The frequency of missed breast cancers in women participating in a high-risk MRI screening program

Our results show that, of 131 breast cancers with negative prior MRI exams retrospectively evaluated in this study, 31% were already visible on this negative exam. In fact, 29% of the 131 lesions were rated as BI-RADS-4 or BI-RADS-5 and should thus have been recalled based on these prior exams. When including cancers presenting with minimal signs, 65% of the lesions were already recognizable on the prior MRI exam. Both from a learning perspective and in terms of liability, it is essential that these figures are available.

These results are in line with reports from Yamaguchi et al. [13], and Pages et al. [14] who, in smaller cohorts (15 and 58 patients), reported that 56 and 47% of breast cancers were already visible on prior MRI examinations and retrospectively assessed as BI-RADS-3 or higher. Both studies, however, only included screen-detected cancers on MRI. To our knowledge, this study is the first to show that also 31% of interval cancers and even 35% of incidental cancers can be identified at the last negative MRI scan.

The program sensitivity of our high-risk screening cohort was, as reported earlier, 89.7% [19], which is comparable to recent prospective studies on MRI screening in women at increased risk such as the Italian Hibcrit trial (91%) [3] and the German EVA trial (93%) [4]. Therefore, it is likely that our findings are applicable to all breast MRI screening settings. Our findings related to visibility of cancers in prior examinations are also similar to those found in mammography screening. Previous studies [7] have reported that up to 50% of the cancers detected in mammography screening with double reading were already visible at earlier screening examinations and approximately half of these were suspicious.

The clinical consequences of missing a cancer in an MRI scan obtained before prophylactic mastectomy are relatively minor. Most of these cancers are DCIS only. This is a relative frequent finding in prophylactic mastectomy specimen (approximately 5% in our institution) [20]. Since the period between the MRI scan and the subsequent prophylactic mastectomy is usually short, it is unlikely that missing these lesions will alter the patients’ prognosis. Nonetheless, the psychological impact might be substantial and should be taken into account [21].

The consequences of missing breast cancers that present subsequently as interval cancer, on the other hand, are dire. These cancers usually are invasive, poorly differentiated and fast growing [22]. Detection of these cancers at the MRI examinations could still have had a significant effect on subsequent prognosis and warrants investigation of methods to reduce false-negative reporting of MRI scans.

In the group of mammography-detected cancers, the number of visible cancers is somewhat lower than in the other groups, though this did not reach statistical significance. This result was expected as the examinations are in practice often reported simultaneously. The fact that we did find four cancers on the current MRI scan that were re-evaluated as visible in retrospect can therefore only be explained by underreporting of subtle findings on the MRI scan due to already significant findings on the mammogram that would warrant a biopsy anyway. In addition, of note is that of four invasive cancers in this mammography category only one cancer was also regarded as invisible in the consensus reading. The finding that the frequency of invisible in situ cancers is substantial (50%) underlines that, although the sensitivity of breast MRI is high, it is not 100%, as we are certain these cancers are present in the breast.

The fact that the frequency of visibility of late(r) detected cancers is virtually independent from the eventual mode of cancer detection implies that the problem should be sought in the evaluation of MRI scans itself. Radiologists in general may make two types of mistakes. The first is known as overlook error. The abnormality is simply not seen, and therefore cannot be classified correctly. The second is known as interpretation error. In this case, the lesion is seen but falsely interpreted as benign finding, and might therefore not even be reported, although the most evident cases of interpretation errors obviously mention the lesion as benign in the report [13, 14, 23, 24]. In the evaluation of screening breast MRI, likely both processes play a role. In retrospect, four cases were classified as BI-RADS-5. It is unlikely that these lesions were seen but not recalled (Fig. 2). It is far more likely that these cases were overlooked. However, the vast majority (N = 35) of visible cases were classified in the prior exam as BI-RADS-4. In these cases, interpretation errors might have prevented recall in clinical practice. This is also supported by the fact that the strongest modifier of the frequency of visible lesions is the presence of a BRCA mutation (49% of the lesions were truly invisible in BRCA mutation carriers compared to 18% in others). The much higher a priori chance of these women to develop breast cancer compared to other women at increased risk [25], along with the fast growth of cancers in these women [26], leads to lowering the threshold for recall. In other words, women with BRCA mutations are recalled for lesions that would have been ignored in women without BRCA mutations. This might also explain the slightly younger age of women with invisible lesions at the last negative MRI, as the BRCA population in general is younger than the non-BRCA population.

We also observed that IQ was associated to the visibility score, with more frequent visible findings in negative reported MRI scans of less than perfect quality. This implies that striving for excellent image quality is important. When a scan is of inferior quality, for example due to motion artifacts, rescanning the patient should be considered. We did not detect an effect of field strength on the visibility of lesions in prior negative MRI scans, and therefore underline that adequate MRI scans can be performed at both 1.5- and 3-T systems. BPE was not statistically related to the likelihood of visibility of lesions in the last negative MRI scan. Nevertheless, in women with marked BPE the frequency of visible lesions was with 14% substantially lower than in other groups. We assume that very strong BPE might in fact sometimes obscure lesions that would otherwise be detected [27], though larger numbers are required for statistical analyses.

While an obvious solution to reduce misinterpretation of eventual cancers would be to lower the threshold for recall in non-BRCA mutation carriers, this cannot be straightforwardly applied. We earlier analyzed the positive predictive value of biopsy (PPV3) in women in this cohort [19], which were 0.38 and 0.29 for BRCA1 and BRCA2 carriers, respectively. However, in women with a positive family history but no BRCA mutation PPV3 was only 0.14, which implies that lowering the threshold for recall in these women might lead to unacceptable high biopsy rates for benign lesions, and might further jeopardize the cost-effectiveness of MRI screening. The optimal balance between recall and cancer detection for women in different risk categories has yet to be determined.

We acknowledge that in the consensus read a bias was introduced by the non-blinded fashion of the evaluation. Readers were aware of the breast cancer location on the current MRI scan, mammography, or in the excised specimen. While this is common practice in auditing, it has been postulated that this approach is too harsh and leads to more lesions classified as visible than what would have been the case in a blinded setting where cases are mixed with normal scans [28]. However, in a previous study by Gubern-Merida et al. [29], a CAD system was used for the detection of breast cancer in a subset of 40 of the cancer cases that were used in the current evaluation, mixed with 120 normal scans. At four false-positives per normal scan, the sensitivity of the CAD system was 0.71 and 0.31 for visible and minimally visible lesions in the prior negative scans, respectively, whereas sensitivity was 0.82 for the respective current scans in which the cancers were actually diagnosed. This implies that indeed a substantial subset of the cancers is at least detectable in the prior negative scans and patients could therefore indeed benefit from CAD-assisted reading or a blinded second read.

This study has some other limitations. It is a single-center study based upon retrospective analysis of screening data. However, the fact that the program sensitivity is in line with published prospective trials implies that the findings are likely generalizable to similar bi-modal screening programs for women at increased risk. In addition, the prospective reads of the scans were conducted by one of eight radiologists with a strong variability (6 months to 23 years) in experience with breast MRI, which also reflects clinical practice. The numbers of missed cancers are too small to assess the effect of experience on false-negative reporting. MRI protocols changed over time. However, all examinations were performed with at the time state-of-the-art equipment and imaging quality was in adherence with international recommendations [30, 31]. Our numbers are likewise too small to detect whether MRI protocol changes might have influenced the frequency of visible findings in negative MRI scans over time.

In conclusion, our results show that almost one-third of breast cancers can be detected at the last reported negative MRI. Since this is true for both screen-detected and interval cancers, it is essential to find methods in order to reduce reading errors. As a first step, regular auditing of clinical practice seems to be indicated. In addition, structural double reading of breast MRI exams may be of value, although further research on the added cancer detection yield is required. Computer-aided detection tools for cancer detection in breast MRI might play a significant role in the prevention of reading errors in the future.