 512 pixels. The DCE sequence involved the injection of a contrast agent (0.1mmol/kg GD-DTPA-MBA, Omniscan, GE Healthcare) through an elbow vein at a rate of 3ml/s using a power injector, followed by a washout of 20ml of saline at the same rate. Immediately after the intravenous injection, continuous acquisitions were performed in eight phases without interruption, with a scanning time of 42s for each phase.

DWI employed echo-planar imaging (EPI) sequences, with a TA of 96s, a TR of 6000ms, a TE of 65ms, a diffusion sensitivity coefficient (b) of 800s/mm², a slice thickness of 6mm, a slice spacing of 7.5mm, a FOV of 34cm × 34cm, and a matrix size of 256 × 256.

T2WI was performed using a fast spin-echo (FSE) fat-suppressed sequence, with a TA of 101 seconds, a TR of 4040ms, and a TE of 84ms, a slice thickness of 5mm, a slice spacing of 6mm, a FOV of 22cm × 22cm, and a matrix size of 512 × 512.

2.3. Regions of interest (ROIs) segmentation, quantitative feature calculation and selection

Prior to feature calculation, all MRI images underwent preprocessing steps including normalization, noise reduction, and alignment to ensure consistency across the dataset. Then, two-dimensional ROIs segmentation was performed on MRI images of four parameters. Initially, a region-growing algorithm was used for preliminary segmentation to obtain the contours of the ROIs. Subsequently, with the guidance of two radiologists with seven years of clinical experience, the ITK-SNAP software was utilized to correct the contours of the preliminarily segmented ROIs, and the segmented two-dimensional ROIs were stacked into three-dimensional volumes of interest (VOIs). As shown in Fig. 1, ROIs were extracted from DCE, DWI, and T1WI, including the tumor portion, the tumor-side breast portion, the contralateral portion, and the contralateral breast portion. From T2WI, only the tumor and its edge portions were extracted. Each case could yield a total of 43VOIs, which included: 32VOIs from the DCE images (8 sequences × 4 parts), 8VOIs from the DWI images (2 sequences × 4 parts), 4VOIs from the T1WI images (1 sequence × 4 parts), and 1VOI from the T2WI images (1 sequence × 1 part).

The three-dimensional VOIs were processed using a three-dimensional wavelet transform (WT) and then fed into a radiomics feature extractor software written in MATLAB to extract three-dimensional radiomics features. The features were selected based on their relevance to breast cancer characteristics, including texture, shape, and enhancement patterns observed in MRI images. Extracted features include 13 morphological features, 31 first-order statistical features, texture features, dynamic enhancement features, and wavelet features. Texture features include 22 gray-level co-occurrence matrix (GLCM) features,^19,20,21,22 13 gray-level run-length matrix (GLRLM) features,^23,24 13 gray-level size zone matrix (GLSZM) features,²⁵ and 5 neighborhood gray-tone difference matrix (NGTDM) features.²⁶

Dynamic enhancement features are unique to DCE. Malignant tissues typically cannot effectively absorb metabolic contrast agents, resulting in significant signal changes after injection. These changes can be used to plot the time-intensity curve (TIC), which reflects the hemodynamic characteristics of the region and is crucial for disease diagnosis. In this paper, we employed a total of nine-time sequences, including eight post-injection sequences and one pre-injection T1 sequence, to construct the TIC. By using time as the horizontal axis and the average gray level of MRI images from each time series as the vertical axis, we obtained nine points on the curve derived from the average gray levels of the nine-time sequences, allowing us to fit the following polynomial :

Based on the TIC, this paper extracted 16 features as follows: Gray Ratio Parameter, Absorption Rate, Expurgation Rate, Time to Peak, Maximum Concentration, Maximum Slope, Maximum Curvature, Curve Fitting Parameters, Signal Decrease Rate During Washout, Signal Enhancement Ratio, Initial Enhancement Percentage, Initial Enhancement Rate, Maximum Enhancement Percentage, Maximum Enhancement Rate, Outflow Percentage, Outflow Rate.

WT classifies different spectral information in images into separate components, with high-frequency information decomposed into high-frequency images (H) and low-frequency information decomposed into low-frequency images (L). In the two-dimensional WT, decomposition is performed first along the x-axis and then along the y-axis, resulting in four wavelet images (LL, LH, HL, HH). In the three-dimensional WT, decomposition is carried out sequentially along the x-axis, y-axis, and z-axis, generating eight wavelet images (LLL, LLH, LHL, LHH, HLL, HLH, HHL, HHH). The WT significantly increases the quantity of various features, and this experiment utilizes this method to achieve a substantial increase in feature quantity.

We used the sequential forward floating selection (SFFS) feature selector²⁶ to filter out the optimal feature subset for four-class molecular typing from the training set. In this experiment, we selected four evaluation metrics: accuracy (acc), sensitivity (sen), specificity (spe), and F1-score (scr) to assess the selected features. Taking the support vector machine (SVM) classifier as an example, the variations of the evaluation metrics with changes in the number of features in the effective feature subset are shown in Fig. 2. Each classifier’s hyperparameters were optimized based on these metrics.

2.4. Construction and evaluation of molecular subtyping prediction models based on single-parameter MRI

We selected five traditional classifiers to develop a four-class prediction model for breast cancer molecular subtyping: SVM, Logistic Regression (LR), Naive Bayes (NB), K-Nearest Neighbors (KNN), and Random Forest (RF). The code for this experiment was written in Python, utilizing the function interface provided by scikit-learn to construct the machine learning prediction models. Each classifier performed feature selection using the optimized SFFS method mentioned above, resulting in effective feature sets for all seven classifiers. The normalized optimal feature subsets from DCE, DWI, T1WI, and T2WI were then separately fed into the prediction model for training and predicting breast cancer molecular subtyping based on single-parameter MRI.

To thoroughly evaluate the models’ performance and mitigate the effects of data imbalance in the test set, we utilized several metrics: accuracy, balanced accuracy (BA), AUC values for each of the four molecular subtypes, and the average AUC value. These metrics enabled a comprehensive assessment of the predictive capabilities of the models.

2.5. Molecular subtyping prediction results based on single-parameter MRI

According to the data displayed in Tables 2–5, in the four-class molecular typing model constructed based on a single parameter, T2WI demonstrates the best predictive performance, with an average AUC value reaching 0.794. In contrast, other imaging techniques such as DCE, DWI, and T1WI have average AUC values of only 0.548, 0.601, and 0.611, respectively, indicating a significant gap compared to the predictive performance achievable by T2WI.

In the further analysis of the prediction model constructed based on T2WI, the results indicate that among the five classifiers, LR demonstrates the most significant predictive performance. We found that the AUC values for Luminal A, Luminal B, Her-2 overexpression, and triple-negative subtypes were 0.936, 0.806, 0.733, and 0.701, respectively. Notably, the predictive outcome for Luminal A is the best, highlighting its advantage in the single-parameter molecular subtype prediction model. Figures 3 and 4 present the confusion matrix and ROC curve for the LR classifier based on T2WI, providing a visual assessment of the model’s performance.

2.6. Construction and evaluation of molecular subtyping prediction model based on multi-parameter MRI

Although the performance of the molecular subtyping prediction models constructed based on DCE, DWI, and T1WI is significantly weaker than that of the classification model built on T2WI, MRI images captured under four different parameters reflect information about breast cancer from various perspectives. By merging the optimal feature subsets extracted from these multi-parameter images, more comprehensive feature subsets can be formed, allowing for better utilization of MRI image information for breast cancer molecular subtyping prediction. In this experiment, the optimal feature subsets from multiple parameters will be input into the classifiers mentioned earlier for training, in order to compare the predictive performance of the model constructed based on multi-parameter MRI with that of the model constructed based on T2WI. The process of merging multi-parameter MRI feature fusion for molecular subtyping prediction is illustrated in Fig. 5.

3. Results

In this experiment, we combined MRI images with different parameters to obtain seven results of multi-parameter fusion for comparison. The combinations including: (1) DCE+DWI+T1WI+T2WI; (2) DCE+T2WI; (3) DWI+T2WI; (4) T1WI+T2WI; (5) DCE+DWI +T2WI; (6) DCE+T1WI+T2WI; (7) DWI+T1WI +T2WI. The four-class classification results for these seven combinations are presented in Tables 6–12, respectively.

Through the analysis of the four-class classification results, we found that the optimal predictive parameter combination after feature fusion is DWI+T2WI, which achieved the highest average AUC value of 0.823. The differences among the various classifiers are minimal, with average AUC values of 0.822, 0.800, 0.776, and 0.749, as shown in Table 8. Additionally, the highest average AUC values for the combinations DCE+DWI+T2WI, DWI+T1WI+T2WI, and DCE+DWI+T1WI+T2WI are 0.822, 0.801, and 0.803, respectively, as presented in Tables 10, 12, and 6.

From Table 10, it can be seen that although the average AUC value of DCE+DWI+T2WI based on the SVM classifier is 0.822, which is close to that of DWI+T2WI, the predictive performance of the other four classifiers is inferior to that of DWI+T2WI, with average AUC values of 0.785, 0.769, 0.714, and 0.798, respectively. As well, for DWI+T1WI+T2WI, only the RF classifier performs relatively well, achieving an average AUC value of 0.801. Compared to the model based solely on T2WI, these two combinations show a slight improvement in average AUC values, but their overall classification performance still does not match that of DWI+T2WI.

Furthermore, in the parameter combinations of DCE+T1WI+T2WI, DCE+T2WI, and T1WI+T2WI, the highest average AUC values did not improve but instead decreased to 0.785, 0.793, and 0.784, respectively. In summary, the feature fusion of DWI+T2WI significantly enhanced the model’s predictive performance; however, after introducing DCE or T1WI, the model’s predictive capability not only failed to improve but also showed a certain degree of decline.

The improvement in predictive performance achieved by the multi-parameter model was statistically significant ($p<$ 0.05) compared to single-parameter models, indicating the robustness of the proposed approach. Compared to previous studies using single-parameter MRI, our multi-parameter approach achieved a higher average AUC, demonstrating the added value of feature fusion.

4. Discussion

This study first explores the feasibility of predicting molecular subtypes of breast cancer using quantitative features extracted from single-parameter MRI images. By comparing four types of MRI parameter images, the results indicate that T2WI performs best in the four-class prediction of breast cancer molecular subtypes. The analysis of this phenomenon mainly involves two aspects:

(1)

Differences in Image FoV: Although the MRI images of DCE, T1WI, and T2WI all have a size of $512\times 512$, T2WI focuses more specifically on the breast region of the patient, while the other two imaging modalities encompass the entire thoracic cavity. This results in a higher proportion of lesions and their margins in T2WI images, providing richer details that enhance the model’s predictive capability.

(2)

Differences in Imaging Direction: DCE, DWI, and T1WI utilize longitudinal imaging, whereas T2WI employs transverse imaging. These different imaging modes allow the images to reflect signal variations of the lesions in both longitudinal and transverse directions; therefore, the molecular subtype prediction for breast cancer may be more sensitive to changes in transverse signals, leading to better predictive performance.

This study focuses on predicting breast cancer molecular subtypes by employing a multi-parameter MRI feature fusion approach, aiming to effectively leverage information from multi-parameter MRI images in constructing accurate prediction models and analyzing the effects of various parameter combinations on prediction outcomes. The experimental results indicate that the model combining DWI+T2WI performs best, showing an improvement in predictive performance across five classifiers compared to the model based solely on the single T2WI parameter. However, when further incorporating DCE or T1WI into the feature fusion, the predictive performance of the model did not improve and even declined. The reasons for this phenomenon may include the following:

(1)	The optimal feature subsets from DCE and T1WI did not provide effective information that would assist the DWI+T2WI combination, making it difficult to enhance model performance by adding these two parameters.
(2)	Increased Risk of Overfitting: The introduction of DCE and T1WI increased the dimensionality of the training optimal feature subset, thereby raising the risk of model overfitting.
(3)	Limited Sample Size: Due to the relatively small sample size, extracting useful information from DCE and T1WI images that could effectively support molecular subtype prediction became more challenging.

As shown in Table 13, existing research findings indicate that some studies on breast cancer molecular subtype prediction^9,13,15 focused solely on binary classifications and did not involve four-class predictions. In contrast, among the studies that addressed four-class molecular subtype predictions,^7,10,16 the multi-parameter model proposed in this paper demonstrated relatively satisfactory predictive performance. This suggests that our research method provides valuable insights for the precise diagnosis and treatment of breast cancer molecular subtypes.

One limitation of this study is the potential bias introduced by the retrospective nature of the dataset. Additionally, the model’s generalizability to other populations or imaging settings requires further validation.

Future work should explore the integration of additional imaging modalities, such as PET or CT, to further enhance the predictive accuracy of the model. Moreover, prospective studies are needed to validate the clinical utility of the proposed approach.

5. Conclusions

This paper proposes a molecular subtype prediction model based on multi-parameter MRI and compares the predictive performance differences between models based on single-parameter MRI and those based on multi-parameter MRI. The experimental results indicate that in the molecular subtype prediction models established using single-parameter MRI, T2WI exhibits the best predictive performance; meanwhile, in the study of multi-parameter MRI, seven different parameter combinations were used to construct molecular subtype prediction models, among which the model based on DWI+T2WI demonstrated the best performance. Currently, most existing studies on breast cancer molecular subtype prediction have focused solely on binary classification, while in studies that conducted four-class predictions, the multi-parameter model proposed in this paper achieved relatively satisfactory predictive results.

The implementation of this multi-parameter MRI model in clinical practice could significantly improve the accuracy of breast cancer subtyping, leading to more personalized and effective treatment plans.

Acknowledgment

This study was supported in part by the National Natural Science Foundation of China (NSFC) (Grant 82260505).

ORCID

Xiaoguang Pan  https://orcid.org/0009-0008-1842-6698

Yi Liang  https://orcid.org/0009-0000-3106-7927

Hongfei Yu  https://orcid.org/0009-0003-7210-5581

Lu Han  https://orcid.org/0009-0002-1062-0994

Hong Li  https://orcid.org/0000-0002-2460-8841

Tao Yu  https://orcid.org/0000-0003-4936-8861