Our challenge was to understand what serous ovarian cancers look like before signs or symptoms lead to their clinical detection. PBSOs, in women with no evidence of serous cancer, provide the only immediately available window on these occult cancers. By aggregating and analyzing the results of all available prospective PBSO studies, we estimated the size and stage distribution of occult serous ovarian cancers, the duration of the “window of opportunity” for early detection, and the location of occult serous cancers (see ).
With respect to tumor stage, the results were encouraging for early detection; most occult cancers in BRCA1
carriers were still at an early stage, either CIS, stage I, or stage II. It is important to note, however, that an overwhelming majority (>90%) of these cancers were only discovered postoperatively by microscopic examination, indicating that surgical staging was incomplete; it is therefore likely that some of these cancers were understaged. Indeed, several studies have reported development of primary peritoneal serous cancers within a few years of a PBSO in which no malignancy was found, suggesting that some cancers are missed or understaged in these studies 
Our analysis of the duration of time that tumors spend at each stage while still occult was also encouraging with regard to the feasibility of early detection. By comparing the prevalence of occult tumors at each stage to the incidence of serous ovarian cancer in a matched population of women (BRCA1 carriers, similar age), we estimated that serous tumors spend approximately 4.3 y as histopathologically detectable but clinically occult early stage tumors. This 4.3-y window represents the period of time during which we could presumably save lives through early detection.
We also estimated that, on average, serous ovarian cancers have already progressed to a late stage nearly 1 y prior to their discovery. This conclusion is consistent with the results of several studies of BRCA1
carriers, which found that most serous cancers present at an advanced stage 
, and with screening studies, which also find that most of these cancers are already advanced when discovered, whether by conventional screening or by symptomatic presentation 
. Presumably, the median interval between progression to advanced stage and clinical diagnosis, and the overall duration of the occult period, would be correspondingly longer in women who are monitored less vigilantly than the high-risk women on which we based this analysis.
With respect to tumor size, our results were sobering: we estimated that occult, early-stage (CIS, stage I, or stage II) serous ovarian cancers (in BRCA1
women) have a median diameter of less than 0.3 cm and spend, on average, more than 90% of the duration of the window of opportunity for early detection at a diameter of less than 0.9 cm (Figure S9
). Indeed, many occult cancers had already progressed to stage III or IV when discovered at PBSO, and the majority of these advanced cancers were 1 cm in diameter or smaller. Remarkably, more than 90% of the occult cancers were missed both during surgery and on gross examination and discovered only upon microscopic examination (Table S1
It is also worth noting that our analysis adds to a growing body of evidence that “ovarian” cancers occur frequently in the Fallopian tubes—we found that occult serous “ovarian” cancers in BRCA1
mutation carriers are approximately twice as likely to be found in the Fallopian tubes as in the ovaries 
We further used the PBSO data, together with data on tumor sizes at clinical diagnosis, to build statistical models for the growth, progression, and diagnosis of serous ovarian cancers. The results of our growth rate estimation highlighted the explosive growth of advanced stage serous ovarian cancers. We estimated that late stage occult ovarian tumors double in volume every 2.5 mo, a conclusion that is reinforced by a previous study that found that most advanced high-grade serous tumors were undetectable by transvaginal ultrasonography several (2–12) mo prior to diagnosis, despite their relatively large size (mean 5.8 cm) at diagnosis 
. It's interesting to note that if we assume that early-stage serous ovarian cancers grow exponentially at a constant rate beginning with the first cancerous progenitor cell, and that the cancer cells are roughly 15 µm in diameter, extrapolation from this growth model suggests that on average these cancers “originate” about 8 y prior to progression to stage III, and 9 y prior to clinical diagnosis.
We estimated the performance of hypothetical screening tests as a function of their tumor-size threshold for detection and of screening frequency using our models for tumor growth, progression, and diagnosis. Our most important conclusion with respect to early detection was that an annual screening test of normal-risk women would need to detect tumors less than 1.3 cm in diameter in order to achieve 50% sensitivity, or less than 0.4 cm in diameter to reach 80% sensitivity in detecting these cancers before they progress to stage III+ ().
Combining the modeling results with stage-specific survival statistics suggests that to achieve a 50% reduction in 5-y mortality from serous ovarian cancer in a normal-risk population, an annual screen would need to detect tumors 0.5 cm in diameter or smaller (). Our estimates of potential survival benefits of early detection assumed that the stage-specific survival rates for screen-detected tumors would be the same as those for tumors diagnosed on the basis of signs and symptoms; however, it is quite possible that detecting even advanced tumors at a smaller size might yield additional survival gains. Our analysis of potential mortality reductions from screening ignored the effects of lead-time bias on stage-specific 5-y survival rates, and did not take into account the potential hazards of overdiagnosis. Had we considered either of these factors our estimates of the survival gains for a given early-detection test performance would almost certainly have been lower.
Sensitivity analyses (; Figures S7
) suggest that our major conclusions regarding the performance requirements for a successful early detection strategy are remarkably robust; even improbably large systematic biases or sampling errors in the critical data-driven parameters of the model, (size distribution of tumors at diagnosis, extent of understaging, duration of the occult period), lead to only slight changes in the results, which do not alter the important conclusions.
We based our analyses on women carrying a mutation in BRCA1
, as they comprise the largest reasonably homogenous defined subset of women who undergo PBSO. The extent to which our findings apply to other ovarian cancers is uncertain. Clearly, it would be a mistake to assume that they apply to ovarian cancers of histological types other than serous; these are distinctly different diseases, clinically and molecularly 
. Whether our findings in BRCA1
carriers can be generalized to serous ovarian cancers in general, including sporadic or non-BRCA1
familial cases, is still an important question. We were able to assess the latter directly, using data from BRCA2
and other unspecified familial cancers, and found no significant difference in our estimates of prevalence, incidence, duration, size distributions, or locations of occult cancers between these high-risk populations (Table S4
). Furthermore, on the basis of current literature, the cancers in women with BRCA1
mutations appear to be a reasonable model for sporadic serous ovarian cancers. Several studies have compared hereditary and sporadic ovarian cancers and found few significant differences in either clinical or molecular characteristics. The most consistently reported difference is in the distribution of tumor histologies; BRCA1
mutation carriers tend almost exclusively to develop serous ovarian cancer, whereas a considerable fraction of sporadic ovarian cancers are of endometrioid, mucinous, or clear-cell histology 
. A few studies have reported small differences in survival and in grade between sporadic and hereditary serous ovarian cancers 
. Unfortunately, most studies comparing sporadic to hereditary cancers failed to stratify results by histological type, making it difficult to compare serous cases in BRCA1
mutation carriers to their sporadic counterparts. Although we cannot take for granted that serous ovarian cancers in general can be represented by the natural history model described here, we are not aware of any data regarding sporadic serous cancers that is inconsistent with this model. Moreover, irrespective of their value as models for sporadic serous cancers, understanding the early natural history of serous cancers in BRCA1
women is important. These women account for up to 10% of all serous ovarian cancers and, as a readily identified high-risk population, are likely to be the first group in which any potential early detection strategy will be evaluated.
For many cancers, early detection is widely believed to be the most promising strategy to save lives 
. The parameters that we estimated here based on a review of PBSO studies, and the resulting model for the natural history of ovarian cancer, have important implications for rational design of an early detection strategy. The relatively long window of opportunity for detection prior to progression to stage III or IV suggests that an annual screen could be a viable screening strategy. Indeed, a universal annual screen capable of detecting serous cancers 0.5 cm in diameter might reduce 5-y mortality from this disease by 50% (). However, our analysis suggests that in order for an annual screen of normal-risk women to achieve a moderate 50% sensitivity in early detection of cancers that would otherwise be diagnosed at an advanced stage, the assay would need to reliably detect tumors of 1.3 cm in diameter ().
The need to detect very small tumors has important consequences for both the type of biomolecule and the type of assay technology that are likely to be effective in an early detection strategy. For a blood-based biomarker to be useful for early detection, there must be a significant difference between the maximum biomarker levels observed in the blood of almost all cancer-free women and the levels characteristically observed in women with early, occult cancers. The incremental change in biomarker levels in the blood as a consequence of a 0.5-cm tumor is likely to be extremely small, implying that the baseline levels must be consistently lower still in an overwhelming majority of ovarian cancer-free women in order for the difference to be detectable and robust. A recently published mathematical model relating secreted blood biomarker levels to tumor sizes illustrates that tumors in the millimeter diameter range can only be detected under ideal conditions of extremely high rates of biomarker secretion by tumor-associated cells and essentially zero background from healthy cells; under more realistic conditions, using known parameters of CA125 and PSA, detection limits are in the several to many centimeter diameter range 
. An addition implication of the need to detect tiny tumors is that biomarker discovery and validation will both require extremely sensitive biomarker assays. This requirement is especially challenging in the discovery phase, particularly for proteomic markers, as mass spectrometry-based proteomics approaches are only recently dipping into the nanogram/milliliter range for detection and quantification—well above the range of serum concentrations expected for proteins produced exclusively by subcentimeter-sized tumors.
Most published studies have evaluated ovarian cancer biomarkers on the basis of their ability to detect ovarian cancer in women with clinically apparent tumors 
. An important implication of our analysis is that these clinically apparent tumors are not good models for the tumors that an effective early detection test would need to detect. Indeed, our analyses show that early-stage serous ovarian cancers are rarely more than a few millimeters in diameter. In contrast, the infrequent serous cancers that present clinically while still stage I or II are typically 8 cm or more in diameter (P. Shaw and B. Rosen, unpublished data). Thus, by the time they become clinically apparent, most serous ovarian cancers are more than 200 times larger than the presymptomatic tumors a successful early detection strategy must detect. We must therefore be extremely cautious in extrapolating from the performance of serum biomarkers in detecting clinically apparent tumors to utility in early detection. Unfortunately, the magnitude of this leap is often ignored; many reports of candidate blood biomarkers for ovarian cancer assume that the detection of symptomatic stage I and II tumors will translate to an effective early detection test. Some have even used this fallacious reasoning to promote useless commercial tests 
. None of the biomarkers or biomarker panels reported to date has come close to demonstrating the performance that, according to our analysis, would be required for useful early detection. Indeed, the few large prospective trials of screening for ovarian cancer have yielded disappointing results 
It is likely that some combination of new biomarkers and new approaches will be needed to meet the challenge of early detection. Novel biomarkers that are truly unique to cancers (e.g., RNA or protein products of oncogenic gene fusions)—if they exist—are one attractive possibility. Detection of tumor biomarkers in alternative biospecimens like proximal fluids (e.g., for ovarian cancer, vaginal, or uterine lavage), rather than in blood, may be a useful strategy for boosting signal to noise by both reducing background from nonmalignant tissues and avoiding the problem of biomarker dilution inherent in blood-based assays. Positron emission tomography (PET) or ultrasound imaging of specific molecular targets is another potentially promising approach to early detection of ovarian cancers currently under investigation 
. Identification of adequately specific molecular markers and development of assays or imaging probes that provide the necessary sensitivity and signal to background ratio is still a great challenge.
Given the critical importance of the parameters that we estimated here for rational design of an early detection strategy, we believe that it would be extremely valuable to perform similar investigations into the early natural history of ovarian cancer in normal-risk women, and of other cancers. In theory, it should be possible to characterize occult “ovarian” cancers in the normal-risk population by incorporating meticulous microscopic examination of the ovaries and Fallopian tubes into the pathological examination of the resected tissues from routine hysterectomies, as over 600,000 such operations are performed annually in the United States alone (approximately half of which include salpingo-oophorectomy 
). For other cancers that develop in organs that are rarely or never removed from healthy individuals, large-scale autopsy studies including meticulous microscopic examination of target organs could potentially provide the necessary data. Such a strategy would be extremely challenging, due not only to the significant cost, but also to the relative rarity of autopsies today, the lack of detailed microscopic examination of apparently normal tissues, and the lack of a coordinated framework for systematically collecting data from autopsies. Nevertheless, we believe that understanding the early natural history of cancer is of such critical importance that the cost and effort would be justified.