Combining Dynamic Contrast-Enhanced Magnetic Resonance Imaging and Apparent Diffusion Coefficient Maps for a Radiomics Nomogram to Predict Pathological Complete Response to Neoadjuvant Chemotherapy in Breast Cancer Patients

Abstract

Objective 

The objective of this study was to develop a nomogrom for prediction of pathological complete response (PCR) to neoadjuvant chemotherapy in breast cancer patients.

Methods 

Ninety-one patients were analyzed. A total of 396 radiomics features were extracted from dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) and apparent diffusion coefficient (ADC) maps. The least absolute shrinkage and selection operator was selected for data dimension reduction to build a radiomics signature. Finally, the nomogram was built to predict PCR.

Results 

The radiomics signature of the model that combined DCE-MRI and ADC maps showed a higher performance (area under the receiver operating characteristic curve [AUC], 0.848) than the models with DCE-MRI (AUC, 0.750) or ADC maps (AUC, 0.785) alone in the training set. The proposed model, which included combined radiomics signature, estrogen receptor, and progesterone receptor, yielded a maximum AUC of 0.837 in the testing set.

Conclusions 

The combined radiomics features from DCE-MRI and ADC data may serve as potential predictor markers for predicting PCR. The nomogram could be used as a quantitative tool to predict PCR.

Breast cancer is one of the most common malignant tumors in women,^1,2 and the incidence and mortality of breast cancer have been increasing in China.³ For the 10% to 20% of breast cancer patients who are diagnosed⁴ with locally advanced breast cancer (LABC), it is imperative to receive effective treatment as quickly as possible. Even with treatment, most patients with LABC will develop distant metastases.⁴ Neoadjuvant chemotherapy (NAC) is the standard of care for the management of early and LABC. The main objectives of this approach are to treat distant metastases early⁵ and to reduce the size of inoperable tumor, thereby enabling conservative breast surgery.⁶ A pathological complete response (PCR) after NAC strongly indicates a better prognosis than pathological partial response (PPR)^7,8 for patients with breast cancer. However, a significant proportion (up to 30%) of patients do not benefit from NAC because of poor pathological response and yet suffer from the toxicity and adverse effects associated with chemotherapy.⁹ Therefore, there is a need for reliable noninvasive pretreatment predictors of PCR that can enable effective targeting of NAC and prevent the delay in effective treatment for patients with poor response.

Clinical trials have described clinical biomarkers of breast tumor, which may predict the response to NAC. For instance, the proliferation index of Ki67¹⁰ and the status of estrogen receptor (ER)¹¹ has been shown to correlate positively with response. Unfortunately, the diagnostic modalities require tissue specimens typically obtained by needle biopsy. Because of the relatively small tissue sample and tumor heterogeneity, the clinical biomarkers and histological features assessed from a needle biopsy sample may not be representative of the entire tumor. Clinical imaging is a noninvasive approach that has the ability to cover a wider range and better captures tumor heterogeneity than a single tissue sample.¹² A number of studies have investigated the value of dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) and apparent diffusion coefficient (ADC) maps for assessing or predicting treatment response to NAC in breast cancer.^13–16 Most of them were limited by not considering the spatial information because the image features were calculated as aggregate measures in the entire tumor volume.

Recently, many studies have focused on detecting breast cancer subtypes or characterizing and predicting PCR after NAC using abundant features extracted from radiological images.^17,18 However, most of studies in radiomics are focused on extracting features from a 2-dimensional region of interest instead of a whole 3-dimensional tumor. As a result, the visualization or interpretation of tumor heterogeneity has not been fully explored.¹⁸ Furthermore, many previous studies in radiomics reported on either DCE-MRI or ADC maps but have not evaluated the performance of the combination of both techniques.

Therefore, the aim of this study was to develop and validate a radiomics nomogram that incorporated both the radiomics signature and clinical risk factors for PCR after NAC in patients with breast cancer and to evaluate the performance of the combination of DCE-MRI data and ADC data.

MATERIALS AND METHODS

Patients

This study was conducted in accordance with the Declaration of Helsinki and was approved by the ethics committee of our hospital, with the requirement for informed consent waived. All patients with biopsy-proven primary breast cancer who received NAC before breast surgery and had breast DCE-MRI and ADC maps before the initiation of NAC between January 1, 2016, and July 31, 2018, were retrospectively included in this study. The detailed selection criteria were as follows: (1) patients who went through NAC, (2) patients who had breast DCE-MRI and ADC mapping before NAC, and (3) patients who were assessed for their pathological response after NAC. The exclusion criteria were as follows: (1) occult breast cancer and (2) artifacts that affected the lesion visualization. Figure 1 depicts the patient selection process. These patients were randomly divided into training and testing set at the rate of 7:3.

FIGURE 1: Patient recruitment and study design.

Immunohistochemistry

The baseline clinicopathological data included age, expression of ER, progesterone receptor (PR), Ki67, HER2, and pathological response. We defined tumor with nuclear staining in less than 1% of tumor cells as ER/PR negative and 1% or greater of tumor cells as ER/PR positive.¹⁹ The cutoff value for Ki67 expression was set at 20%.²⁰ For HER2, tumor with immunohistochemistry staining of 3+ were defined as HER2 positive. For tumor with immunohistochemistry staining 2+, further confirmation was obtained with molecular testing (fluorescence in situ hybridization). Positive HER2 status was defined as an HER2 gene/chromosome 17 ratio of 2.0 or greater.²¹

NAC and Pathological Assessment of Response

All patients received 6 to 8 cycles of NAC before breast surgery. HER2− patients received 4 cycles of doxorubicin (60 mg/m²) plus cyclophosphamide (600 mg/m²) every 2 weeks followed by 4 cycles of paclitaxel (175 mg/m²) every 2 weeks. All HER2+ patients received treatment with docetaxel (75 mg/m²) every 3 weeks for 6 cycles and trastuzumab every 3 weeks for 6 cycles (at a loading dose of 8 mg/kg once and then 6 mg/kg every 3 weeks).

The histopathologically examination and analysis were dedicated by breast pathologists who were blinded to the MRI data. Miller-Payne grading system²² is a 5-level classification method: grade 1, no change or some alteration to individual malignant cells but no reduction in overall cellularity; grade 2, a minor loss of tumor cells but overall cellularity still high, up to 30% loss; grade 3, between an estimated 30% and 90% reduction in tumor cells; grade 4, a marked disappearance of tumor cells such that only small clusters or widely dispersed individual cells remain, more than 90% loss of tumor cells; and grade 5, no malignant cells identifiable in sections from the site of the tumor, only vascular fibroelastotic stroma remains often containing macrophages. However, ductal carcinoma in situ may be present. Grades 1 to 4 are categorized as a PPR, and grade 5 was a PCR.

MRI Data Acquisition

All examinations were performed with a 3.0-T MRI scanner (Magnetom Skyra; Siemens Healthcare, Erlangen, Germany) with 16-channel phased-array breast coils. The patients were in prone position with the bilateral breasts naturally suspended in the coil. The scanning protocol included quantitative time-resolved angiography with interleaved stochastic trajectories DCE-MRI, readout segmentation of long variable echo trains DWI. After an intravenous injection of 0.2 mmol/kg of gadopentetate dimeglumine with an injection flow rate of 3.0 mL/s, the time-resolved angiography with interleaved stochastic trajectories DCE-MRI sequence was performed with the following parameters: repetition time/echo time, 6.4/3.3 milliseconds; flip angle, 9.0; field of view, 288 mm × 384 mm; 2 mm slice thickness without slice gap; and 34 phases as first phase (17.7 seconds) and later phase (8.7 seconds). per phase within 5 minutes and 5 seconds. The readout segmentation of long variable echo trains DWI sequence had the following parameters: b values (50, 800 s/mm²); repetition time/echo time, 4800/56 milliseconds; flip angle, 180; field of view, 170 mm × 340 mm; and 4 mm slice thickness with 0.8 slice gap.

Radiomics Feature Extraction

Pretreatment MRI data were collected for tumor masking and feature extraction. The regions of interest were manually delineated via ITK-SNAP software (version 3.6.0, https://itk.org/) on each slice of the ADC maps and T1 + C (the peak enhanced phase of the DCE-MRI selected according to the time intensity curve) data. The largest lesion was selected if there were multiple lesions in a breast. The tumor was manually delineated by one experienced radiologist and then verified by another experienced radiologist to exclude the border of the lesion and any other irrelevant components.

After a tumor was segmented, the volume of interest images (digital imaging and communications in medicine format) were transferred to the AK software (Artificial Intelligence Kit V3.0.0. R; GE Company). A total 792 radiomics features were extracted from the DCE-MRI scans (396 radiomics features) and ADC maps (396 radiomics features) for each patient. The features included the following: (1) histogram parameters, (2) morphological features, (3) gray-level cooccurrence matrix, (4) gray-level run-length matrix, and (5) gray-level size zone matrix. All these features were normalized with Z scores to get a standard range before being used in a machine learning model for pathological response classification.

Feature Selection and Radiomics Signature Construction

Spearman correlation analysis (SPM) combined with the least absolute shrinkage and selection operator (LASSO) method²³ were used to select the most useful predictive features from the training data set. The threshold of SPM is 0.9. The LASSO logistic regression model was used with penalty parameter tuning that was conducted by 10-fold cross-validation based on minimum criteria. Multivariable logical regression was used to construction the predicted model by selected features. A radiomics score (Radscore) was calculated for each patient via a linear combination of selected features that were weighted by their respective coefficients in the constructed model. The mathematical expression can be described as follows as follows: , where b is the intercept, X_i is the value of ith selected feature, and C_i is the coefficient of the ith selected feature listed in Table 2.

TABLE 2: Selected Radiomics Features in DCE, ADC, and Joint Models

Multivariable logistic regression analysis began with the following clinical candidate predictors: age, ER expression, PR expression, HER2 expression, and Ki67 status. Then, the radiomics signature combined with independent clinical predictors was applied to develop a diagnostic model for PCR. A backward step-wise selection was applied by using the likelihood ratio test with Akaike's information criterion as the stopping rule.²⁴ Furthermore, we built a radiomics nomogram based on multivariable logistic analysis in the training set to provide the clinician with a quantitative tool to predict the individual probability of PCR.

To develop an optimal model with the highest accuracy, we evaluated the radiomics performance of DCE-MRI, ADC maps, and the combination of both techniques and then incorporated the independent clinical predictors to build the combination model. The classification performance of the model was evaluated by the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity.

Statistical Analysis

Decision curve analysis was conducted to determine the clinical usefulness of the radiomics model by quantifying the net benefits at different threshold probabilities in the validation data set.²⁵Figure 2 depicts the flowchart of the proposed analysis pipeline described previously.

FIGURE 2: Radiomics prediction pipeline for PCR.

Statistical analysis was conducted with R software (version 3.5.1,http://www.r-project.org/). The reported statistical significance levels were all 2-sided, and the statistical significance was set at 0.05. The multivariate logistic regression analysis and nomogram construction was performed with the rms package. Decision curve analysis was performed using the dca.R package.

RESULTS

Clinical Characteristics

In total, 63 patients were identified and comprised the training set, and 28 patients were identified and comprised the testing set, at a ratio of 7:3. The patient characteristics in the training and testing set are shown in Table 1. There was a significant difference in ER and PR status between the PCR and PPR patients in the training set (P < 0.001), which was then confirmed in the testing set (P < 0.05). There were no significant differences in age, HER-2 status, and Ki67 status between the PCR and PPR patients in the 2 cohorts (P > 0.05). There was a significant difference in the DCE-Radscore between the PCR and PPR patients in the training set (P < 0.001), which was then confirmed in the testing set (P < 0.05). There was a significant difference in the ADC-Radscore between the PCR and PPR patients in the training set (P < 0.001), but this difference was not confirmed in the testing set (P > 0.05).

TABLE 1: Characteristics of Patients in the Training and Testing Set

Feature Selection and Radiomics Signature Construction

Of the texture features, 396 or 792 texture features in 3 groups were reduced to form the predictors based on 63 patients in the training set. Table 2 lists the features selected by SPM and LASSO. The Radscore of each prediction model was calculated according to the coefficients in the multivariable logistic regression model.

Development and Performance of the Individualized Prediction Model

The prediction model based on DCE-MRI, ADC maps, and combination of DCE-MRI and ADC data was developed and quantitatively integrated into 3 Radscores: DCE-Radscore, ADC-Radscore, and Joint-Radscore. Univariate analyses identified ER and PR as independent clinical predictors (Table 1). Therefore, 3 Radscores combined with the ER and PR (namely, Clinical) are used to develop the predicted model. The receiver operating characteristic curve (ROC) of DCE-Radscore, ADC-Radscore, and Joint-Radscore model with or without clinical predictors is shown in Figures 3 to 5, and related AUC, accuracy, specificity and sensitivity are listed in Table 3. The Clinical-Joint Radscore yielded a maximum AUC of 0.931 in the train cohort. Therefore, we developed the Clinical-Joint Radscore nomogram (Fig. 6) and the decision curve (Fig. 7). The decision curve showed relatively good performances for the model with Clinical-Joint Radscore compared with that for the Joint-Radscore model. Across majority of the range of reasonable threshold probabilities, the decision curve analysis showed that Clinical-Joint Radscore model had a higher overall benefit than the Joint-Radscore model.

FIGURE 3: The ROC curves of DCE-Radscore and Clinical-DCE Radscore for predicting. Pathological complete response in the training and testing set.

FIGURE 4: The ROC curves of ADC-Radscore and Clinical-ADC Radscore for predicting PCR in the training and testing set.

FIGURE 5: The ROC curves of Joint-Radscore and Clinical-Joint Radscore for predicting PCR in the training and testing set.

TABLE 3: Performance of the Individualized Prediction Models

FIGURE 6: The radiomics nomogram to predict PCR after NAC. The nomogram was developed in the training set using the Joint-Radscore, ER status, and PR status. 0, PPR; 1, PCR.

FIGURE 7: Decision curve analysis for the Joint-Radscore model and Clinical-Joint Radscore model.

DISCUSSION

The utility of breast MRI in the context of NAC is already established to predict PCR using pre-NAC and post-NAC scanning to assess radiological response to treatment and guide surgical decision-making.^4,13,14 In this study, we built and validated a radiomics model that incorporated pretreatment DCE-MRI and ADC data for the noninvasive, preoperative, individualized prediction of PCR in breast cancer patients after NAC. Incorporating the radiomics signature and clinical risk factors into an easy-to-use nomogram facilitates the noninvasive prediction of PCR. The Clinical-Joint Radscore model performs well (AUC, 0.931) and thereby provides an effective tool for clinical decision-making.

Although feature extraction methods have been developed and tested, the important issues of predicting the response to NAC and optimally combining these features have not been fully investigated. Moreover, most prior studies used features from breast lesions^9,26–28 or background parenchymal enhancement alone.^29,30 Our study aimed to investigate the performance of the combined features with pretreatment DCE-MRI and ADC data to predict PCR in breast cancer.

In our study, radiomic features were used, including breast lesion and texture features based on multiparametric MRI (including DCE-MRI and ADC maps). The Joint-Radscore based on pretreatment DCE-MRI and ADC data were able to predict PCR with a greater accuracy (AUC, 0.848) than either DCE-Radscore (AUC, 0.750) or ADC-Radscore (AUC, 0.785) alone. Because the clinical information may only take into consideration certain aspects of the tumor (the AUC in training set for Clinical [namely, ER + PR] is 0.823), multiparametric MRI may better reflect all information on the tumor.³¹ Thus, when we combined the clinical information and imaging features for nomogram. By incorporating the clinical risk factors and the Joint-Radscore into the model, the overall predictive ability was quite strong in both the training and testing sets with AUCs of 0.931 and 0.837, respectively. One recent study investigated the PCR prediction capabilities of radiomic features with DCE-MRI and derived an AUC of less than 0.80,⁹ which was a lower value than that of the independent validation results obtained in our study. Meanwhile, we also found that the sensitivity of ADC-Radscore in the testing set was 1. The main reason was that the number of PCR patients was smaller than the number of PPR patients in our study, which reflects the real situation in which fewer patients have PCR after NAC. According to the existing research and actual situation, it is a common phenomenon that the number of PCR patients was smaller than the number of PPR patients after NAC,^32,33 which may inference verification of the technology. However, the effectiveness of radiomics in this issue had been proved in previous studies.^34,35 Our models could achieve the effective prediction performance for PCR patients. In the future, we will include more patients to further validate our models.

Intertumoural biological heterogeneity may influence response to treatment, and the assessment of these changes in heterogeneity may provide additional information to that captured by standard imaging assessments of tumor size or enhancement changes.³⁶ We observed that imaging features predict PCR after NAC. According to previous studies, kurtosis and skewness features were identified as biomarkers of tumor heterogeneity,³⁷ and high values of the expression of these biomarkers have been associated with treatment failure,³⁸ whereas low values indicate a response to treatment.³⁹ In our studies, kurtosis and skewness are selected. Moreover, the Cluster Shade, Long Run High Gray Level Emphasis and High Intensity Small Area Emphasis features are also included in the model. Cluster Shade is a measure of the skewness and uniformity of the gray-level cooccurrence matrix. A higher cluster shade implies greater asymmetry. Cluster Shade, Long Run High Gray Level Emphasis is a measure of the joint distribution of long run lengths with higher gray-level values. High Intensity Small Area Emphasis is a measure of Gray Level Zone Size Matrix, which is particularly efficient to characterize the texture homogeneity, nonperiodicity, or speckle like texture. All features described previously extract information about tumor heterogeneity effectively. As a result, the AUCs achieve 75% to 85% by using imaging features only. When ER and PR status were incorporated into the prediction model, the AUCs achieved a 10% increase approximately.

Our study had several limitations. The study results were assessed in a single institution in a relatively small set of patients. Therefore, the results in this study should be validated in larger cohorts. Second, we did not examine the imaging features of background glandular enhancement on DCE-MRI,⁹ although such lesions may be important predictive biomarkers. It is an interesting direction that the research on relationship between background glandular enhancement on DCE-MRI and the response of NCA and can be studied in-depth in the future.

In conclusion, the combined radiomics features from pretreatment DCE-MRI and ADC data could be used as potential biomarkers for predicting the response to NAC. The nomogram that incorporated the combined radiomics signature and clinical risk factors could be used as a quantitative tool to predict the PCR to NAC.