AI Systems for Mammography with Digital Breast Tomosynthesis: Expectations and Challenges

See also article by Resch and Lo Gullo et al in this issue.

Masako Kataoka, MD, PhD, is a lecturer in diagnostic imaging and nuclear medicine at Kyoto University Graduate School of Medicine and a chief of breast imaging. Her research interests are in breast imaging that includes ultrafast MRI, high-resolution diffusion-weighted imaging, breast PET, and image-based risk prediction. She is an associate editor of Radiology: Imaging Cancer and a board of member of the International Society for Magnetic Resonance in Medicine and the Japanese Society of Magnetic Resonance in Medicine.

Takayoshi Uematsu, MD, PhD, serves as the head of the department of breast imaging and breast interventional radiology at Shizuoka Cancer Center Hospital in Japan. His extensive research encompasses all breast imaging, biopsy, and screening. With more than 30 years of experience in the field, he has amassed a wealth of knowledge on the various modalities and their clinical applications, contributing to the reduction of breast cancer mortality.

Recent advancements in artificial intelligence (AI) have revolutionized medical imaging, leading to a transformative impact on health care. This is particularly evident in radiology, where clinical radiologists are tasked with interpreting a rapidly growing number of complex cases. With several promising future applications of AI in radiology, there is a need for ongoing research, the adoption of innovative imaging technologies, and the cultivation of strong partnerships between radiologists and AI developers. Advances in AI and machine learning, combined with increased computational power and large data sets, have enabled progress in computer-aided detection (CAD) and diagnosis, disease prediction, and image segmentation. Some AI algorithms have even been shown to match the performance of radiologists. Breast imaging is one of the areas of clinical radiology that benefits most from use of AI because it deals with a large number of normal images in the screening setting.

Digital breast tomosynthesis (DBT) is a substantial technical advancement due to its improved diagnostic accuracy compared with conventional full-field digital mammography, as reported in several large-scale trials (1,2). In a meta-analysis, DBT increased sensitivity (84%–90% vs 69%–86%) and cancer detection rate (relative risk, 1.16; 95% CI: 1.02, 1.31) versus digital mammography alone in a diagnostic setting (3).

However, the main drawback of DBT is its prolonged image interpretation time. DBT typically covers one breast with dozens of reconstructed sections, requiring the reader to scroll through these images. Early reports showed that the reading time was twice as long as that of full-field digital mammography (1). Efforts have been made to decrease image interpretation time by adopting slab-reading (4) and computing synthetic mammograms from DBT. Since wider use of DBT is anticipated in the next edition of the Breast Imaging Reporting and Data System (BI-RADS), the development of deep learning–based AI CAD systems is also sought after.

Two AI CAD systems for DBT have already been approved for clinical use in Europe and the United States. Both Transpara version 1.7.0 and ProFound AI 3.0 are deep learning based and have been trained on over 1 million and 6 million images, respectively. The two systems rely on different deep learning models, which were developed using different training sets. However, no head-to-head intraindividual comparison has previously been made available.

The study by Resch and Lo Gullo et al published in this issue of Radiology: Imaging Cancer examines the two AI CAD systems for DBT in breast cancer detection, providing preliminary data on commercially available AI software performance in a real-life clinical setting, including intermethod comparison and comparison with benchmark human interpretation (5). This retrospective study included consecutive asymptomatic patients who underwent mammography with DBT (2019–2020). Both the Transpara 1.7.0 and ProFound AI 3.0 AI systems were used to evaluate the DBT examinations. The systems were compared with one another and with human double-reading using receiver operating characteristic analysis, including mammographic breast density–based subgroup analysis.

An enriched cohort of 419 female patients (median age, 60 years) with 58 histologically proven breast cancer cases was analyzed. The area under the receiver operating characteristic curve (AUC) for Transpara, Profound AI, and human double-reading was 0.86, 0.93, and 0.98, respectively. For Transpara, a rule-out criterion of a score 7 or lower yielded 100% sensitivity and 60.9% specificity, and a rule-in criterion of a score greater than 9 resulted in 96.6% sensitivity and 78.1% specificity. For Profound AI, a rule-out criterion of score less than 51 yielded 100% sensitivity and 67.0% specificity, and a rule-in criterion of score greater than 69 yielded 93.1% sensitivity and 82.0% specificity.

Perhaps the most important results of the study were the relatively excellent performance of the two commercially available software systems and an even better performance of the human double-reading. When the patients were grouped by breast density, ProFound AI performed better for nondense breasts, while both performed similarly for dense breasts. In a screening setting, missed cancers should be avoided while a reasonable number of false-positives may be acceptable. Thus, rule-out criteria with 100% sensitivity are a realistic option.

There are also notable study limitations. Although the authors initially included patients consecutively, the study results should be interpreted with caution as the final cohort was an enriched cohort with 25% cancer cases. In breast cancer screening for average-risk individuals, a less than 1% cancer detection rate is expected. Also, the current study sample included a small number of postoperative patients. Therefore, the presented results may not be applicable to the real screening setting. Using AI as stand-alone software may not be common. The strategies proposed for the implementation of AI systems in breast cancer screening are as a concurrent decision support and as a triage tool (6). If the aim is to improve diagnostic accuracy, using AI as support for human readers would result in better diagnostic performance than human readers alone.

The study was also limited by missing information on lesion characteristics that might affect diagnostic performance. For example, tumor diameter was not documented. Tumor size is a known prognostic factor for breast cancer and is also associated with lesion detectability. Another missing piece of information was the mammographic features of the lesions. In the categorization of breast tumors using the BI-RADS system, it is crucial to detail specific mammographic features, such as mass, calcification, and architectural distortion. These characteristics are instrumental in guiding the assessment process and may hold more value than the final BI-RADS category assigned. Notably, the performance evaluation of AI software in identifying calcification lesions, particularly in DBT, is of paramount importance. DBT presents challenges in the detection and interpretation of calcifications, especially those that are amorphous. However, it also offers the advantage of improving the detection of subtle architectural distortions by reducing the impact of overlapping breast tissue (7). Consequently, the relative presence of each mammographic feature within tumors must be considered, as it could introduce bias in the study outcomes.

According to the subanalysis based on density, Resch and Lo Gullo et al found the two AI systems performed differently. The AUCs for Transpara were 0.92 and 0.86 for dense and nondense groups, respectively. The AUCs for Profound AI were 0.92 and 0.93 for dense and nondense groups, respectively (5). It is not certain why such differences occurred. Nonetheless, neither AI system surpassed the performance of human readers, and an AUC of 0.92 was not sufficiently high.

For mass lesions in extremely dense breasts without architectural distortion, DBT may be less effective due to electronic noise and low contrast in low-dose projections. Furthermore, DBT, which involves reconstructing images from multiple low-dose x-ray projections, does not match the lesion contrast quality of breast MR and US images. Therefore, mammographic density substantially influences the efficacy of breast cancer detection via mammography and DBT, with established evidence showing reduced sensitivity as breast density increases (8). The learning data of AI systems, originally comprising cases that are challenging due to dense breasts, cannot lead to the development of AI software with superior diagnostic abilities for individuals with dense breasts. The previous study findings underscore a critical fact: Mammography is a modality best used for individuals with nondense breasts. This reaffirms the essential role of mammography in the early detection and management of breast cancer for individuals with nondense breasts, and “like parent, like child,” DBT inherits both its strengths and weaknesses.

With this in mind, the ultimate solution to improve the diagnostic performance of mammography efficiently may be identifying cases requiring special attention. For example, individuals with nondense breasts could be diagnosed using AI software, while those with dense breasts may be better suited for diagnosis by clinical radiologists assisted by AI software. Such a triaging approach could potentially be an effective method to implement AI software for mammography with DBT in clinical and screening settings. Furthermore, using AI to select patients for supplemental breast cancer screening is being investigated. Such an innovation would represent a substantial advancement in breast imaging, providing a more comprehensive tool for oncologists and radiologists in their fight against breast cancer. With AI software just emerging in the market, the accumulation and examination of real-world evidence related to mammography and DBT-based screening and diagnosis with AI support is becoming more important than ever.