Improving the efficacy of mammography screening: the potential and challenge of developing new computer-aided detection approaches

Abstract

In this editorial article, we share our visions of how to improve the efficacy of breast cancer screening using quantitative image feature analysis methods with the medical imaging research community. Although earlier cancer detection and treatment can help reduce the mortality rates of patients with breast cancer, the most existing long-term (or lifetime) risk models do not have clinically acceptable discriminatory power at the individual level to select women for current screening. Because of such a clinical dilemma and low efficacy of breast cancer screening, it is of great importance to establish a new and more effective personalized breast cancer screening paradigm. The new computer-aided detection (CAD) approach can be applied in two application fields: analyzing negative mammograms can provide useful information to assess cancer risk, and implementing case-based cancer risk scores can develop an adaptive CAD cueing method to yield higher sensitivity and lower false-positive detection rates.

Keywords:

• beast cancer screening
• cancer risk prediction model
• computer-aided detection for breast cancer
• early cancer detection and treatment
• risk analysis

Breast cancer is the most prevalent cancer in women worldwide, and scientific data supported that earlier cancer detection and treatment helped reduce breast cancer mortality rates in the past four decades. Currently, mammography is the only clinically accepted imaging modality for screening general population. However, accurately interpreting screening mammograms is difficult due to large variability of breast abnormalities, overlapped dense fibroglandular tissue and low cancer prevalence (<1%). As a result, the efficacy of the current uniform population-based mammography screening remains quite controversial due to its low cancer detection sensitivity and specificity [1]. For example, a recent multi-institutional prospective study reported a 53.2% sensitivity of using mammography [2]. Study also reported that during a 10-year screening period, >50% women would receive at least one false-positive recall and 7–9% have at least one false-positive biopsy [3]. The higher false-positive recall rates add anxiety with potentially long-term psychosocial consequences [4] and harms to many cancer-free women due to cumulative x-ray radiation, injection of contrast agents and unnecessary biopsies [5]. In addition, current mammography screening is also associated with a high economic burden on the healthcare system [6], which is not sustainable [7]. To solve this clinical dilemma and improve efficacy of breast cancer screening, establishing a more effective personalized breast cancer screening paradigm has been attracting wide research interest [8] in which the screening interval of the individual women should vary based on the personalized risk assessment. The goal is to increase cancer detection yield and reduce unnecessary screenings along with the higher false-positive recall rates.

However, the majority of the previous studies in an attempt to develop personalized breast cancer screening use the existing genomic biomarkers (e.g., BRCA1/2 gene mutation, which only applies to a very small portion of women) and/or epidemiology-based long-term (or lifetime) risk factors that do not have clinically acceptable discriminatory power at the individual level to select women for current screening [9]. Thus, it is important to identify and develop a new approach that can generate a clinically acceptable marker to stratify women with short-term risk of developing image-detectable breast cancer after a negative mammography screening. As a result, only a small fraction of women with high short-term risk needs to be screened annually or even every 6 months, while the majority of women with low short-term risk can be screened in the longer interval until their short-term cancer risk scores significantly increase in the future assessment. The purpose of this editorial article is to share our visions with the medical imaging research community in the effort to improve efficacy of breast cancer screening using the new quantitative image feature analysis methods.

In the past two decades, computer-aided detection (CAD) schemes of mammograms have been developed and applied in the clinical practice to assist radiologists in reading and interpreting screening mammograms and help detect more cancers. Since current CAD fails to add verified clinical value to help detection of soft tissue abnormalities, many scientists call for continuing to explore new approach to improve CAD-alone performance and optimize CAD-cueing methods [10]. We believe that although the two fundamental limitations, namely, using a single-image processing method and training datasets dominated by the ‘easier’ cases, make the current lesion cueing-based CAD have higher false-positive detection rates and correlation with radiologists’ detection results, a different case-based CAD approach can overcome such limitations and provide a new quantitative image feature analysis tool to help improve efficacy of screening mammography. Recent studies demonstrated that the new CAD approach can be applied in two application fields.

First, quantitative analysis of mammographic image feature difference and/or variation between the bilateral negative mammograms provides useful information to assess cancer risk. Studies have shown that mammographic density has the highest discriminatory power besides women’s age in the existing epidemiology-based risk models [11]. However, subjectively assessing mammographic density using a Breast Imaging-Reporting and Data System standard is often inaccurate and unreliable due to the large interobserver variability and many other factors (i.e., women’s life style/cycle and variation of imaging condition including difference in x-ray exposure and breast compression) [12,13]. To overcome these issues, CAD schemes can be applied to quantitatively analyze bilateral mammographic image density feature distribution and/or difference in both craniocaudal and mediolateral oblique views of the left and right breasts. From the computed quantitative image features, several research groups, including ours, developed and tested new risk models to predict breast cancer risk [14–16]. For example, one recent study demonstrated that applying a four-view-based CAD scheme to detect and analyze image feature difference of the negative mammograms can yield a significantly higher discriminatory power to predict the risk of women having mammography-detectable cancer in the next mammography screening (12–18 months later) [17].

Second, due to the higher false-positive detection rates, studies have shown that using CAD as ‘a second reader’ in the clinical practice reduced overall performance of radiologists in reading and interpreting screening mammograms [18]. Before a technology breakthrough happens, which enables to reduce case-based false-positive detection rates to a level that is comparable or lower than the false-positive recall rates of radiologists, alternative approaches in both CAD development and cueing methods should be investigated and tested. For example, one research group demonstrated that using an interactive CAD cueing method, in which CAD marks and their associated detection scores remain hidden unless their locations are queried by a radiologist, could help in significantly improving the radiologists’ performance compared to the use of traditional CAD cueing method [19]. However, hiding false-positive cues may still be unable to increase CAD sensitivity in detecting subtle cancers due to the limitation of CAD training datasets and have limited impact in helping radiologists significantly reduce false-positive recalls. Thus, another alternative that we recently investigated is to develop a new case-based CAD scheme that does not detect, segment and cue any individual suspicious lesions. The new CAD scheme focuses on detection and quantitative analysis of global mammographic image feature distribution and difference among all four-view images of the left and right breast. The scheme then generated a case-based likelihood score of being high risk for positive [20]. Using the case-based cancer risk scores, one can implement an adaptive CAD cueing method to increase sensitivity of cueing more subtle cancers without increase of false-positive detection rates [21]. In addition, we also hypothesized that showing the case-based high-risk scores can ‘warn’ radiologists to pay more attention to read and analyze the suspicious regions depicted on the images of these cases, and thus more sensitively detect early cancers that may be missed or overlooked without using CAD.

In summary, improving efficacy of mammography screening depend on many factors, which should include: identifying more effective short-term breast cancer risk markers to optimally determine the screening intervals for the individual women, and assisting radiologists in reading and interpreting mammograms to detect more early cancers and also reduce false-positive recalls. For these purposes, the quantitative image feature analysis of mammograms can play an important role. Recent development of radiomics [22] also demonstrated that quantitative image features computed from the radiographic images associated well with many genomic biomarkers or provided supplementary information to the genomic biomarkers and other patients’ clinical or demographic information. Hence, using quantitative image features has a number of significant advantages, including being noninvasive, cost-effective (requiring no additional tests) and highly reliable or consistent (without reporting errors from patients and inter-reader variability in reading images). Despite the disappointment of using CAD to add verified clinical value in the past [10], CAD remains a fundamental base or tool for radiomics or other quantitative image feature analysis approaches.

For the new applications, the researchers also recognized that using CAD in radiomics still face many technical challenges from tumor segmentation, feature selection/classification and fusion with other nonimage features [23]. In addition, although quantitative image feature analysis can eliminate inter-reader variability, CAD is much more sensitive to the image noise than human eyes, which could reduce reproducibility of CAD-based quantitative image feature analysis results [24]. Despite these challenges, we believe that the trend of developing radiomics and the initiative for precision medicine [25] creates many unique opportunities for the researchers in the CAD field to explore new CAD development and application approaches. In addition, integrating image feature analysis-based CAD schemes with other genomic risk factors or clinical information can help further improve performance of cancer risk assessment and efficacy of breast cancer screening in the future.

Financial & competing interests disclosure

W Qian, W Sun and B Zheng received support from the Texas University System Star Award. The authors have no other relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript apart from those disclosed.

No writing assistance was utilized in the production of this manuscript.