Digital breast tomosynthesis (DBT) plus synthesised two-dimensional mammography (s2D) in breast cancer screening is associated with higher cancer detection and lower recalls compared to digital mammography (DM) alone: results of a systematic review and meta-analysis

A systematic literature search was performed to identify relevant studies published on PubMed between January 1, 2010, and September 2, 2020. Search terms are available in ESM, S1. The search was limited to studies published in English language and with available abstracts. This study is reported in accordance to the Preferred Reporting Items for Systematic reviews and Meta-Analysis (PRISMA) guidelines [15, 16]. Our protocol was registered on INPLASY under registration number INPLASY202140073.

Prospective and retrospective studies with a comparative design reporting original data were evaluated eligible if, first, asymptomatic women with an average risk of breast cancer presenting for screening were considered, including women with symptoms at a regular screening. No restriction regarding age, gender, or country was set (Patients). Second, studies were included if DBT plus s2D (Index test) was compared to DM alone (Comparator). Third, at least data of CDR, recall rates, ICR, biopsy rates, or PPV 1–3 (Outcomes) had to be reported. Studies without human subjects and studies evaluating solely women with symptoms, findings suspicious of breast cancer, or screened for diagnostic work-up only were excluded.

Records identified through database search were first screened for eligibility based on information provided in title and abstract and second, using the full texts of articles. In case of any disagreement among reviewers, a third reviewer assessed the record and consensus was reached by discussion. Data of studies assessed to be eligible after full-text screening were extracted into a pretested spreadsheet by two independent reviewers. Reviewers were not blinded to the authors and institutions of studies undergoing review. The data extraction spreadsheet was designed according to the checklist of the data extraction for complex meta-analysis (DECiMAL) guide [17] (ESM, S1 for detailed information).

The RoB and applicability were evaluated by two reviewers independently using QUADAS-2 [18]. In case of any disagreement, a third reviewer was asked for assessment and consensus was reached by discussion. Studies were assessed for RoB regarding the dimensions (I) patient selection, (II) index test, (III) reference standard, and (IV) flow and timing. Applicability was evaluated by dimensions I to III.

Since all outcomes were dichotomous, RR with a 95% CI were used as summary statistics (inverse variance) to express the outcome in women screened with DBT plus s2D in relation to women screened with DM alone. p values less than 0.05 were defined as an indicator of statistical significance. Heterogeneity among the studies for each outcome was ascertained visually by forest plots and statistically using the Higgins I² for quantification inconsistency. Random effects models (REM) were used to pool an effect size when I² > 50%—indicating a moderate to high probability of heterogeneity, otherwise (I² ≤ 50%) fixed effects models (FEM) were used. Studies reporting outcomes based on the same population were included once, whereas the study with the larger sample size was included. Analysis was conducted using Microsoft Excel and Review Manager 5.4. To address heterogeneity among studies [19] and to test the robustness of results, sensitivity analyses were done using the leave-one-out approach if at least three studies were included in analysis. Each study was excluded once from meta-analysis and results were recorded to verify whether findings are depending on any study. The results are presented in summary tables in ESM, S2.

A total of 766 records were identified in PubMed. Five hundred eight records were excluded after title and abstract screening and 258 records were assessed in full text for eligibility. A total of 246 studies were excluded after full-text screening for eligibility. Twelve studies [5, 20‐30], representing the results of 10 populations, were included in at least one meta-analysis of CDR [5, 20‐28], recall rate [20, 21, 23‐27, 29], ICR [21, 30], biopsy rate [20, 23‐25, 27], PPV-1 [20, 21, 23‐27, 29], PPV-2 [20, 23], or PPV-3 [20, 23‐25, 27]. The selection process is shown in Fig. 1.

Table 1 summarises characteristics of studies comparing DBT plus s2D versus DM alone. Twelve studies [5, 20‐30] represent results of 10 unique study populations with 414,281 women. Two studies were conducted in the USA (79,209 women) [20, 23], one in Australia (10,146 women) [26], and 9 in Europe, representing results of 7 populations (324,926 women) [5, 21, 22, 24, 25, 27‐30]. Two of the European studies reported outcomes from Trento, Italy (Bernardi et al. 2020 [21] and Bernardi et al. 2016 (STORM-2) [5]). Women who were previously enrolled in STORM/STORM-2 were excluded from analysis in Bernardi et al. (2020).

Figure 2 shows the RoB and applicability assessment. All studies were evaluated having a high RoB in ‘flow of timing’, as not every woman received the same reference test after screening. Low RoB would require that all women, including women with inconspicuous findings in screening, subsequently undergo histopathological assessment for verification. Since this is ethically not acceptable, studies assessed with a high RoB in the domain ‘flow and timing’ only were assessed with an overall low RoB (ESM, S3).

Twelve studies comparing DBT plus s2D to DM alone in screening were included in our meta-analyses. We found that screening with DBT plus s2D compared to DM alone is associated with a higher CDR ([RR, 95% CI] 1.35, 1.20–1.52), decreased recalls (0.79, 0.64–0.98), and a higher cancer detection among recalls (1.69, 1.45–1.96). Cancer detection after recommended and performed biopsies was higher with DBT plus s2D compared to DM alone (PPV-2: 1.57, 1.08–2.28; PPV-3: 1.36, 1.17–1.58). We did not identify any differences in biopsy rates and ICR.

Our results regarding CDR, recall rates, and PPV-1 were in line with Alabousi et al. [13]. In comparison to Giampietro et al. [14], we found a statistically significant difference in recalls with fewer recalls for DBT plus s2D than for DM alone. However, the inclusion criteria of the latter study differ. The better results in our and Alabousi’s study may indicate lower recall rates for DBT plus s2D than for DBT plus DM. They may also reflect the learning curve from prior to more recent studies. Statistical significance was lost in our sensitivity analyses of recalls, if, e.g. the US studies Aujero et al. [20] or Freer et al. [23] were excluded. Since screening characteristics like reading procedure or screening intervals differ in the USA, these could be potential factors impacting heterogeneity. In contrast to the other studies, Houssami et al. [26] reported a statistically significantly higher risk of being recalled for one pilot screening trial when using DBT plus s2D compared to DM alone. In this population, women screened with DBT plus s2D were younger, reported symptoms more often, and participated more often in the prevalent screening round compared to women screened with DM alone [26]. Even if recall rates are contrary between studies, the number of cancers detected per 100 recalled women is consistently higher in women screened with DBT plus s2D. DBT plus s2D is associated with higher CDR and concurrently with fewer recalls. Furthermore, higher cancer detection per 100 women with recommended or performed biopsy underlines that DBT plus s2D is more precise in identifying cancers than DM alone. A 9%-point higher sensitivity (83%, 95% CI: 78–87%) for DBT plus s2D compared to DM alone (74%, 95% CI: 65–81%) was also reported by Abdullah et al. [34].

Since high CDR may be related with overdiagnoses, ICR is the more clinically relevant outcome parameter as it reflects potentially important delays in diagnosis and treatment. We identified two European studies [21, 30] reporting ICR comparing DBT plus s2D to DM alone. While Bernardi et al. [21] defined interval cancers as ‘cancers identified over two-year follow-up’ [21], Hovda et al. [30] defined interval cancers as ‘cancers diagnosed 0–24 months after negative screening findings or 6–24 months after false positive baseline screening findings’ [30]. An inconsistent trend of ICR per screening modality and small sample sizes resulted in no statistically significant difference. The same results are shown in a recently published meta-analysis by Houssami et al.[35]. They assessed ICR in women screened with DBT compared to DM. Sensitivity analyses had shown no statistically significant differences in ICR comparing DBT plus s2D to DM alone [35].

Published data on interval cancers following DBT plus s2D are limited and inconsistent. In principle, high CDR and unchanged ICR may be associated with a smaller than expected improvement of mortality reduction and with the risk of increased overdiagnosis. However, the effect of over-detection on mortality reduction and the risk of overdiagnosis can only be estimated after results of follow-up rounds, cancer stages, and biology become available from appropriately designed studies. For improved interpretation of the increased CDR, cancer biology and results for different breast densities may play a role. While Winter et al. [36] reported a lower rate of node positive interval cancers after DBT screening, Bahl et al. [37] reported comparable biology of interval cancer after DBT versus FFDM. Both of these study designs, however, cannot exclude bias. One very recent study, the only study using wide-angle DBT (without s2D), presented reduced ICR after screening with DBT compared to DM [38]. While differences might be associated with the different technologies, possible bias in the control group must also be discussed. Considering the differing results of the limited data on interval cancers, a possible correlation between the amount of recall reduction, additional detection, and effect on ICR might also be worth discussing. Also, most of the included studies were originally not designed and powered to show differences in ICR. Meta-analyses with pooled estimates based on data from underpowered studies with small sample sizes are also likely to be underpowered [39]. Finally, results concerning additional detection and effect on ICR may vary for different ranges of breast densities. To date, these data are not yet available from large studies. Given the fact that the European Commission recommends mammography screening for women aged 45–49 years old [40], DBT could be a more effective alternative, since younger women tend to have more dense breasts and the accuracy of mammography may be poorer. Furthermore, overdiagnoses may be lower in younger women with a longer remaining life-time, as small cancers have a higher risk, or longer time, respectively, for negative development.

This study has several limitations. First, search was carried out in only PubMed and studies that did not have available abstracts and English full text were excluded. Second, in 9 of the ten underlying trials women were not assigned randomly to screening modalities (concerning ~ 385,532 women among a total of 414,281). Also, 4 of the ten underlying trial study groups differ in time periods (concerning ~ 226,419 women among a total of 414,281). Unpaired and non-randomised study designs, for example in which participant characteristics (e.g. breast density, family history, or availability of screening modalities) may differ, lead to potential bias by confounding variables. Since a systematic assessment of potential confounding variables was beyond the scope of our work, a comparison of screening performance of modalities in women with dense breast tissue only seems to be useful in further subgroup analyses. Third, heterogeneity among studies was observed. We used REM and strived to interpret results only considering potential factors impacting heterogeneity. However, our study did not address other influencing or limiting factors. Valuable further data from randomised designs will become available in the near future, for example from a large RCT for which the recruitment of a prospectively acquired study population of 80,000 women [41] was recently completed.