Mammography Topics

AI and DBT in Screening: Approaches to Improve and Accelerate Mammography Reading

High breast density is associated with an increased risk of breast cancer. However, qualitative measurements of breast density by radiologists may vary and be subjective. Automated VBDA was developed to provide objective and reproducible measurements. To explore the possibilities and clinical use of VBDA, previous studies have described the agreement between different methods of measuring volumetric density (e.g., by software such as Volpara and Quantra) and radiologists’ assessments in mammography (e.g., qualitative measurements such as BI-RADS and a visual analogue scale) with varying results. DBT is a promising technique and a potential screening modality, and the possibility to measure breast density on DBT images is important. Our group has previously compared breast density that was measured by radiologists to measurements obtained from an automated VBDA tool from Siemens using the central projection image in DBT. The results suggested that VBDA could be used in DBT in addition to mammography. Taken together, the use of a robust VBDA is important and seems possible in both mammography and DBT, enabling it to be used in individualized screening programs and in breast cancer risk scores.

Learning objectives:

1. To understand the clinical basics of volumetric breast density analysis (VBDA) based on previous studies

2. To acknowledge the difference between radiologists’ assessment of breast density and software measurements

3. To discuss VBDA’s potential use for mammography and digital breast tomosynthesis (DBT) in clinical practice

There’s a need to increase cancer detection with mammography and to increase interpretation efficiency, especially because tomosynthesis is becoming the standard of care. In this session we’ll review the impact of artificial intelligence (AI) on cancer detection and compare AI to radiologists’ interpretations of 2D and 3D mammograms. You’ll experience the benefits of the AI-based Transpara®¹ decision-support tool from ScreenPoint Medical. It’s been integrated in the advanced visualization software on MAMMOVISTA B.smart to support 2D and 3D mammography reading.

¹ Transpara® powered by FusionAI™, ScreenPoint Medical, is available in combination with MAMMOVISTA B.smart and MAMMOMAT Revelation

The use of radiomics in cancer diagnosis has seen exponential growth in interest over the past few years. Either via machine learning (ML) algorithms fed with engineered features or deep learning (DL) with convolutional neural networks, research in radiomic-based computer-aided diagnosis (CADx) is exploding. This isn’t just thanks to advances in artificial intelligence (AI); it’s also due to improvements in imaging, which is being continuously optimized to increase image quality and diagnostic performance. For X-ray breast imaging, 3D technologies like digital breast tomosynthesis (DBT), and more recently dedicated breast CT have been introduced to overcome the inherent 2D limitation of digital mammography. As a result of including the third dimension, increased insights into breast tumor characterization can be achieved. From the perspective of radiomics, these can translate into significant imaging biomarkers that may be used to develop CADx systems to help improve the diagnostic accuracy of breast cancer imaging. 

In this lecture, the potential of ML and DL radiomics for quantitative tumor malignancy prediction in breast CT images is discussed. A variety of radiomic solutions, based on both independent ML and DL algorithms and on fusion strategies that integrate the two approaches are presented and compared. Promising results were obtained, especially when incorporating ML and DL radiomics into a single AI system. This suggests that ML and DL may complement each other – at least for some clinical applications and for specific image datasets – and the most effective fusion strategies can continue to be investigated in depth in the future. 

Learning objectives: 

• Describe the potential of 3D X-ray imaging and radiomics for breast cancer diagnosis
• Understand the main differences and the potential synergy of machine learning and deep learning in breast image analysis

Abstract: Globally, over a half a million women die of breast cancer annually. In the US, the death rate has decreased by 40%, due to improved risk-based screening and targeted therapies. There is a need to increase cancer detection with mammography as well as increase interpretation efficiency, especially as tomosynthesis is becoming the standard of care. This presentation will review the impact of Artificial intelligence (AI) in cancer detection as well as compare AI to radiologist interpretation with 2D and 3D mammograms. As we become more aware and sensitive to the disparity of care among underserved populations, both in the US and around the globe, this presentation will discuss the use of AI has a strategy to mitigate the disparate breast cancer care and work toward equalizing care for all.

Prospective trials have shown the improvement in detection performance of digital breast tomosynthesis (DBT) compared to mammography at screening. However, these trials have also repeatedly shown that reading a DBT stack takes about twice as long as it does to interpret a 2D mammography image. This is one of the major roadblocks for DBT to be introduced for use in screening programs.

However, many strategies to reduce DBT reading time are being investigated. Some trials have investigated acquiring only the medio-lateral oblique (MLO) view during screening. Other trials have estimated the impact on performance of single-reading DBT cases, as opposed to current standard double reading. The capabilities of artificial intelligence (AI) to identify suspicious cases could be used to avoid the human reading of normal cases. These strategies effectively reduce the number of images to be read. Alternative methods that have been proposed to read an individual case faster include the use of thick slabs instead of thin slices to represent the entire breast, and the use of the synthetic mammogram as a guide to determining which areas of the DBT stack to review. Finally, AI -assisted navigating and reading could also have a substantial impact on reading time.

All these strategies, their technical requirements, and their potential impact on reading efficiency and detection performance will be discussed. The current evidence and availability for incorporating some of them will be presented, while the remaining advances and trials that need to be performed to prove the efficiency and safety of the others will also be reviewed.

Learning objectives

1. Describe the proposed strategies to reduce the number of cases and images that need to be interpreted.

2. Describe the methods proposed to decrease the time required to interpret a given case.

3. Discuss the remaining issues and future developments needed to implement these time-saving strategies in the clinic.

The superiority of digital breast tomosynthesis (DBT) compared to digital mammography (DM) for cancer detection in screening is undoubted, as indicated by the results from several large, prospective screening trials. One of the challenges for implementation of DBT in screening is that longer reading times are reported for DBT, up to twice as long as for DM. In high-volume screen reading, ways to improve reading times with sustained sensitivity and specificity are warranted. The image protocols vary between trials, from two-view DM+DBT, two-view synthetic DM+DBT, and one-view DBT, although with quite similar results on detection and somewhat mixed effects on recalls. This presentation will include a discussion on what is the optimal image protocol. Further, does the use of narrow- versus wide-angle DBT make a difference? Do we need double reading with DBT? Will artificial intelligence systems replace one reader? Are thicker slices, slabbing, a way forward? How much does experience add? Finally, some of the final results from the Malmö Breast Tomosynthesis Screening Trial will be presented.

Learning objectives:

1. To become familiar with the different image protocols used in prospective trials

2. To acknowledge ways to accelerate and standardize DBT screen reading

In BI-RADS 2003, the composition was based on the overall density resulting in ACR category 1 ( <25% fibroglandular tissue), category 2 ( 25–50%), category 3 (50–75%) and category 4 (>75%). In BI-RADS 2013, the use of percentages is discouraged, because in individual cases it is more important to take into account the chance that a mass can be obscured by fibroglandular tissue than the percentage of breast density as an indicator for breast cancer risk. Four groups are used: a, b, c, and d. The patterns c and d are considered “dense.” Dense breasts reduce the sensitivity of mammography up to 50%. This is the main limitation of mammography. Fortunately, tomosynthesis can significantly increase the sensitivity of mammography, especially if wide angle is used (increment of detection rate up to +43%). Tomosynthesis is able to reduce the superimposition of tissue and the anatomic noise, allowing the detection of occult lesions. However, at least a small amount of fat surrounding the lesion is needed to be detected. Breast US is widely used as an adjunct to mammography and it improves the sensitivity in dense breasts. But US is a time-consuming, operator-dependent technique that detects too many benign lesions (false-positive results). This is why US cannot be used for population-based screening. MRI is not routinely used for the evaluation of dense breasts, although it can be very useful in some particular cases (preoperative planning, high-risk patients, etc.).

Learning objectives:

1. To become familiar with the limitations of mammography in dense breasts

2. To learn the role of tomosynthesis in overcoming the limitations of mammography in dense breasts

3. To understand the role of breast US in dense breasts

High breast density is associated with an increased risk of breast cancer. However, qualitative measurements of breast density by radiologists may vary and be subjective. Automated VBDA was developed to provide objective and reproducible measurements. To explore the possibilities and clinical use of VBDA, previous studies have described the agreement between different methods of measuring volumetric density (e.g., by software such as Volpara and Quantra) and radiologists’ assessments in mammography (e.g., qualitative measurements such as BI-RADS and a visual analogue scale) with varying results. DBT is a promising technique and a potential screening modality, and the possibility to measure breast density on DBT images is important. Our group has previously compared breast density that was measured by radiologists to measurements obtained from an automated VBDA tool from Siemens using the central projection image in DBT. The results suggested that VBDA could be used in DBT in addition to mammography. Taken together, the use of a robust VBDA is important and seems possible in both mammography and DBT, enabling it to be used in individualized screening programs and in breast cancer risk scores.

Learning objectives:

1. To understand the clinical basics of volumetric breast density analysis (VBDA) based on previous studies

2. To acknowledge the difference between radiologists’ assessment of breast density and software measurements

3. To discuss VBDA’s potential use for mammography and digital breast tomosynthesis (DBT) in clinical practice