A randomized controlled trial of digital breast tomosynthesis versus digital mammography in population-based screening in Bergen: interim analysis of performance indicators from the To-Be trial

Digital breast tomosynthesis (DBT) in combination with digital mammography (DM) has been claimed to be superior to DM alone in prospective studies of cancer detection in European breast cancer screening programs [1‐4]. However, recall rates have been shown to vary between studies.

Globally, a limited number of studies using DBT for screening have reported complete data on interval breast cancers [5‐7], and there is presently limited knowledge about the characteristics of the cancers detected with DBT versus DM [5, 7‐9]. Further, most studies have evaluated results of DBT in addition to DM, which substantially increases the radiation dose [10‐12]. Replacing the DM in DBT + DM with synthetic mammograms (SM), a 2D mammographic image reconstructed from the projection data obtained during the DBT acquisition, has been suggested as a solution and has recently shown promising results with respect to early performance measures in European screening programs [3, 8, 9, 13]. In addition, the sensitivity of DBT among women with dense breasts has been questioned [14‐16].

Logistical aspects including increased examination and reading times, the burden on IT systems related to storage, power and speed, and the financial costs are additional aspects that need to be explored to fully evaluate the cost-effectiveness and feasibility of using DBT + SM in organized screening programs.

To address some of the aforementioned gaps in knowledge, we conducted a randomized controlled trial (RCT) using DBT + SM versus DM only: the Tomosynthesis trial in Bergen (the To-Be trial). The To-Be RCT started in January 2016 and spanned one screening round (2 years). Our study objectives for this paper were to describe the design of this RCT and to report results of interim analyses after the first year of the trial. We compared selected secondary screening outcomes, such as examination time, time spent on screen reading and consensus, rates of cases discussed at consensus, recall rates due to abnormal mammographic findings, and mean glandular dose for DBT + SM (hereafter referred to as DBT) and DM, by mammographic density.

The To-Be trial is approved by the Regional Committees for Medical and Health Research Ethics and registered at ClinicalTrials.​gov (NCT02835625).

Among women included in the interim analyses, 1% (185/14,274) were excluded because of missing mammographic density data. Information from 14,089 women was thus included in analyses: 7037 screened with DBT and 7052 screened with DM. Women were, on average, 59 years old at screening in both groups (p = 0.469) (Table 1). The distribution of characteristics detailed in Table 1 did not differ between the two groups.

Women spent an average time (minutes:seconds) of 5:24 (median 5:13) for DBT and 4:19 (median 4:07) for DM in the screening examination room (p < 0.01) (Table 2). Average and median times spent on initial screen reading and consensus were generally higher for DBT compared to DM.

The rates of cases discussed at consensus were 6.4% for DBT and 7.4% for DM (p = 0.03) (Table 3). These rates did not differ among prevalent examinations (13.0% for both DBT and DM, p = 0.97), which was in contrast to the subsequent examinations, where the rate was 5.2% for DBT and 6.3% for DM (p < 0.01). We observed an increasing rate of cases discussed at consensus by VDG for DBT (p for trend < 0.01), but not for DM (p for trend = 0.078).

The eight radiologists’ reading volume before and during the trial period varied (Appendix, Table 5). A score of 2+, resulting in a consensus meeting, was given for an average of 4.5% of the DBT and 5.4% of the DM screen reads for each of the radiologists (Appendix, Table 6). The consensus rate decreased with 0.1% for DBT (p = 0.4) and 0.2% for DM (p = 0.05) per 1000 DM screen reads prior to start-up of the trial.

The recall rate was 3.0% for DBT and 3.6% for DM (p = 0.03) (Table 3). This rate did not differ for the two techniques among prevalently screened women (6.3% for DBT and 6.2% for DM, p = 0.95), in contrast to subsequently screened women where the rate was 2.3% for DBT and 3.1% for DM (p < 0.01). For DBT, recall rates increased from 2.2% for women with VDG 1 to 3.6% for women with VDG 4 (p for trend < 0.01). No statistically significant difference was observed for women screened with DM (p = 0.93). The number of DM screen reads before the trial period did not significantly alter the recall rates for DBT or DM (p = 0.6 for DBT and p = 0.8 for DM) (Appendix Table 5).

The cumulative reading volume of DBT during the trial showed a non-significant trend of a decreasing consensus rate (RR = 0.95, p = 0.3) (Appendix Table 6 and Fig. 4). For DM, this trend reached statistical significance (RR = 0.93, p = 0.04). For recall rates, a non-significant trend of decreasing value with cumulative reading volume during the trial period was observed both for DBT and DM (p = 0.8 for DBT and p = 0.4 for DM).

The adjusted risks of consensus and recall were lower for DBT than for DM: RR 0.71 (95% CI 0.52–0.97) for consensus and 0.58 (95% CI 0.38–0.89) for recalls (Table 4). The interaction between screening technique and mammographic density was not stastistically significant when modelling the risk of consensus. However, the risk of recall among women screened with DBT increased for VDG 3 versus VDG 1 (p = 0.033), and displayed a trend toward increased values for VDG 4 versus VDG 1 (p = 0.061), compared with DM.

MGD per examination was 2.96 mGy for DBT and 2.95 mGy for DM (p = 0.433) (Fig. 3). It did not differ with mammographic density, nor within the density groups or between screening techniques.

In the first year of this RCT using DBT and DM in population-based breast cancer screening, we found lower consensus and recall rates among women screened with DBT than with DM. Our density-stratified analyses identified that recall rates were lower for DBT only for women with non-dense breasts (VDG 1 and VDG 2). Time spent both on screen reading and consensus was longer for DBT than for DM. Average MGD did not differ between the two techniques.

The lower recall rate for DBT compared to DM found in our interim analyses supports results from other studies, although recall rates have been shown to vary [1‐4, 8, 9]. Different reading protocols and screening logistics might be some of the reasons for this variance [23‐26]. Reducing recall rates below 3% in organized screening programs seems more challenging than reducing a recall rate of 10% or higher. Regardless of screening technique, there is limited evidence on what the ideal recall rate is, according to false positive screening results, cancer detection and breast cancer mortality [27, 28].

More than 65% of the women in our study were classified as having non-dense breast (VDG 1 or VDG 2). Women with non-dense breasts had a lower recall rate when screened with DBT than when screened with DM. However, recall rates did not differ between DBT and DM for women with dense breasts (VDG 3 or VDG4). Moreover, the effect of mammographic density on the risk of recall tended to be larger for DBT than for DM, a relevant finding in a breast cancer screening program given that it applies to the larger proportion of screening attendees in our population. Given the established knowledge about the increasing risk of breast cancer with mammographic density, the increase in recall rate with density seems reasonable.

The consenus rates were also higher for women with dense rather than fatty breasts, both for DBT and DM. This is possibly related to the complex parenchyma and the need for a second opinion. The consensus meeting used in BreastScreen Norway can be considered an educational activity where “positive” cases are discussed and prior screening exams are carefully considered before a final decision about recall is made. In a broader perspective, our results, demonstrating a lower percentage of cases needing to be discussed at consensus, suggest that DBT may reduce the percentage of cases needing third arbitrating reads in other programs. As far as we know, no other studies have reported consensus rates for DBT previously. It is possible that the dense cases discussed at consensus were more obvious to recall than the fatty cases. The radiologists might thus need less time to agree about recall for the dense versus the fatty cases.

The burden of increased examination and screen reading time from DBT is a critical issue for screening programs. The increased examination time was mainly due to time spent on explaining to the women how the x-ray machine would move and to make the x-ray tube ready for exposure. This extra time is expected to be reduced or resolved in subsequent screening rounds. We demonstrated that the average reading time was 30 s longer for DBT than for DM at initial screen reading (1:11 versus 0:41, respectively). The Oslo Tomosynthesis Screening Trial (OTST) reported that an additional 41 s was needed for reading DBT compared to DM [2], while results from the STORM trial, Malmo trial and a study by Dang et al showed an increase of 44 s [1], 30 s [4] and 54 s [29], respectively. Our results therefore represent the minimum increase in time spent on initial screen reading reported in the literature to date. However, the reading time varied between radiologists. We found that some radiologists were fast readers while other used more time. We consider this variability amongst the radiologists as individual-related rather than trial-related since the findings were independent of screening technique and volume of screen reads during their career.

In our study, time spent on screen reading and consensus was lowest 5–8 months after the start of the trial. This could be because this period was during the summer months, when fewer women were screened, resulting in low power in the estimate. The low reading and consensus time could also be related to a learning effect. A workshop reviewing cancer cases dismissed by one of the two readers was performed 7–8 months after the start of the trial, as a part of the usual quality assurance in the program. This might have contributed to readers deliberating longer at screen reading and may account for the increased reading time in the third period, 8–12 months after trial commencement.

Results from other studies indicate the need for training and workshops before reading DBT in screening [30, 31]. In our study, the radiologists’ experience in DM screen reading before the trial period varied from beginners to very experienced, the latter with more than 100,000 screen reads during their career as a breast radiologist. Not all radiologists participated in screen reading DBT in the pilot, which was performed 8 weeks before the trial commenced. We identified a significant decreasing trend of consensus with reading volume during the trial for DM, but not for DBT. The volume of screen reads prior to the trial did not show any correlation with either consensus or recall rate, neither for DBT nor DM. Our study presents results only for the first year of the trial, which might be considered the learning period. Further analyses including a longer study period might shed a different light on the issue. In this trial radiologists without experience in screen reading did training on test sets, shadow reading within the trial and performed clinical mammography with DBT. In retrospect, the pilot could have been extended to 6 months to enhance reader preparation, and additional workshops could have been held to make sure all participating radiologists had read a minimum number of negative and false positive examinations, screen-detected and interval breast cancers before the trial started. Although a roster was established at the start of the trial to ensure all radiologists read equal numbers of DBT and DM cases, this plan was not strictly followed because of varying individual work speeds and an unforeseen high volume of mammography outside of the screening program. Moreover, participating radiologists were not all exposed to the same number of DBT cases. The issues encountered in the implementation of the To-Be trial represent real-world screening challenges and provide novel insights that should inform other breast screening programs when planning DBT evaluations.

We found no statistically significant difference in radiation dose per examination between DBT and DM. Gennaro et al [10] reported doses per view (CC, MLO), also calculated by Volpara, for examinations acquired using a different unit/system and found the doses to be statistically significantly higher for DBT than for DM for both views. In a per view comparison (DBT and DM exposures of the same breasts during the same compression session) they found an average increase in DBT dose compared to DM of 38% (range 0–75%). Similarly, the Oslo Tomosynthesis Screening Trial used DBT systems from the same vendor as Gennaro et al and found, on average, dose per view to be 23% higher with DBT than DM when machine-reported doses were compared [12].

Using a system from yet another manufacturer, Lång et al [4] did not report dose values; instead, the automatic exposure control was set to yield an average dose of 1.2 mGy for DM and 1.6 mGy for DBT for a standard breast model. This gives an expected per view ratio of MGD_DBT/MGD_DM of 1.33. For our system the manufacturer stated that the target MGD for DBT using automatic exposure control was equivalent to the MGD per view for DM, i.e. an expected ratio of approximately 1. The absence of a difference between MGD with DBT and DM observed in our study is therefore in line with how the system is set to operate by the manufacturer.

During the study period, routine quality assurance of the collected data and control activities were performed. We consider this to be one important strength of this study. We used an RCT design, the most reliable research design to compare screening modalities, and embedded this in a population-based screening program; these features of our trial minimize bias and increase the generalizability to other organized screening programs.

A limitation of this study is the short time spent on training and workshops in DBT for radiographers before the start of the trial, which could have influenced the results in either direction [30, 31]. Moreover, we have not presented breast cancer detection data; this decision was based on per protocol power estimation, which showed that 2 years of screening—one screening round—was needed to show a 25–30% difference in the rate of screen-detected breast cancer between DBT and DM. The moderate number of cases included in the analyses also represents a limitation in this study, particularly when stratifying into subgroups. Despite these limitations, we present our interim results to inform other population-based screening programs of selected secondary screening outcomes from an RCT of DBT and DM, in particular the estimated recall rate, screen reading time and radiation metrics, all of which matter to screening practice and research planning. To the best of our knowledge, there are no published secondary screening outcomes from other RCTs of DBT screening.

In conclusion, after the first year of running an RCT comparing DBT and DM, including about 7000 screened women in each arm, we showed a lower recall rate for women screened with DBT than DM. Our RCT sheds further light on the burdens of interpretation time and radiation dose, which are key factors in population-based screening. Time spent on screen reading and on consensus was longer for DBT than for DM. MGD measured by automated software on a GE SenoClair machine did not differ between the two techniques. Our results are somewhat different from other published studies and call for RCTs from different screening populations and with equipment from different vendors in order to gain evidence about the consequences of implementing DBT with synthesized mammograms, as a screening technique in population-based screening programs.