Background
The aim of breast cancer screening is to detect cancers at an early treatable state to reduce mortality. In Norway the introduction of organized biennial mammography screening in the age group 50–69 years was associated with a reduced breast cancer mortality [
1], but also a marked increase in the incidence of invasive breast cancer in women invited to participate in the screening program [
2,
3]. The incidence increased from 175 per 100,000 in the five-year period prior to screening to 350 per 100,000 during the prevalence screening in 1996–1997, leveling out at 300 per 100,000 in the succeeding period. Similar high levels during repeated screenings have occurred during introduction of mammography screening in other countries [
4,
5].
The high detection rates observed during screenings may be due to several combined reasons. At the prevalence screening, many slow-growing tumors are detected due to the long time period from screening detectable size to clinically detectable size. In addition, there will be an age-shift in the detection affecting all rounds of screening. Accumulation of experience and improvements in the screening methodology may further enhance the detection rates somewhat over time. The breast cancer incidence may also change due to causes not related to screening, such as exposure to postmenopausal hormone replacement therapy (HRT). Finally, high detection rates over time may reflect that mammography is associated with overdiagnosis, i.e. the diagnosis of tumors that will never cause clinical disease in the life time of the women.
Breast cancer has traditionally been seen as an exclusively progressive disease [
6], but recent observations of high incidence at repeated screenings have questioned these findings [
7]. Hence, we aim to test whether the observed pattern of incidence changes observed in Norway can be explained using current progressive growth models, adjusting for the changes in HRT use at population level. To address this issue, we have developed a simple simulation model. Data from the pre-screening period and the prevalence screening are used to estimate growth rates and screening sensitivity. The model then predicts incidences in the subsequent period that can be compared to the observed levels.
Discussion
We aimed at reproducing the observed breast cancer incidence trend in the AORH-counties in Norway after the introduction of organized mammography screening based on the classic progressive tumor growth assumption. The simulated incidence levels were markedly below the observed levels for the subsequent rounds of screening. The results were robust to changes in most parameters such as screening attendance, demographics and size distributions of clinical detected cancers.
The effects of mammography screening on cancer incidence have been studied since the 1970s, with diverging results [
1,
20‐
27]. One key factor for the effect of mammography screening on cancer incidence is breast cancer tumor growth and development. The growth and development of breast cancer has been studied using both analytic [
16,
17] and simulation approaches [
28,
29]. In the CISNET collaboration [
29], collaborative modelling have been used to investigated many aspects of mammography screening including mortality reduction, treatment effects and overdiagnosis. All breast cancer simulation models are, however, based on some set of assumptions that will influence the results of the model. In the CISNET collaboration, several different models have been applied to show the effect of different model assumptions. We developed a model partly inspired by the CISNET simulation models [
29‐
31], and combined it with estimates of tumor growth and screening sensitivity based on pre-screening and prevalence-screening data. As the results show, we are unable to reproduce the observed data without introducing so-called LMP-tumors, which were also used by the Wisconsin CISNET group [
19] to explain the trends in their US data. On the other side, several CISNET models have shown decent fit using US data simulating only progressive tumors [
32,
33]. Our model is, however, based on much more detailed screening data than US CISNET breast cancer modelling.
For an estimate including use of both progesterone-estrogens combinations and pure estrogen, our 2.2 relative breast cancer risk is relatively high [
12,
13,
34] compared to earlier studies, as pure estrogens use probably have less effect on breast cancer risk than progesterone-estrogens use. Hence, we could be fairly certain that at least the majority of the potential increased risk due to hormone treatment use is accounted for in our hormone treatment adjusted rate.
Following the Wisconsin-model [
19] we introduced so called LMP-tumors that typically stop growing at a sub-clinical size and later become undetectable. The LMPs are modeled in a simple way and without confirmed biological parallels. However, the simulations indicate that the existence of tumors that reach a stage where they are detectable on a mammogram, but subsequently undergo changes that make them markedly less detectable, can help explain the observed incidence pattern.
The datasets on tumor sizes are essential to the modeling. However, some tumors in the datasets lack a size measurement; 8% of the screen-detected tumors and 3% of the clinically detected tumors. When looking at the information concerning screen-detected tumors lacking size, most have a classification code stating the tumors have grown into skin and/or chest wall, hence measuring a diameter is impossible. Moreover, although we have tried to remove cases detected at non-diagnostic mammography from the clinical dataset, some such cases may remain. Thus the sizes of both clinical and screen-detected tumors may be underestimated. On the other hand, the average size reported from four randomized trials is about as the same as for our data (mean 21 mm in the non-mammography arm, 16 mm among screened tumors [
21]). Furthermore, although there may be some systematic errors in the size measurements, the sensitivity analyses using the alternative clinical dataset with larger measured tumors resulted in similar qualitative results (Additional file
1: Figure S2).
A strength in our model is the high-quality data from the Norwegian Cancer Registry [
35]. The reporting of cancer to the Norwegian cancer registry is mandatory, and diagnostic information is obtained separately from clinicians, pathologists, and death certificates, with only 0.2% of all cancers ascertained only from death certificates. The Registry is also responsible for NBCSP, and the introduction of screening in Norway occurred in a systematic fashion with county wise introduction of mammography and systematic collection of data, including data on tumor size and questionnaire data on earlier mammography.
Our model aims at simulating the female population in the AORH counties. Some of the data are, however, based on statistics from all of Norway. Comparisons of results from mimicking the real population and using uniform numbers born each year strongly indicate that small or moderate changes in demographics are of limited importance. The results are also robust to moderate changes in underlying incidence trends.
Opportunistic mammography poses specific challenges when modeling the incidence trends. Estimates based on information obtained in the NBCSP program as well as in The Norwegian Women and Cancer Study indicate fairly high levels of mammography prior to the organized screening [
36]. Given these estimates, the empirically observed increase in incidence seen prior to 1996 is surprisingly low, possibly indicating low sensitivity of the unorganized screening. If one believes there is a large effect of opportunistic mammography this would mainly affect the prevalence peak, making it even harder to reproduce the observed data, see Additional file
1: Figure S13.
The available estimates on tumor growth are not vastly different despite being based on entirely different types of data [
16,
17]. Still, robustness tests of the model are important. Lower growth rates would in the presence of strictly progressive tumors not only increase the observed incident during subsequent screening, but also the level on initial screening (Fig.
4). Generally, our tests of various growth rates indicate that variations in such rates cannot fully explain our missing fit when assuming only progressive tumors.
The use of HRT increased dramatically in Norway from the mid-90ies, about simultaneously as the introduction of organized screening, and remained high until a sharp decrease from around 2002 [
37]. The coincidence of the increase in the use of hormones and the introduction of screening poses a challenge when interpreting the data. We did a separate set of simulations based on incidence rates adjusted for the effect of HRT. Our main conclusions remained unchanged, we are able to reproduce the background incidence and the prevalence peak, but unable to reproduce the subsequent level. The gap is smaller, indicating that HRT is a likely contributor to the higher observed level in subsequent screening rounds, but the gap is still substantial especially from around 2005.
Another factor possibly contributing to increasing incidence rates is the introduction of improved screening techniques, including the introduction of digital mammography. However, the fact that the size of the detected tumors have not decreased significantly over time [
38] indicate a moderate effect since a marked effect is necessary to maintain a high incidence level over a long time period.
The introduction of mammography has led to a substantial increase in the detection and surgical removal of ductal carcinoma in situ (DCIS) [
39]. The justification for treating DCIS is that some cases will develop into invasive cancer if left untreated [
40]. Early treatment of DCIS should therefore reduce the future incidence of invasive cancers. Consequently, it is even harder to explain the incidence during subsequent screening if a substantial number of screening detected DCIS cases would have progressed to invasive cancers if left untreated [
28].
Introducing tumors that regress, or enter an undetectable state, enable us to approximate observed incidence rates. Regressing tumors are known for certain cancer types, including neuroblastoma [
41], but is regarded as uncommon for most cancer types. There are some case reports of likely clinical spontaneous regression of breast tumors [
42,
43], but since cases are confirmed through surgery or biopsy, interfering with the biology, there are few good possibilities to firmly observed spontaneous regression clinically.
One possible explanation for regressive changes being more common in breast cancer might be the dependence on growth factors, particularly estrogens. About 70% of breast tumors are reported to carry estrogen receptors [
44]. The hormone levels drop dramatically around menopause, followed by a slow further decline [
45]. The age group invited to screening overlaps the population undergoing menopause. Thus, it seems likely that many tumors will encounter a decreasing growth stimulus, potentially resulting in slower growth rates, and maybe opening for the possibility of regressive changes. The effect of hormones on breast tumor growth is also supported by the shape of the age-incidence curve [
46], the effect of anti-estrogen drugs on relapse of breast cancers [
47] and the effect of HRT [
48]. It is also conceivable that the occurrence of regressive changes is increased in periods with frequent use of external hormones, due to hormone levels being increased for a period of use with a marked drop in hormone levels at HRT termination. There are, however, no direct biological observations of tumor regression as a consequence of natural changes in hormone levels. One may also argue that the changes in hormone levels after 60 years of age are too moderate to be a likely cause of regression. Existence of such tumor changes would, however, give a possible explanation of the high incidence during repeated screenings and should be studied further.
Other model-based studies have also pointed at regressive tumor changes as possible explanations for trends seen in the considered data. As earlier stated, Fryback et al. [
19], in their Wisconsin CISNET model of incidence trends, concluding that high numbers of LMP-tumors were necessary to obtain satisfactory fits to their data. Moreover, Zahl et al. reached similar conclusions in two studies comparing cohorts with repeated screenings over a 6 years period to cohorts with only one screening at the end of the 6 years period [
2,
7].
Breast cancer risk is known to decline when women leave mammography screening programs [
14], but the key question regarding mammography screening overdiagnosis is to what degree this decline eradicates the earlier increase in breast cancer incidence during mammography screening. Working with strictly progressive natural history models, breast cancer overdiagnosis will always be low for any mammography screening with substantial life expectancy beyond the last screening exam [
49]. On the other hand, there are many epidemiological studies indicating substantial levels of overdiagnosis [
50]. If some tumors are not strictly progressive, as indicated in this study, future overdiagnosis estimates should be made with great care not fixing the study to assumptions of only progressive breast cancer tumors.
To summarize, the observed trends in Norwegian breast cancer incidence around screening introduction is not possible to reproduce by simulating data with only strictly progressive breast cancer tumors. There is hence indications that some screening detected tumors might had regressed to a non-screening detectable phase in the absence of screening detection. Epidemiological data have, however, many potential pitfalls and there is great need of more studies looking into breast cancer incidence around screening introduction, based on data from other countries with good cancer registries.