In 2020, female breast cancer (BC) became the most common type of cancer with an estimated 2.3 million new cases (11.7%), followed by lung cancer (11.4%) (1). Treatment of BC relies on conducting an accurate diagnosis, including histological, molecular, and clinical phenotypes. Non-invasive imaging techniques such as mammography, ultrasound, and magnetic resonance (MR) are available for qualitative and quantitative analysis of BC in clinical practice. The American College of Radiology Breast Imaging Reporting and Data System (BI-RADS) is a standardized assessment structure that enables radiologists to clearly and concisely communicate results of breast imaging to referring physicians (2). In the fifth edition of the BI-RADS atlas (3), category 4 and 5 breast lesions are defined as suspicious cancerous lesions, and a biopsy is recommended for further diagnosis. Recent studies have shown that a large number of benign lesions are present in category 4 and 5 breast lesions, particularly in the mammography reporting system, exposing these patients to invasive biopsies (4–6). Depending on the technique, the sensitivity values of biopsy results ranged from 87% to >97% (7–9).

Radiomics is a high-throughput image mining technique that aims to enhance the predictive power of medical images by quantifying the morphology, intensity distribution, and texture patterns of lesions. Recent investigators have examined the role of mammography, ultrasound, and MR radiomics in the prediction of molecular subtypes (10–12), lymph node metastasis (13–15), response to neoadjuvant chemotherapy (16–18), recurrence risk (19, 20), and disease-free survival (21, 22) of BC. However, with the initiative of precision medicine (23–25), the reduction in overdiagnosis and overtreatment of breast lesions through non-invasive radiomics method is also a topic worth investigating.

There are four main findings of breast lesions on diagnostic mammography images: masses, calcifications, architectural distortion, and asymmetries. One large sample study (26) showed that BC most often presented as mass at 56%, followed by calcifications at 29%, asymmetry at 12%, and architectural deformities at 4%, and another small sample study (27) suggest that approximately 50% of breast lesions presenting as a mass were ultimately confirmed benign lesions. In addition, mass may be the only finding or one of the combined findings of breast lesions in mammography (we defined these masses as mammographic masses). Although experienced radiologists have a high diagnostic accuracy in identifying benign and malignant mammographic masses, less experienced radiologists sometimes make excess errors (28) such as benign mammographic masses are over-classified as BIRADS category 4 or even 5.

A previous study has shown that combining both craniocaudal (CC) and mediolateral oblique (MLO) positions radiomics data had good classification performance between HER2-enriched BC and non-HER2-enriched BC (11). Here, we combined both CC and MLO positions radiomics data aimed to explore a model with a sensitivity more than 98% for the characterization of BI-RADS category 4 and 5 mammographic masses, thereby achieving a reduction in biopsies of benign lesions at a very low rate of missed malignant lesions.

In this study, we built a radiomics model with similar sensitivity to biopsy for predicting malignancy of BI-RADS category 4 and 5 mammographic masses by using the combination of CC and MLO position images from DM. In both training and test sets, the radiomics model obtained specificity by over 60% while preserving sensitivity more than 95.8%. Both AUC and sensitivity were relatively stable, while the specificity was not so stable. These experimental results suggest that the non-invasive imaging radiomics method could achieve similar sensitivity to biopsy while avoiding some benign mammographic masses to undergo unnecessary invasive biopsy.

We are aware of several papers that use mammographic radiomics data to differentiate between benign and malignant breast lesions (30, 36–38), and we are aware that several of these papers incorporated morphological features (36, 37), and several that did not (30, 38). However, most breast lesions that present alone as calcification, architectural distortion, or asymmetries tend to lack distinct grayscale contrast boundaries, which compromises the masking of tumors. Besides, non-invasive imaging techniques with similar biopsy sensitivity are important for differentiating between benign and malignant BI-RADS category 4 and 5 breast lesions for which invasive biopsy was already indicated. To the best of our knowledge, this is the first study to use mammographic radiomics data to predict the malignancy of BI-RADS category 4 and 5 mammographic masses with a sensitivity similar to that of a biopsy, and we believe that the radiomics model is more limited but more relevant.

When using radiomics data to predict the malignancy of breast lesions, commonly used mammography images include CC position alone (37, 39), mixed CC and MLO position (30, 36), and combination of CC position and MLO position (40, 41). Gupta et al. demonstrated that the corresponding first-order and texture features of mammographic masses between CC position and MLO position were not strongly correlated (42), suggesting that the inclusion of first-order and texture features from multiple mammographic positions may impact the accuracy of diagnosis of mammographic masses. Ma et al. has shown that by combining both CC and MLO position, radiomics data had higher classification performance between HER2-enriched BC and non-HER2-enriched BC than using CC position alone and MLO position alone (11). Hence, we applied both CC and MLO position radiomics data to predict the malignancy of mammographic masses suspicious for cancer, and the results showed that this method had good classification performance.

Recent studies have shown that the use of a single random training–test set split may lead to unreliable results in small sample radiomics machine learning studies (43). In our study, we divided the training and test sets based on the chronological order in which patients underwent mammography examinations, and we performed random fivefold cross-validation in training and test sets, respectively, to assess the stability of the radiomics model. Although the specificity gives a large error (12.5%, 8.6%), the AUC (0.038, 0.062) and sensitivity (3.9%, 4%) were relatively stable. This is consistent with previous findings that cross-validation may lead to large error bars for small sample sizes (44).

Of the 188 mammographic radiomics features, the feature selection method screened out 14 useful features in the training set (n=192, ratio 14:1). As suggested by Gillies et al., each radiomics feature requires at least 10 samples to support in a classifier (45). Previous publication has shown that maximum 2D diameter is a useful feature for predicting BC (37). In this study, we found a significant positive correlation between maximum 2D diameter and gray-level size zone matrix-based gray level non-uniformity in the CC position images (r=0.82, p<0.001) and MLO position images (r=0.924, p<0.001) of the training set ( Figure 5 ). Thus, the maximum 2D diameter feature was removed when filtering features in the Pearson’s correlation analysis. It is worth noting that although the size feature was not included in our radiomics model, we do not consider it unimportant in predicting malignancy of mammographic masses suspicious for cancer.

Of the 14 useful mammographic radiomics features, perimeter-to-surface ratio had the greatest weight value of 1.234, suggesting that morphological features remain important in identifying benign lesions or malignancies for mammographic masses suspicious for cancer. The mean density obtained moderate weight values of 0.389 and 0.079, which was higher in the malignant group than in the benign group, indicating that the malignant mammographic masses tend to be higher density than benign mammographic masses. In addition, the malignant group had higher value for inverse difference normalized, zone variance, gray level non-uniformity, and busyness, and lower value for uniformity, skewness, coarseness, small-dependence high gray-level emphasis, and large-dependence low-gray level emphasis, indicating more heterogeneity of malignant mammographic masses in the density and texture patterns (31) ( Figure 6 ). This experimental result is consistent with the pathological fact that malignant tumors tend to be heterogeneous, while benign lesions tend to be homogeneous (46).

Admittedly, our study has several limitations. First, mammographic radiomics data for this study were collected from a single center with a limited number of participants, and further multicenter testing is needed. Second, the mammographic masses were masked by manual method; however, good inter-observer reproducibility was obtained in feature extraction and model diagnosis. Some publications indicate that semi-automatic segmentation method had higher inter-observer reproducibility (47, 48). Further work is needed to construct a semi-automatic segmentation method for mammographic masses. Finally, the radiomics model was constructed by 2D images, which may lose some important information of mammographic masses. However, some publications showed that 2D radiomics features had higher classification performance than 3D radiomics features in lung cancer (47, 48).

In conclusion, a mammographic radiomics model combining both CC and MLO position images had excellent sensitivity and moderate specificity in differentiating malignancies and benign lesions for BI-RADS category 4 and 5 mammographic masses. It may help reduce the temporarily unnecessary invasive biopsies for benign mammographic masses while providing similar diagnostic performance for malignant mammographic masses as biopsies.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by the Ethics Committee of Xiang’an Hospital of Xiamen 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: GW and DS. Acquisition of data: HZ, QG, and SW. Analysis of data: GW, DS. Drafting of the manuscript: GW, DS, HZ, and SW. Critical revision: QG and KR. Statistical analysis: GW and DS. Study supervision: QG and KR. All authors contributed to the article and approved the submitted version.

This work was supported by the Scientific Research Foundation for Advanced Talents, Xiang’an Hospital of Xiamen University (no. PM201809170011).

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.