Introduction

Mammography and dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) are commonly used clinical imaging modalities for diagnosis of breast lesions, which are known to reveal different aspects and provide complementary information for improved accuracy [1, 2]. However, they are limited by high false positive rate and low sensitivity, respectively. Since different imaging examination methods can evaluate different pathological characteristics of the abnormal tissue, combining them may improve the diagnostic accuracy [1]. Breast Imaging Reporting and Data System (BI-RADS) [3] is used to indicate the level of suspicion in detected abnormality. However, subjective reading using the BI-RADS lexicon only achieved moderate levels of inter-reader agreement [4]. To circumvent this problem, computer-aided diagnosis systems have been proposed to develop quantitative models that are not subject to high variations as potential diagnostic tools [5, 6]. Radiomics allows high-throughput extraction of numerous quantitative features from radiographic images [7]. It has become a very popular machine learning-based method for medical applications. Artificial intelligence (AI) based radiomics study has been applied for diagnosis of breast cancer and proven as a feasible approach for mammography [8] and MRI [9, 10]; however, combined radiomics models using both was rarely reported. Therefore, the purpose of this study was to evaluate the performance of radiomics models built based on DCE-MRI and mammography for diagnosis of breast cancer, compared to results determined by BI-RADS reading.

Methods

A total of 202 patients who had both DCE-MRI and mammography for diagnosis were retrospectively reviewed, including 146 malignant and 56 benign lesions, all confirmed by histopathology. DCE scan was acquired using the volume imaging for breast assessment sequence, consisted of 6 frames: one pre-contrast (F1) and 5 post-contrast (F2-F6). BI-RADS score was obtained from the radiology reports, classified into 2, 3, 4A, 4B, 4C, and 5. Lesions were segmented based on contrast-enhanced maps using FCM clustering algorithm on each DCE slice containing the lesion. Then, the ROIs from all slices were combined, and 3D connected-component labeling and the hole-filling algorithms were applied to generate the final 3D mask [11, 12]. For corresponding mammography, two experienced radiologists manually outlined the lesion on cranio-caudal or medio-lateral oblique view by consensus, with the guidance of the lesion shown on the maximal intensity projection of MRI, projected from different angles. Three heuristic DCE parametric maps were generated according to: the early wash-in signal enhancement (SE) ratio [(F2-F1)/F1]; the maximum SE ratio = [(F3-F1)/F1]; the wash-out slope [(F6-F3)/F3] [9]. Four cases are illustrated in Figures 1 to 4. The radiomics analysis was performed using the PyRadiomics on heuristic DCE parametric maps and mammography. A sequential forward feature selection method based on support vector machine was used for optimal feature subset selection. After feature selection, five models were built based on the selected features from: 1) DCE-MRI; 2) mammography lesion; 3) mammography margin; 4) mammography lesion+margin; and 5) all combined. The diagnostic performances were tested using 10-fold cross-validation. The accuracy predicted by different models was compared by using the Chi-square test.

Results

For individual models, 8 DCE-MRI, 4 mammography-lesion, and 3 mammography-margin features were selected, respectively. In the combined model, a total of 9 features including 7 DCE-MRI and 2 mammography-lesion were selected. The achieved accuracy using 10-fold cross-validation was 83.2%, 75.7%, 64.4%, 77.2%, 89.6% for the model built based on DCE-MRI, mammography lesion, mammography margin, combination of mammography lesion and margin, all features combined, respectively (Table 1). For BI-RADS, The final diagnostic results evaluated using the cut-off of BI-RADS 4A or 4B are shown in Table 2, showing higher sensitivity and lower specificity based on 4A than 4B, but the overall accuracy was comparable, 82-83% for MRI, and 79-82% for Mammography. The combination model based on all features (89.6% accuracy) performed significantly better than the BI-RADS reading using 4A or 4B as threshold on MRI (82.7% p=0.044, 83.2% p=0.059) and mammography (82.2% p=0.032, 79.2% p=0.004).

Discussion

Although BI-RADS is a well-established scoring system for assessment of breast lesions, it is limited by moderate inter-reader agreement, and also that the accuracy depends on the experience level of the reader. In this study, MRI was found to have a higher sensitivity and lower specificity compared to mammography, which was consistent with their clinical role. When using the highest BI-RADS score between MRI and mammography to assign to each lesion, the specificity was further decreased, lowing overall accuracy. Since both imaging modalities are commonly performed for diagnosis, there is a wealth of data that can be explored using the AI based image analysis methods. We built 5 different radiomics models to investigate the features extracted from MRI, lesion and margin of Mammography, and all of them combined. The model based on all features yielded an accuracy of 89.6%, which was higher than all other models based on DCE-MRI or mammography alone, suggesting their complementary role. The results demonstrated the feasibility and effectiveness of the AI analysis using multimodal imaging to improve the diagnostic accuracy for breast cancer.

Acknowledgements

This work was supported in part by Foundation of Wenzhou Science & Technology Bureau (No. Y20180187 and Y20180144), Medical Health Science and Technology Project of Zhejiang Province Health Commission (No. 2019KY102), and NIH/NCI R01 CA127927 and R21 CA208938.