Optimizing mammography computer-aided detection: from bilateral learning to single-view inference

Breast cancer remains a significant global health concern, affecting millions of women worldwide. Globally, one in eight women are affected by breast cancer, underscoring its status as a major health issue [1]. Early detection plays a crucial role in improving prognosis, with many countries implementing regular screening programs. Mammography has proven to be an effective tool in detecting breast cancer at its early stages [2]. In Japan, The Ministry of Health, Labour and Welfare of Japan, for instance, recommends biennial mammography screenings for women aged 40 and above, recognizing its importance in identifying breast cancer early when it is most treatable [3]. Mammography interpretation relies on two primary approaches: single image and bilateral image reading. Standard screening protocols typically involve two-view (mediolateral-oblique, MLO) or four-view (MLO and craniocaudal, CC) imaging. These methods allow radiologists to enhance detection accuracy through bilateral comparison, which is crucial for assessing architectural symmetry and identifying focal asymmetric densities (FADs) [4], also known as Breast Asymmetry. While FADs are important diagnostic indicators, they often lead to increased recall rates and significant variability in radiologists' interpretations. Prior research has shown that 27% of missed or misinterpreted cases in mammograms were due to asymmetric density, with FADs being the second most frequently overlooked feature in interval cancers [5]. However, comprehensive four-view mammography screenings are infeasible in all clinical scenarios. For example, approximately 46% of patients diagnosed with breast cancer choose partial mastectomy [6]. In patients who have undergone this surgery, part of one breast has been excised, making it difficult to compare the left and right breasts. While many studies have performed better accuracy using bilateral images, these methods struggle with unilateral imaging scenarios. This study proposes an innovative method to enhance the detection of Mass-like lesions (e.g., mass, architectural distortion, and FADs) in mammography screening, with a particular focus on FADs. Our approach involves training a model using bilateral images but applying it for inference on single-view images (B-SI). This novel technique aims to maintain a pseudo-detection capability and identify features characteristic for bilateral differences, even when only unilateral imaging is feasible. By developing a model that accounts for situations where only single-side mammography is possible, we seek to improve the detection of various breast abnormalities, especially FADs, in a wider range of clinical situations. This research addresses a critical gap in current mammography practices and has the potential to significantly impact breast cancer screening efficacy.