Applications of artificial intelligence-based computer-assisted diagnosis in breast radiology: a narrative review

Article information

J Korean Med Assoc. 2025;68(5):281-287

Publication date (electronic) : 2025 May 10

doi : https://doi.org/10.5124/jkma.25.0045

Si Eun Lee, MD, PhD, Eun-Kyung Kim, MD, PhD

Department of Radiology, Yongin Severance Hospital, Yonsei University College of Medicine, Yongin, Korea

Corresponding author: Eun-Kyung Kim E-mail: ekkim@yuhs.ac

Received 2025 March 13; Accepted 2025 April 10.

Abstract

Introduction

Mammography is the standard imaging technique for breast cancer screening. Due to its proven efficacy in reducing mortality, it has been widely adopted as an essential screening tool in numerous countries, as demonstrated by large-scale randomized controlled trials. Despite its relative simplicity and cost-effectiveness, mammography's diagnostic accuracy varies considerably based on individual patient characteristics and the interpreting radiologist's expertise. Notably, approximately 70% of Korean women aged 40 to 59, who constitute the primary screening population, have dense breast tissue [1]. This high density significantly limits mammographic interpretation by obscuring breast lesions. Furthermore, mammographic accuracy varies substantially with radiologist experience; previous research indicates that radiologists attain optimal diagnostic accuracy only after interpreting more than 20,000 mammograms [2]. In clinical practice, however, it is challenging for all radiologists to achieve this level of proficiency. Recently analyzed large-scale national cancer screening data in Korea revealed mammographic sensitivity rates below 70%, implying that approximately 30% of breast cancers are missed by mammography alone [3].

In clinical practice, such limitations can delay the early detection of breast cancer and frequently result in unnecessary recalls and biopsies. Over the past decade, computer-aided diagnosis/detection (CAD) programs have been implemented in mammography to address these shortcomings. Unfortunately, traditional CAD systems have failed to significantly enhance diagnostic performance due to excessive false-positive markings, which have limited their clinical utilization [4]. However, recent advancements in machine learning technologies and big data have facilitated the development of sophisticated artificial intelligence (AI) algorithms to improve breast cancer detection [5]. Various AI programs have been introduced globally, with several undergoing rigorous validation through extensive clinical studies. Some of these validated AI systems have been subsequently commercialized and are increasingly integrated into clinical practice [6]. Recent large-scale prospective studies have demonstrated that these AI programs improve diagnostic accuracy across radiologists of varying experience levels while simultaneously reducing radiologists' workloads by prioritizing mammogram interpretation based on cancer risk [7–10].

This review article aims to discuss methods for improving diagnostic accuracy and efficiency in breast cancer screening mammography using AI-based computer-assisted diagnosis (AI-CAD), and to consider the implications of AI-CAD for clinical practice within the Korean healthcare system.

Improvement in Diagnostic Accuracy

Currently, breast cancer screening through mammography relies entirely upon radiologists' interpretations. However, even experienced radiologists inevitably encounter cases of missed subtle lesions (false negatives) or recommend additional examinations based on abnormal findings that ultimately prove non-cancerous (false positives), especially in populations like Korea, where dense breasts predominate. Given that the primary purpose of screening is early cancer detection, emphasis is placed on achieving high sensitivity. Nevertheless, certain occult cancers remain undetectable by mammography, and some malignancies cannot be clearly distinguished from benign lesions. According to a recent Korean national cancer screening report (2009–2020 data) published in 2025, mammography demonstrated sensitivities ranging from 64% to 69%, and specificities from 82% to 89% [3].

In recent years, highly sophisticated AI algorithms have demonstrated diagnostic accuracy comparable to expert radiologists, effectively increasing cancer detection rates (Figures 1, 2) [11–13]. As of 2024, two AI-CAD systems have received regulatory approval in Korea. Notably, the program developed by Lunit Inc., first approved domestically in 2019, has been predominantly utilized. A pivotal 2020 reader study showed statistically significant improvements in diagnostic accuracy among 14 participating radiologists when aided by AI [14]. The accuracy enhancement was particularly pronounced among radiologists who did not specialize in breast imaging [14]. Traditional CAD systems previously faced challenges due to excessive false-positive markings, adversely affecting efficiency and user satisfaction. In contrast, contemporary AI-CAD maintains high sensitivity with significantly fewer false positives [15]. Nonetheless, variations persist in the performance of commercially available AI-CAD programs, depending on differences in the quantity and quality of training data and underlying algorithms [16].

Figure 1.

A 48-year-old patient underwent screening mammography (A), which revealed a small mass in the left upper central breast detected by artificial intelligence-based computer-aided diagnosis (B). This finding corresponded to a 7 mm suspicious nodule located at the 12 o'clock position in the left breast on ultrasound (US) examination (C, arrow). The lesion was subsequently confirmed as invasive ductal carcinoma via US-guided biopsy and surgery.

Figure 2.

A 58-year-old patient underwent screening mammography (A), which demonstrated a small mass in the right upper outer breast detected exclusively by artificial intelligence-based computer-aided diagnosis (B). This finding corresponded to a 3 mm suspicious nodule located at the 10 o'clock position in the right breast on ultrasound (US) examination (C, arrow). The lesion was confirmed as invasive ductal carcinoma by US-guided biopsy and subsequent surgical excision.

Prospective validation is required to confirm reader-study outcomes. Recently published large-scale prospective European studies have demonstrated increased cancer detection rates without elevated recall rates when AI-CAD results were integrated into mammographic interpretations [7–9]. Similarly, in March 2025, results from a Korean prospective study involving 24,543 participants undergoing mammography at six university hospitals indicated significantly higher cancer detection rates when AI-CAD results supplemented radiologist interpretations. However, while recall rates remained unchanged among radiologists with over ten years of mammographic interpretation experience, less experienced radiologists showed increased recall rates when referencing AI-CAD findings [10]. Follow-up outcomes over two years are forthcoming.

Improvements in Diagnostic Efficiency

In screening mammography, fewer than 1% of cases result in a breast cancer diagnosis, meaning most radiologists' efforts are directed toward evaluating normal mammograms. As screening utilization increases and radiologist staffing remains constrained, improving interpretative efficiency using AI-CAD becomes critical.

Unlike Korea or the United States, where single radiologist interpretation is standard, many European nations practice double reading, involving consensus or arbitration by a third radiologist when discrepancies arise. Recent European prospective studies demonstrated that replacing the second reader with AI-CAD resulted in a 4% to 17.6% higher cancer detection rate than standard double reading, reducing the total reading volume by more than half [7,8]. Given superior performance regarding cancer detection and recall rates, AI-CAD is expected to see broad clinical adoption in Europe.

However, in single-reader environments like Korea, reducing interpretation volume is challenging unless AI findings replace radiologist reviews entirely—a scenario currently impractical. Nevertheless, referencing AI-CAD can reduce reading times, notably in digital breast tomosynthesis, which inherently requires more interpretation time than traditional mammography [17]. Maximizing efficiency requires seamless integration of AI-CAD results within Picture Archiving and Communication System (PACS), minimizing additional effort and facilitating rapid prioritization of higher-risk images. Some modern PACS directly integrate AI-CAD scores, enabling radiologists to prioritize efficiently and allocate interpretive resources effectively (Figure 3).

Figure 3.

Artificial intelligence-based computer-aided diagnosis embedded in the Picture Archiving and Communication System. Before opening an examination, radiologists can view AI-generated abnormality scores directly from the worklist interface (highlighted by squares). These scores can be sorted in ascending or descending order, facilitating prioritization and efficient interpretation of mammograms.

Prediction of Breast Cancer Risk

Historically, breast cancer risk prediction models have relied on clinical factors such as family history, hormonal influences, and prior biopsies. Mammographic breast density has emerged as a significant risk factor and has been incorporated into prediction models, enhancing their predictive accuracy [11]. Breast density, typically classified into four categories through subjective assessment according to the Breast Imaging Reporting and Data System, often shows limited intra- and inter-observer agreement [18]. AI-CAD offers consistent, quantitative breast density assessments and potentially identifies additional mammographic features indicative of breast cancer risk beyond density alone [1].

Yala et al. [19] recently applied an AI model to mammograms, achieving significantly higher predictive accuracy than traditional methods (e.g., Tyrer-Cuzick model version 8). Although the specific features leveraged by AI algorithms remain unclear, broader imaging characteristics, such as breast asymmetry, are presumed to contribute substantially [20]. Considering the prolonged latency of breast cancer, AI risk prediction encompasses both the identification of subtle early lesions and the estimation of future cancer development. Indeed, retrospective analyses revealed rising AI-CAD scores in breast imaging preceding clinical cancer diagnoses, suggesting predictive capability [21,22]. Research continues toward personalized screening strategies informed by AI-generated risk assessments, potentially guiding supplementary imaging or individualized screening intervals [23].

Clinical Implementation in Korea

Hospitals may purchase AI-CAD software approved as diagnostic aids by the Korean Ministry of Food and Drug Safety, integrating it into existing PACS or using dedicated programs. According to a June 2023 Korean survey, 95% of breast radiologists reported experience with AI-CAD, with actual usage likely increasing further [24]. The principal barrier remains associated costs. Notably, in the third quarter of 2024, AI-CAD in mammography was temporarily approved by the Korean Ministry of Health and Welfare’s National Evidence-based Healthcare Collaborating Agency as an new medical technology under defered evaluation, allowing hospitals to offer it as a non-reimbursed service for up to 5 years. Clinical evidence generated during this period will inform potential future insurance reimbursement decisions.

As temporary non-reimbursement alleviates immediate financial concerns, AI-CAD usage is expected to increase significantly. Consequently, clinical practice must address pragmatic considerations. Radiologists remain legally responsible for diagnostic interpretations, necessitating caution and clarity when discrepancies between AI-CAD and radiologist interpretations arise. Proper documentation noting the use of AI-CAD in clinical reports is advisable.

Establishing reliable program performance and building physician trust are prerequisites for broader clinical application. Continuous monitoring of system performance and developing guidelines for appropriate clinical use are crucial for successful integration.

Conclusion

AI-based diagnostic assistance programs trained on big data significantly improve diagnostic accuracy and efficiency in mammographic interpretation. Their ability to improve cancer detection rates while maintaining acceptable false-positive rates justifies their integration into clinical mammography workflows. As AI-CAD continues to evolve, attention must shift from purely technical advancements toward examining clinical utility and optimizing integration strategies.

Article information Continued