Prediction of the clinicopathological subtypes of breast cancer using a fisher discriminant analysis model based on radiomic features of diffusion-weighted MRI

Breast cancer is the second most common cancer cause of cancer death in females [1]. Based on gene expression profiling, four intrinsic molecular subtypes can be defined: luminal A, luminal B, human epidermal growth factor receptor 2 (HER2)-enriched, and basal-like [2‐4].

One clinicopathological classification of breast cancer focused on therapeutic purposes has been adopted by the 12th International Breast Cancer Conference [5]. These clinicopathological subtypes are similar but not identical to the intrinsic molecular subtypes. There are five clinicopathological subtypes including luminal A, luminal B_HER2- (luminal B/HER2 negative), luminal B_HER2+ (luminal B/HER2 positive), HER2 positive and triple negative [5] (Table 1). Four immunohistochemical (IHC) biomarkers, including oestrogen receptor (ER), progesterone receptor (PR), HER2, and Ki-67, are recommended to define the clinicopathological subtypes. This classification is aimed at systematic therapy: luminal A cases require endocrine therapy; luminal B_HER2- cases require endocrine therapy with or without cytotoxic therapy; luminal B_HER2+ cases require cytotoxic, anti-HER2 and endocrine therapy; HER2 positive cases require cytotoxic and anti-HER2 therapy; and triple negative cases require cytotoxic therapy. At least two advantages of the clinicopathological subtypes are as follows: first, in contrast to high cost and time-consuming gene expression array testing, clinicopathological subtyping is simplified and can be conducted easily in clinical practice; second, this subtyping undoubtedly contributes to the treatment selection of breast cancer.

Diffusion-weighted imaging (DWI) is an essential sequence that can monitor the mobility of water molecules. With restricted water diffusion, breast cancer usually shows hyperintensity on DW images [6]. DWI contributes to the differential diagnosis of breast lesions and may be a promising tool in breast cancer detection [7]. In differentiating malignant and benign breast lesions, the diagnostic performance of contrast-enhanced magnetic resonance imaging (MRI) with DWI is higher than that of contrast-enhanced MRI with time-intensity curves [8]. In addition, DWI also has the potential to monitor radiation-induced treatment response and neoadjuvant treatment response [9, 10]. More importantly, DWI can be an alternative for breast cancer screening without contrast media [6].

Radiomics is a process of converting digital medical images into mineable high-dimensional data [11]. It has been used in the detection and diagnosis of cancer, assessment of prognosis, prediction of response to treatment, and monitoring of disease status [11, 12]. The applications of radiomics in breast cancer include the prediction of molecular classification [13, 14], assessment of tumour recurrence [15], and response to treatment [16]. Recently, radiomics, by using the image phenotyping of breast cancers and their surrounding parenchyma on dynamic contrast-enhanced MRI, was used to identify triple-negative breast cancer [13]. Another radiomic study showed a positive trend between the molecular cancer subtype and breast tumour phenotype of size and enhancement texture based on dynamic contrast-enhanced (DCE) MRI [14]. Unfortunately, the differential diagnosis of diverse molecular subtypes was not explored in this study. Mammographic radiomic features could also be used for the prediction of breast cancer molecular subtypes with oversimplified classifications such as triple-negative and non-triple-negative, HER2-enriched and non-HER2-enriched, and luminal and non-luminal [17]. Radiogenomics is a novel approach that can correlate imaging characteristics with underlying genes, mutations and expression patterns at the genetic level [18]. Radiogenomics can be imaging surrogates for genetic tests and can reflect tumour biology [19]. One recent study demonstrated that radiogenomics of breast cancer could infer underlying gene expression by using RNA sequencing [20]. Breast tumours with higher expression levels of the JAK/STAT and VEGF pathways had more intratumour heterogeneity in image enhancement texture detected by using dynamic contrast-enhanced MRI. Fan et al. showed that the image features of DCE-MRI were associated with gene expression modules and could predict the prognosis of breast cancer patients [21]. In addition, for breast cancer subtyping, the combination of correlated miRNAs and imaging features has better classification power in differentiating luminal A and other breast cancer subtypes than using miRNAs or imaging alone [22]. By means of deep learning or novel algorithms, the breast cancer molecular subtypes were differentiated based on image features from DCE-MRI [23, 24].

As mentioned above, the clinicopathological subtypes of breast cancer are essential for treatment selection. Thus, it is urgent to develop a reliable method for the prediction of the clinicopathological subtypes. Diffusion-weighted imaging has high sensitivity in detecting breast cancer and is widely used in clinical practice. Herein, our study aims to verify the feasibility of using a radiomic approach based on DWI for the prediction of clinicopathological breast cancer subtypes.

There were 29 luminal A cases, 31 luminal B_HER- cases, 17 luminal B_HER+ cases, 24 HER2-positive cases and 11 triple-negative cases in our study (Table 2).

A total of 162 radiomic features of four statistical methods were included after preprocessing. Of the 162 radiomic features, there were 42 histogram parameters, 18 texture parameters, 52 GLCM parameters and 50 RLM parameters. Each of 162 features would be evaluate in the Fisher discriminant analysis to determine its contribution, with which feature contributed to modelling would be held. Finally, 84 features were used for the model-building.

The overall accuracy for predicting the clinicopathological subtypes was 96.4% by Fisher discriminant analysis. Based on class-weighted accuracy calculating, the overall weighted accuracy was 96.6%. For predicting diverse clinicopathological subtypes, the prediction accuracies ranged from 92 to 100%. When predicting subtypes of luminal B_HER2- and triple negative, both accuracies were 100% (Table 3).

A leave-one-out cross-validation was performed to test the Fisher discriminant analysis model. The overall accuracy of the model was 82.1% in the prediction of the clinicopathological subtypes of breast cancer. The accuracies of the model for predicting the luminal A, luminal B_Her2-, luminal B_Her2+, Her2-positive and triple-negative subtypes were 79, 77, 88, 92 and 73%, respectively (Table 3).

The areas under the ROC curve (AUROCs) of histogram parameters, texture parameters, GLCM parameters and RLM parameters for predicting different IHC biomarkers are shown in detail in Table 4 and Fig. 3. Furthermore, the AUROCs of histogram parameters, GLCM parameters and RLM parameters were higher than those of texture parameters in assessing the status of ER, PR, HER2 and Ki-67 (p < 0.001) (Table 5).

Our study has shown that the Fisher discriminant analysis model with radiomic features of DW images can be used for predicting the clinicopathological subtypes of breast cancer. Furthermore, the model had excellent accuracies ranging from 92 to 100% in the prediction of clinicopathological subtypes. Based on the leave-one-out cross-validation, the overall accuracy was 82.1% when testing the Fisher discriminant analysis model.

In our study, each clinicopathological subtype could be distinguished from others with high accuracy. Furthermore, we applied Fisher’s discriminant analysis to resolve a multiple classification problem, i.e., five clinicopathological subtypes of breast cancer. The clinicopathological subtypes of breast cancer are defined according to their therapeutic purposes. Our findings may contribute to decision-making and treatment selection in clinical practice, such as endocrine therapy, cytotoxic therapy, and anti-HER2 therapy [5]. As mentioned above, the molecular luminal B subtype is divided into two new clinicopathological subtypes, luminal B with or without HER2 overexpression (i.e., luminal B_HER2- and luminal B_HER2+). Luminal B_HER2+ cases require cytotoxics, endocrine therapy and anti-HER2 therapy, whereas a luminal B_HER2- cases does not require anti-HER2 therapy. Therefore, the clinicopathological subtype can provide detailed therapeutic information. Moreover, compared with molecular subtypes from gene assays, the clinicopathological subtypes from immunohistochemistry can be easily obtained with lower costs.

In addition, leave-one-out cross-validation was performed to evaluate the accuracy of the Fisher discriminant analysis model. Our results showed that the overall accuracy of the model was 82.1% in predicting the clinicopathological subtypes. The accuracy for predicting HER2 positivity was up to 92%. We confirmed that the Fisher discriminant analysis model is a reliable method that can be used for predicting the clinicopathological subtypes of breast cancer. More importantly, our study validated the feasibility of using the Fisher discriminant analysis model to settle a multi-classification problem.

In our study, we provided a noninvasive method for the prediction of the clinicopathological subtypes of breast cancer by using radiomic features of DWI. DWI is a widely used method that can measure the Brownian motion of water molecules. The motion of water molecules in tissue can be affected by tissue cellularity and membrane integrity [6]. DWI has been applied in the detection of breast cancer [34], differentiation of benign and malignant lesions [35], and monitoring the response to neoadjuvant chemotherapy [36, 37]. However, there is no consensus that DWI is a reliable stand-alone method. One recent study showed that the combination of T2-weighted fat suppression and DWI textural features could predict sentinel lymph node metastasis [38]. By using diffusion MRI, a radiomic signature was shown to differentiate malignant from benign lesions [39]. Furthermore, another study showed that ADC values correlate with the biological features of breast cancer [40]. In addition to DWI mentioned above, a number of studies have focused on radiomic analysis from DCE-MRI. The imaging features from DCE-MRI can predict the luminal A and luminal B molecular subtypes [24] and can also differentiate between the histological and immunohistochemical subtypes of breast cancer [41]. One DCE-MRI feature that quantifies the relationship between lesion enhancement and background parenchymal enhancement is associated with the luminal B subtype of breast cancer [42]. However, these imaging feature-based studies provide insufficient information for the differential diagnosis of the five clinicopathological subtypes of breast cancer. To resolve this problem, our study focused on the prediction of clinicopathological subtypes by using DW imaging features and showed a high diagnostic performance with an overall accuracy of 96.4%.

For predicting breast cancer receptor status and molecular subtyping, Leithner et al. showed that the breast tumour segmentation approaches could affect the classification accuracy by using radiomic signature of DWI [43]. In their study, two segmentation approaches included: (1) segmentation ROI performed on high b value DWI and copied to ADC map; (2) segmentation ROI drawn directly on ADC map. The results indicated that tumour segmentation directly on ADC map was of better classification accuracy. However, in their study, some lesions could not be identified on ADC map. Our study only drawn segmentation ROI on DWI without ADC map, on which distinct tumour margin could easily be confirmed.

The multiparametric MR radiomics using DCE and DWI can also be performed for predicting breast cancer subtypes [44]. Text features were extracted form DCE images of six contrast-enhanced phases and DWI with three b-values. The best accuracies of multiparametric MR radiomics model were 72.4 and 91.0% for the 4-IHC classification task and for the TN vs. non-TN cancers, respectively. However, for 4-IHC classification task, the accuracy of DWI with linear discriminant analysis model was 53.7% by using minor dependence emphasis on Kendall-tau-b. And for differentiating triple negative (TN) from non-TN tumours, the accuracy of DWI was 83.6% by using the maximum of variance. It indicated that the multiparametric MR radiomics performed well than DWI for 4-IHC classification task and for the TN vs. non-TN cancers.

Recently, Leithner et al. apply artificial intelligence (AI) to breast cancer molecular subtyping with multiparametric MR radiomics [45]. Texture features extracted from DCE images and ADC maps and a multi-layer perceptron feed-forward artificial neural network (MLP-ANN) were used for differentiation of TN and luminal A breast cancers from other subtypes. Their results indicated that multiparametric MR radiomics could provide prognostic and predictive information derived from the entire tumour before and during treatment.

In predicting the different statuses of IHC biomarkers, excellent diagnostic performance of radiomic features was found in differentiating between the ER-positive and -negative group, PR-positive and -negative group, HER2-positive and -negative group, and Ki-67-low and -high group. However, there were relatively low AUROCs of texture parameters in assessing the status of these biomarkers. This result may be ascribed to a small number of 18 texture parameters. Compared with our results, one study using 38 radiomic features showed a less powerful result, with AUROCs ranging from 0.641 to 0.789 in predicting ER, PR and HER2 statuses [46]. Another study showed a similar result with AUROCs ranging from 0.65 to 0.89 [14]. Therefore, radiomics-based approach can serve as a non-invasive method for predicting breast cancer receptor status.

In our study, histogram, GLCM and RLM features had high diagnostic performance in differentiating different IHC biomarkers. For differentiating ER-positive and ER-negative cases, PR-positive and PR-negative cases, HER2-positive and HER2-negative cases, and Ki-67-low and Ki-67-high cases, the AUROCs of histogram, GLCM and RLM features were 0.963–973, 0.923–0.939, 0.902–0.974 and 0.926–0.975, respectively. Furthermore, the Fisher discriminant analysis model had high accuracy in predicting the five clinicopathological subtypes of breast cancer. We concluded that radiomic features of DWI were highly associated with IHC biomarkers and could be emerging surrogates for IHC biomarkers. With the Fisher discriminant analysis model, we provided a new radiomic approach for breast cancer subtyping, and its findings contribute to treatment selection.

Our study has some limitations. First, the main limitations of the study are its small sample size and lack of independent external validation, which makes the results difficult to reproduce in other populations. We will enlarge the sample size and apply other methods for the clinicopathological classification of breast cancer. Second, the automatic segmentation method was not applied in our study. Hence, the ROI delineation was time-consuming. Third, the apparent diffusion coefficient (ADC) value for each breast tumour was not acquired. Therefore, we failed to compare the diagnostic performance of ADC with that of radiomic features. Fourth, the lack of morphological features could limit our results. Our future work will focus on resolving these limitations.

Our study demonstrated that the Fisher discriminant analysis model based on radiomic features of diffusion-weighted MRI could be used for the prediction of the clinicopathological subtypes of breast cancer with high accuracy. Moreover, the radiomic features had excellent diagnostic performance in differentiating between the ER-positive and -negative groups, PR-positive and -negative groups, HER2-positive and -negative groups, and Ki-67-low and -high groups. More breast cancer cases will be enrolled in our future work to validate this radiomic approach.