Breast cancer is the most common cancer in women, and one main cause of cancer deaths (1, 2). Mammography, ultrasound, and magnetic resonance imaging (MRI) are well-established diagnostic modalities, which are known to reveal different aspects of underlying abnormalities and provide complementary information for diagnosis (3, 4). Dynamic contrast-enhanced MRI (DCE-MRI) can assess angiogenesis (5, 6), which is essential for cancer development and progression (7, 8). The high spatial resolution and 3D imaging capability of MRI allow for detecting early small cancers, and for evaluating the extent of the disease for pre-operative staging and treatment planning. However, some benign diseases may show strong contrast enhancements and lead to a false positive diagnosis (9).

Mammography can detect breast cancer based on the presence of mass, microcalcifications, architectural distortion, or asymmetric density. It is a widely used imaging modality for screening and diagnosis, and crucial for detecting breast cancer at an early, curable, stage to decrease mortality (10). However, mammography is limited by breast density, which may compromise the detection sensitivity. For women with a high-risk of developing breast cancer, the screening is recommended to start from a young age, and to mitigate the problem of high density in mammography MRI is commonly used as a supplementary modality. Since different imaging can evaluate different pathological characteristics of the abnormal tissue, combining them may improve the diagnostic accuracy (3). MRI is also commonly used for problem-solving when other imaging shows equivocal findings. For example, in patients with category 4 mammographic microcalcifications, MRI can decrease false positive findings and unnecessary biopsy (11).

Breast Imaging Reporting and Data System (BI-RADS) (12) 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 (13). For MRI, intra-/inter-observer agreement was particularly worse for non-mass enhancement compared to mass lesions (14, 15). To circumvent this problem, computer-aided diagnosis (CAD) systems have been proposed to develop quantitative models that are not subject to high variations to serve as potential diagnostic tools (16, 17).

Artificial intelligence (AI) based radiomics study has been widely applied for medical applications. The method allows for high-throughput extraction of quantitative features from radiographic images (18), and it has been shown as a feasible approach for diagnosis of breast cancer using mammography (19–22) and MRI (23–25). However, the combined model using different imaging modalities was rarely reported. Features from corresponding lesions on each modality can be extracted, and then combined in the selection process to develop better models based on their complementary information.

The purpose of this study was to evaluate the diagnostic performance of radiomics models built based on DCE-MRI and mammography. The motivation was coming from the high false positive diagnosis of contrast-enhanced benign lesions commonly seen on MRI. It is anticipated that the complementary information provided by the radiomics analysis of the lesion on mammography may help to improve the diagnostic accuracy. In mammography, features extracted from the lesion and the margin were used to build separate models. The complementary role of MRI and mammography was first evaluated by the selected features, and then by comparing the performance of final models built using each modality alone and in combination.

In this study, we developed the radiomics models for diagnosis of breast cancer using DCE-MRI alone, mammography alone, and the combined MRI and mammography. While quite a few studies have reported the radiomics models developed using MRI (23, 24, 34) or mammography (19–22), the combined analysis was rarely reported (35). We further investigated the complementary role of MRI and mammography features in diagnostic sensitivity and specificity. In the training set, the combined model (89.6%) had a higher accuracy than individual ones (83.2% for mammography, 77.2% for mammography). When mammography features were added to MRI features, it could significantly improve specificity from 69.6% (39/56) to 82.1% (46/56); and thus, have the potential to decrease unnecessary biopsy. Interestingly, the sensitivity was also improved, so the higher specificity was not at the expense of compromised sensitivity. Similar findings were seen in the testing set, with slightly lower overall accuracy from 89.6% to 83.3%.

For mammography, we further separated the analysis using features extracted from the lesion-ROI alone, and from the margin-ROI alone by using a bandshell. The results showed that the accuracy was much better for the lesion model than the margin model, but the margin information could help to improve the accuracy. The results were consistent with the knowledge that margin plays an important role in characterization of a lesion for diagnosis.

Since MRI is more expensive than mammography, the most established clinical indication is for pre-operative staging and high-risk screening. It is not always included in the standard diagnostic workup. It has been shown that in the mammography 4 category, particularly in non-palpable lesions presenting only with microcalcifications, MRI can be used to reduce false positives and avoid unnecessary biopsy (11, 36, 37). On the other hand, benign lesions may show enhancements on MRI, and the information from mammography may help to rule out malignancy (38). As in the case examples shown in Figures 4 and 5 , the benign lesions might be inconspicuous on mammography and had low BI-RADS score of 2, and we had to use MIP generated from MRI as a reference to locate them. Since MRI and mammography evaluate different aspects of the underlying pathology, they should be reviewed together to determine which information needs to be weighted more.

Radiomics is becoming an active research field in breast cancer diagnosis. Due to the large number of images acquired using different MR sequences, radiomics provides an efficient analysis method to extract information. Therefore, more MRI radiomics studies were reported than ultrasound, mammography, and 18F FDG PET/CT (34). MRI radiomics was shown to provide better discrimination than conventional parameters for the diagnosis of breast cancer (23, 24). Mammography radiomics analysis has also been performed in several diagnostic studies (19–22). However, since the patient cohort is different, the diagnostic accuracy will be highly dependent on the inclusion/exclusion criteria, and not directly comparable among studies. Mao et al. (19) used four modeling algorithms, including SVM, naive Bayes classifier, k-NN classifier, and logistic regression to differentiate between benign and malignant cases, and showed a high vibration of 0.629-0.978 in the obtained accuracy. The radiologists’ reading accuracy was 0.772. Lei et al. (20) applied radiomics to diagnose patients showing BI-RADS 4 calcifications on mammography, and achieved AUC of 0.80 in the validation cohort. For characterizing microcalcifications, since the lesion area was not well-defined, the ROI drawing will affect the extracted features, and thus, the diagnostic results. Huang et al. (21) applied mammography radiomics for distinguishing male malignant and benign lesions, and reported an AUC of 0.82 – a very unique study in rarely reported male patients. Another study by Niu et al. (22) also analyzed patients showing abnormal lesions on mammography and MRI, close to our patient cohort, but their goal was to evaluate the combined effect of mammography and digital breast tomosynthesis (DBT), as well as the combined effect of DCE and diffusion weighted MRI. The reported accuracy based on the mammography was close to ours, around 0.70. Multi-modal radiomics combining different imaging modalities are rarely reported. In a study by Chen et al., the multimodal classifier achieved a better diagnostic performance than any single modality (35). Since each imaging modality is unique in its acquisition method and parameter setting, the extracted features from a lesion may be different and provide complementary information to improve diagnostic accuracy.

In this study, the cases were identified from the MRI database first, and then only those with mammography performed within one month were further selected for analysis. All lesions showed strong enhancements on MRI, and the information was used to determine a corresponding ROI on mammography. Co-registration of MRI and mammography to ensure that the traced ROI is indeed coming from the same suspicious tissue is not a trivial task. We used maximum intensity projection of MRI as guidance, and it could be projected from different angles to simulate CC view and MLO view to guide the tracing of the suspicious tissues on mammography. Some computer techniques have been proposed for registration between MRI and mammography, e.g., using finite element methods by Hopp et al. (39) and Mertzanidou et al. (40), and the thin-plate spline method by Yang et al. (3). These registration techniques can be considered in future multi-modality radiomics studies. However, since the mammography was acquired using heavily compressed breast tissues in a different body position, it might be difficult to find the precise correspondence. Therefore, in this study we only analyzed the CC or MLO view that had more clear presentation of the lesion.

There were several limitations in this study. First, the models were developed using a dataset from a single institution. The earlier cases were used for training, and the performance was evaluated using 10-fold cross-validation. We assembled an independent testing set using later cases according to time of enrollment, so the developed models from training can be independently tested. Another limitation is that the sample size was relatively small. In our dataset, all benign lesions had to show visible enhancements on MRI and were histologically confirmed, which were very strict criteria and limited the number of eligible cases. However, since the major goal of this study is to investigate whether and how much the addition of mammography radiomics features can complement MRI, using a strict rule to identify eligible cases with histologically confirmed lesions is needed. Third, while all lesions showed enhancements on MRI, lesions not visible with the MRI-guidance on mammography were not included in this study. Since the boundary of these lesions could not be clearly defined, the radiomics features might not be reliably extracted.

In conclusion, the radiomics models built based on combined MRI and mammography had better diagnostic accuracy than models built using single modality alone. The combined model could reach the accuracy of 89.6% in the training and 83.3% in the testing sets. The motivation of this study is to use the complementary information extracted from radiomics analysis of the lesion shown on mammogram to decrease the false positive diagnosis of contrast-enhanced benign lesions on MRI. In the western countries, breast MRI is recommended as a clinical modality for screening of women with a high risk of developing breast cancer, and the false positive diagnosis in a screening population will lead to many unnecessary procedures including biopsy, and patient anxiety. Our study may provide a helpful computer-aided diagnostic tool for such clinical indications. The multimodality radiomics analysis by combining mammography and MRI features has the potential to improve the specificity and reduce unnecessary biopsies, while maintaining a high sensitivity for diagnosis of breast cancer.

The datasets used and analyzed in this study will be made available by the corresponding author on a reasonable request.

The studies involving human participants were reviewed and approved by Ethics Committee of The First Affiliated Hospital of Wenzhou Medical University. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

Study concept and design: Y-FZ and M-YS. Acquisition of data: ZC, JZ, and MW. Analysis of data: YZ, ZC and KL. Drafting of the manuscript: Y-FZ, J-HC, FC, RP, and RM. Critical revision: MW and M-Y S. Statistical analysis: YZ and ZC. Study supervision: MW and M-Y S. All authors contributed to the article and approved the submitted version.

This work was supported in part by Foundation of Wenzhou Science & Technology Bureau (No. Y20180185), Medical Health Science and Technology Project of Zhejiang Province Health Commission (No. 2019KY102), Research Incubation Project of First Affiliated Hospital of Wenzhou Medical University (No. FHY2019085), the National Cancer Institute of the National Institutes of Health under award number P30 CA062203, R01 CA127927, R21 CA208938 and the UC Irvine Comprehensive Cancer Center using UCI Anti-Cancer Challenge funds.

The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health or the Chao Family Comprehensive Cancer Center.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.