Radiomics of contrast-enhanced spectral mammography for prediction of pathological complete response to neoadjuvant chemotherapy in breast cancer

2.1 Patients

This study was approved by the institutional review board of Yantai Yuhuangding Hospital, and the patients right to informed consent was waived.

A total of 196 patients with locally advanced breast cancer who had underwent CESM examination before NAC treatment from March 2018 to January 2021 were enrolled. The inclusion criteria were as follows: (i) the patients had biopsy-confirmed unilateral primary invasive breast cancer before NAC; (ii) the patients received no treatment other than NAC and completed the full course of NAC; (iii) surgery was followed by the completion of NAC, and postoperative pathology was obtained; and (iv) the CESM examination was conducted 2 weeks before NAC. The exclusion criteria were as follows: (i) multifocal, bilateral, or occult breast cancer; (ii) lack of pre-treatment CESM image or clinical data; (iii) the patients who received non-standard treatment or did not complete the NAC regimen; (iv) no surgery was performed or the results of postoperative pathological immunohistochemistry (IHC) were incomplete; and (v) the image quality of CESM is insufficient to obtain measurement results. Finally, 118 patients were included in this study and randomly divided into a training set (n = 82, included 21 pCR and 61 non-pCR) and a test set (n = 36, included 9 pCR and 27 non-pCR) in a ratio of 7:3. We also repeated this training/testing process 6 times to prevent biases from in case partition. The screening diagram of the study population is shown in Fig. 1.

OPEN IN VIEWER

Fig. 1

Flow chart of patient enrolment.

The CESM used in our breast center is based on a full-digital breast machine (Senographe DS SenoBright, GE Healthcare). The contrast agent Omnipaque 300 (GE Healthcare, Inc., Princeton, NJ) is injected through the patient’s upper arm vein at a dose of 1.5 mg/kg and a flow rate of 3.0 mL/s. About 2 min after the contrast agent is injected, the bilateral breasts are pressed in the cranial caudal (CC) view and mediolateral oblique (MLO) view. Moreover, low- and high-energy images are rapidly acquired. Usually, the contralateral mammary gland is photographed first and then the suspicious side is photographed. Bilateral breast CC and MLO views are photographed at low- and high-energy exposures. Low- and high-energy exposures are continuously performed within 1.5 s, and then recombined images are obtained using a post-processing system. The low-energy image is similar to a conventional mammogram, whereas the recombined image shows the contrast medium uptake [13].

2.3 Clinical and histopathological characteristics

Clinical and histopathological data were obtained from the medical record system, including age, carcinoembryonic antigen (CEA), carbohydrate antigen 153 (CA153), carbohydrate antigen 125 (CA125), and molecular subtype (according to the status of estrogen receptor [ER], progesterone receptor [PR], epidermal growth factor receptor-2 [HER2], and proliferation marker Ki-67). Immunohistochemical (IHC) staining shows the stained tumor cell nuclei:< 1% positive tumor cells mean ER/PR negative and≥1% positive tumor cells mean ER/PR positive [23]. The cutoff level for Ki67 is 20%. The IHC results for HER2 status were interpreted as follows:+3 was defined as positive, 0 to+1 was defined as negative, and+2 required a fluorescence in situ hybridization (FISH) of the biopsied tissue for validation. The non-amplified FISH results were considered HER2–, whereas the amplified FISH results were defined as HER2 + [24]. Tumors were stratified into molecular subtypes using IHC surrogate markers. Luminal A was defined as ER positive, PR (20%) positive, HER2 negative, and Ki67 < 20%. Luminal B was defined as ER positive and/or PR (< 20%) positive, HER2 negative or ER/PR/HER2 positive, or ER and PR positive but Ki67 > 20%. HER2 positive was defined as ER/PR negative and HER2 positive. The triple-negative group was defined as ER/PR/HER2 negative.

2.4 NAC and pathological assessment

In accordance with the National Comprehensive Cancer Network (NCCN) guideline and Chinese Society of Clinical Oncology (CSCO) guideline for breast cancer, all patients received six or eight cycles of NAC before breast surgery. The NAC regimens were based on taxane (11 patients, 9.3%) or taxane and anthracycline (107 patients, 90.7%). All HER2-positive patients also received trastuzumab or trastuzumab and pertuzumab. Surgical operations were performed 2–3 weeks after completing NAC.

pCR is defined as the absence of residual infiltrating primary breast cancer (which can be ductal carcinoma in situ) and residual regional lymph nodes after completing NAC [3], while non-pCR refers to patients with partial response or no response, meaning that invasive cancer remains after NAC.

2.5 Region of interest segmentation and radiomic feature extraction

The radiologist 1 (with 9 years of experience in breast imaging) manually segmented the tumor regions of interest (ROIs) by following the outline of the tumor on the low-energy and recombined images of the CC and MLO views on ITK-SNAP software (version 3.8.0). The sample ROI segmentation is shown in Fig. 2.

OPEN IN VIEWER

Fig. 2

Examples of regions of interest (ROIs) segmentation on contrast-enhanced spectral mammography (CESM) images. (A, C) The low-energy and recombined images with cranial caudal (CC) view. (B, D) The low-energy and recombined images with mediolateral oblique (MLO) view. The ROIs of breast lesions were drawn manually on low-energy and recombined images, respectively.

Image preprocessing is required before feature extraction. It includes the standardization of the gray value of the ROI, discretization of gray level, and image resampling. The μ±3σ method is used to standardize the image of the ROI. Image normalization can improve the texture difference features between two image classes [25]. Therefore, normalization was performed first, and the gray value greater than μ±3σ was set to 0. The gray scale of the image was adjusted by gray scale discretization. In this study, 64 gray levels were used to reduce the calculation time and effect of noise [18, 26]. Finally, the image was re-sampled to ensure that the scale of the feature was unchanged, and the anisotropic pixel was converted to 0.1 mm×0.1 mm [27].

Radiomics features were then extracted from each ROI of the low-energy images and the recombined images from CC and MLO views respectively using “RadiomicsFeatureExtractor” by Pyradiomics, including three categories: first-order features (270), shape-based features (14), and texture features (1125).

Intraclass correlation coefficients (ICCs) were used to assess the intra- and inter-observer agreement of ROI feature extraction to evaluate the reproducibility of radiomics feature extraction. The CESM images of 30 patients were randomly selected to calculate ICCs. Radiologist 1 repeated the ROI segmentation after 2 months to evaluate intra-observer agreement. Radiologist 2 (with 13 years of experience in breast imaging) segmented the ROIs manually and extracted features using the same method to evaluate inter-observer agreement. ICC > 0.75 was considered a good agreement, and features with ICCs > 0.75 within the training set were used for further analysis.

2.7 Radiomics feature selection and radiomics score construction

Radiomics features were selected by low variance method, SelectKBest method, and least absolute shrinkage and selection operator (LASSO) in the training set. In the low variance method, the threshold was 0.8, and the eigenvalues of the variance larger than 0.8 were included. In the SelectKBest method, the features with p < 0.05 between NAC pCR and NAC non-pCR were selected. LASSO regression was finally used for selecting the key radiomics features with non-zero coefficients. The 7-fold cross-validation was used to select the optimal tuning parameter (alpha) of LASSO via minimum criteria. Radiomics score (Rad-score) was calculated using a linear combination of the final selected features weighted by their LASSO coefficients. The area under the receiver operating characteristic (ROC) curve (AUC) value of Rad-score was calculated to assess the predictive performance of radiomics features.

2.8 Radiomics nomogram construction and assessment

In the training set, univariate and multivariate logistic regression were used to select independent predictive factors, including clinic radiological characteristics (age, CEA, CA153, CA125, ER, PR, HER-2, Ki67, tumor size, and molecular subtypes) and Rad-score. Factors with p < 0.05 in were selected as independent predictive factors and were integrated into the multivariate logistic regression. Likelihood ratio test with backward step-down selection was applied for logistic regression. With Akaike information as the criterion, when the minimum Akaike information criterion was reached, the step-by-step process was stopped. A radiomics nomogram incorporating Rad-score and independent predictive factors was built by multivariate logistic regression, which could be used as a convenient visible tool to predict the individual probability of NAC pCR. To train the model, we performed a grid search and fivefold cross-validation to find the optimal parameters in the training set and examine the performance of generalization. The accuracy of the model was taken as the indicator for extracting the model parameters. We then trained the model using the entire training set.

The predictive performance of the radiomics nomogram was evaluated by ROC in the training and test sets. In addition, the point on the ROC curve farthest from the diagonal line corresponds to the largest Youden index, which is calculated by the sum of sensitivity and specificity and then subtracting 1 over all the possible threshold values. This index was used to determine the cutoff value for dividing the patients with NAC pCR and NAC non-pCR. Accuracy, sensitivity, and specificity were calculated in the training and test sets according to the cutoff value.

The agreement between the observed results and predictive probability was assessed by calibration curve. Decision curve analysis (DCA) was employed to evaluate the net benefit of the radiomics nomogram under the different threshold probabilities.

2.9 Statistical analysis

The statistical analysis was conducted in Python (version 3.6), R software (version 3.4.1), and SPSS (version 26). The categorical clinical variables (ER, PR, HER-2, Ki67, and molecular subtypes) between training and test sets were compared using the chi-square test or Fisher’s exact test, whereas continuous clinical variables (age, tumor size, CEA, CA153, and CA125) were analyzed by two-sample t test. The “glmnet” and “glm” packages were used for LASSO logistic regression. The “rms” and “pROC” packages were used for nomogram construction and ROC analysis, respectively. The “Calibration Curves” package was used for calibration curve analysis, whereas the “Decision Curve” package was used for DCA. p < 0.05 was interpreted as statistically significant difference.

3.1 Clinical characteristics

A total of 118 patients were finally enrolled in this retrospective study. 82 patients (with 21 pCR, 61 non-pCR) were included in the training set, and 36 patients (with 9 pCR, 27 non-pCR) were included in the test set. The patient characteristics in the training and test sets are shown in Table 1. Significant differences in ER (p = 0.048) and PR (p = 0.002) but no significant differences in age (p = 0.135), diameter (p = 0.624), CEA (p = 0.329), CA153 (p = 0.525), CA152 (p = 0.354), HER2 (p = 0.405), Ki-67 (p = 0.440), and molecular subtype (p = 0.202) were found between NAC pCR and NAC non-pCR in the training set.

A total of 5636 radiomics features were extracted from each patient’s CESM images. The intra- and inter- observer ICCs among 5636 radiomics features were 0.769–0.815 and 0.786–0.853, respectively, which represented good reproducibility in radiomics feature extraction.

Moreover, 4110 radiomics features were selected by low variance method, and 33 radiomics features were further selected by SelectKBest method. Finally, the optimal 13 radiomics features with non-zero coefficients in LASSO logistic regression were selected, whereas the optimal value of alpha was 0.0198, and the optimal –log(alpha) was 1.7 (Fig. 3, Table 2). The Rad-score of each patient was calculated by the 13 radiomics features.

OPEN IN VIEWER

Fig. 3

Lasso algorithm for radiomics features selection. (a) Least absolute shrinkage and selection operator (LASSO) coeffificient profifiles of the 33 features. The y-axis represents coeffificient of each feature. The optimal value of alpha was 0.0198, and the optimal -log(alpha) was 1.7, where 13 features with non-zero coeffificient were selected. (b) Mean square error path using seven-fold cross-validation.

In univariate logistic regression, ER (P = 0.049), PR (P = 0.001), molecular subtypes (P = 0.034), and Rad-score (P < 0.01) were included into multivariate logistic regression. After multivariate logistic regression, Rad-score and PR were remarkably related to pCR after NAC (both P < 0.001). Thus, the radiomics nomogram was developed with the combination of Rad-score and PR as shown in Fig. 4.

OPEN IN VIEWER

Fig. 4

A radiomics nomogram with Rad-score and PR incorporated for predicting the probability of NAC pCR in breast cancer patients.

The radiomics nomogram yielded AUCs of 0.906 (95% CI, 0.840–0.966) and 0.790 (95% CI, 0.554–0.952) in the training and test sets, respectively (Fig. 5). The results showed that the radiomics nomogram has favorable predictive performances. The AUCs of the Rad-score and clinical model were 0.873 (95% CI, 0.808–0.948) and 0.694 (95% CI, 0.581–0.782) in the training set and 0.700 (95% CI, 0.506–0.911) and 0.759 (95% CI, 0.617–0.870) in the test set, respectively. The diagnostic performances of the three models are shown in Table 3. The results revealed that the radiomics nomogram’ performance was better than Rad-score and clinical model in the training and test sets. The average AUC and standard deviation (STD) of the nomogram after repeating training/testing process 6 times are shown in Fig. 6. The results showed a mean AUC of 0.943 and 0.785, and a STD of 0.027 and 0.038, respectively, for the nomogram in the training and test sets. The nomogram was experimentally shown to be robustness.

OPEN IN VIEWER

Fig. 5

Receiver operating characteristic (ROC) curves of Rad-score, clinical model and radiomics nomogram in the training (a) and validation (b) cohorts.

OPEN IN VIEWER

Fig. 6

Receiver operating characteristic (ROC) curves of the nomogram after repeating training/testing process 6 times in the training (a) and test (b) sets.

DCA showed that the radiomics nomogram had net benefit than the “treat all” or “treat none” scheme, whereas the threshold probability ranged from 0 to 0.65 in the test set (Fig. 7). The calibration curves of the radiomics nomogram demonstrated good consistency between predictive outcome and observation in the training and test sets (Fig. 8).

OPEN IN VIEWER

Fig. 7

Decision curve analysis of three models in validation set. The y-axis measures the net benefit. The green line represents the radiomics nomogram. The red line represents the radiomics score. The blue line represents the clinical model. The black line represents the assumption that all patients achieve the PCR after NAC. The horizontal red thin line represents the assumption that no patients reach the PCR after NAC.

OPEN IN VIEWER

Fig. 8

Calibration curves of radiomics nomogram in the training (a) and validation (b) sets. The diagonal line represents the perfect prediction of the radiomics nomogram. The red, blue and green lines represent the calibration curve of rad-score, clinical model, and radiomics nomogram, respectively. The calibration curves of radiomics nomogram are close to the diagonal line both in the training and validation sets, which shows that the prediction probability have good agreement with the actual probability.

Ethics approval and consent to participate

This study was approved by the institutional review board of Yantai Yuhuangding Hospital, and the patients’ right to informed consent was waived.