Evaluating double-reading mammography for long-term surveillance in breast cancer survivors: a retrospective exploratory analysis from a single center.

Background
Breast cancer represents a significant public health concern worldwide [1]. In Italy, according to the Italian Association of Cancer Registries, approximately 55,900 new cases were diagnosed in 2023, with a five-year survival rate of 88% and an estimated 834,200 women currently living with a history of breast cancer [2].
Women with a personal history of breast cancer face a higher risk of mortality (lifetime risk of 15–20%) compared to those without a prior diagnosis, primarily due to loco-regional recurrences and contralateral breast cancer [3]. Early detection significantly improves both prognosis and quality of life: A large observational study showed a 27–47% lower mortality risk in asymptomatic patients compared to symptomatic ones [4, 5]. For this reason, imaging follow-up is crucial to detect true loco-regional recurrences or new breast cancers as early as possible, ideally before symptom onset [6].
Although follow-up guidelines are not fully uniform across institutions and countries [2, 7, 8], mammography remains the most widely recommended imaging modality for surveillance in patients who have undergone breast cancer surgery [9]. However, its sensitivity may be reduced in the postsurgical period due to post-treatment related changes that lead to interpretation challenges [9, 10]. To overcome these limitations, additional imaging methods such as breast ultrasound (US), magnetic resonance imaging (MRI), and, more recently, contrast-enhanced mammography (CEM) have been advocated and are variably employed [11, 12]. Despite their increasing use, there is still no clear evidence regarding their effectiveness and cost–benefit ratio. Moreover, the optimal surveillance strategy for women with non-metastatic breast cancer remains uncertain, as no randomized controlled trials have been conducted to compare different follow-up approaches or assess the effectiveness of tailored surveillance strategies [9, 13].
Thanks to advancements in early detection and treatment, the number of breast cancer survivors is steadily growing, a trend that is expected to continue [14]. As this population expands, there is an increasing need for sustainable and effective follow-up strategies to ensure long-term healthcare support [6, 9].
This exploratory study investigates the clinical feasibility of a long-term surveillance strategy based on double-reading mammography, structured within a population-based screening framework, for breast cancer survivors currently managed with intensive follow-up.

Material and methods

Study design and patient population
This is a single-center retrospective study, with patient consent waived due to its retrospective nature. The study was granted exemption from ethical approval by the institutional review board and was conducted in accordance with the principles of the Declaration of Helsinki. All patient-related data were handled in accordance with the standards of good scientific practice. The study design is illustrated in Fig. 1.
All consecutive women who underwent surgery for breast cancer in 2013 at IRCCS Humanitas Research Hospital were identified through electronic health records. Patients who completed a 10-year oncological follow-up at our institution and had a histologically confirmed recurrence were included in the analysis. Recurrence rates and time to relapse were calculated across histological subtypes.
In this study, the term ‘recurrence’ is used broadly to encompass both true recurrences (ipsilateral tumor events) and new tumors (either ipsilateral or contralateral). Although these represent clinically distinct entities, no differentiation was made, as the primary aim was to evaluate the diagnostic performance of mammography in detecting any form of tumor reappearance.
Available diagnostic mammograms of patients with recurrence were independently reviewed by three radiologists, R1, R2, and R3 with 1, 4, and 25 years of experience in screening mammography, respectively. All readers were blinded to each other’s assessments, as well as to clinical and prior radiological information.

Molecular subtypes classification
Invasive tumors were classified according to their molecular subtype. Hormone receptor status was assessed using immunohistochemistry (IHC), with estrogen and progesterone receptors considered positive when more than 1% of tumor cells showed immunoreactivity.
HER2 expression was by IHC using a standardized 0–3+ scoring system. Tumors with an IHC score of 0 or 1+ were considered HER2-negative. Tumors with an IHC score of 3+ or a 2+ score with confirmed gene amplification by fluorescence in situ hybridization (FISH) were classified as HER2-positive (HER2+).

Radiological review and double-reading simulation
All examinations were acquired using standard digital mammography and included two standard projections per breast (cranio-caudal and medio-lateral oblique). Each reader independently provided a binary judgment (positive or negative) on a per patient basis, in line with reading workflows adopted in population-based mammographic screening. For positive cases, readers specified the lesion type (mass, mass with calcifications, architectural distortion, or isolated calcifications) and anatomical location (breast side and quadrant).
The standard of reference was final pathology. Double reading was simulated using a worst-case scenario, in which a case was considered positive if at least one reader in each pair (R1 + R2, R1 + R3, R2 + R3) identified a suspicious finding. This approach was used to assess whether combined interpretation improved performance compared to single reading. Agreement between readers was assessed to evaluate concordance in lesion classification across independent readings.

Statistical analysis
Data were tabulated in aggregate form, and all statistical analyses were conducted using MedCalc v.23.1.3. A Chi-square test was performed to determine whether loco-regional recurrence rates varied significantly among tumor subtypes and ductal carcinoma in situ (DCIS). Pairwise comparisons between each subtype and the overall population were further analyzed using two-proportion Z-Tests to identify specific differences in recurrence rates.
A Shapiro–Wilk test was first performed to assess the normality of recurrence time distributions across histological subtypes and DCIS. If the distribution was not normal, a nonparametric Kruskal–Wallis test was used.
Sensitivity and positive predictive value (PPV) were calculated for each reader and for double reading (R1 + R2, R1 + R3, and R2 + R3) using a Chi-square test to assess any difference among multiple reading setting.
Inter-reader agreement was assessed using Cohen’s kappa coefficient for each pair of readers and Fleiss’ kappa for the overall agreement among the three. Percent agreement was also calculated to quantify the proportion of cases rated identically across pairs.
Finally, to assess whether simulated double-reading combinations offered a statistically significant improvement in sensitivity over individual readings, McNemar’s test was applied to compare each single reader against the corresponding paired combination. A p-value < 0.05 was considered statistically significant.

Results
Out of 706 patients who underwent breast cancer surgery at IRCCS Humanitas Research Hospital in 2013, 505 completed a 10-year oncological follow-up. The distribution of breast cancer subtypes in this cohort and the corresponding recurrence rates is shown in Table 1. A total of 46 women with a mean age of 68.4 years (range: 47–94; SD: 12.3) relapsed, corresponding to a recurrence rate of 9.1%: 27 (58.7%) in the ipsilateral breast or chest wall, 12 (26.1%) in the contralateral breast, and 7 (15.2%) in other anatomical areas (e.g., skin or lymph nodes). The distribution of time to recurrence, histological subtype at diagnosis, and anatomical site among the 46 loco-regional recurrences observed over the 10-year follow-up is summarized in Table 2.

Regarding the type of initial surgery, 5 out of 46 recurrences (10.9%) occurred in women who underwent mastectomy: 3 were detected in the contralateral breast and 2 in the chest wall on the operated side.

Recurrence patterns across molecular subtypes
Loco-regional recurrence rates over the 10-year follow-up period differed significantly among molecular subtypes of invasive carcinoma (p = 0.0024), with 19/271 recurrences (7.0%) in Luminal A, 6/84 (7.1%) in Luminal B, 3/28 (10.7%) in triple negative (TN), and 7/57 (12.3%) in HER2 + tumors.
A separate analysis was performed for patients initially diagnosed with DCIS. This group showed the highest overall recurrence rate, with 11 out of 65 patients (16.9%) experiencing a relapse during the 10-year follow-up period: 7 ipsilateral and 4 contralateral. Regarding surgical and adjuvant treatment, one patient underwent mastectomy and did not receive radiotherapy. The remaining 10 patients (90.9%) underwent breast-conserving surgery, and 9 of them (81.8%) received adjuvant radiotherapy as part of standard local treatment. Adjuvant endocrine therapy was administered in only one case.
Histopathological features showed comedonecrosis in 6 cases (54.5%) and micropapillary features in 3 cases (27.3%). Four recurrences (36.4%) were associated with high-grade (G3) lesions, while the remaining were intermediate-grade (G2) DCIS. No low-grade (G1) DCIS was reported. Margin involvement was absent in all DCIS, with the exception of one case for which data were unavailable. Lesion extent was reported in only one case (11 mm).
Median time to relapse was 8.2 years (IQR: 5.5–10.3) for Luminal A, 6.9 years (IQR: 4.8–9.0) for Luminal B, 4.6 years (IQR: 2.9–6.5) for TN, and 5.1 years (IQR: 3.4–7.3) for HER2 + tumors. For DCIS, the median time to recurrence was 5.2 years (IQR: 3.3–7.6) (Fig. 2).
Although the median time to recurrence varied among subtypes, these differences were not statistically significant (p = 0.190).

Radiological revision
Mammograms were available for 31/46 patients and were submitted for independent reading. Most relapses (26 cases, 83.9%) involved the breast, with 17 occurring in the ipsilateral and 9 in the contralateral breast. A total of 26 lesions were detected, varying in type and distribution across different tumor subtypes. Details are summarized in Fig. 3, and one case recognized by all the three readers is shown in Fig. 4.
Notably, five cases (16.1%) remained mammographically occult. These included three cutaneous recurrences and two nodal recurrences: one in the axilla (Fig. 5) and one in the supraclavicular region, all identified through clinical breast examination (CBE) or breast US.

Diagnostic performance of individual readers
In the single-reading setting, Reader 1 and Reader 3 detected 23 out of 31 loco-regional recurrences, while Reader 2 detected 21 recurrences. Sensitivity was 74.2% (95% CI 56.7%–86.2%) for Reader 1 and Reader 3, 67.7% (95% CI 50.1%-81.4%) for Reader 2, with no statistically significant differences in sensitivity among readers (p > 0.05).
PPVs were 95.8% for Reader 1 (95% CI 79.7%–99.2%), 95.4% for Reader 2 (95% CI 78.2%–99.1%), and 100% for Reader 3 (95% CI 85.6%–99.9%). One false positive was reported by both Readers 1 and 2; Reader 3 did not report any false positives.
Inter-reader agreement was moderate between Reader 1 and Reader 2 (κ = 0.38), and good between Reader 1 and Reader 3 (κ = 0.66), as well as between Reader 2 and Reader 3 (κ = 0.69). The corresponding agreement rates were 74.2%, 87.1%, and 87.1%, respectively. Overall agreement was moderate (Fleiss’ κ = 0.57).
Each reader missed between 8 and 10 of the 31 recurrences. Among the total false negatives, five lesions (16.1%) were located outside the mammographic field of view. The remaining false negatives were attributable to subtle findings or interpretation errors.
The single false positive recorded by both Reader 1 and Reader 2 involved the same lesion, while Reader 3 reported no false positives.

Diagnostic performance of combined readings
Simulated double-reading combinations demonstrated an overall improvement in diagnostic sensitivity compared to individual interpretations. When Reader 1 was combined with Reader 2, three additional cancer cases were identified, increasing sensitivity from 74.2% (95% CI 56.7–86.2%) to 83.9% (95% CI 67.4–92.9%). A similar gain was observed when combining Reader 1 with Reader 3, leading to the detection of two additional cases, with sensitivity rising to 77.4% (95% CI 60.2–88.6%).
The combination of Reader 2 with the others yielded even more notable gains. In particular, the pairing of Reader 2 with Reader 1 resulted in five additional cases detected compared to Reader 2 alone, increasing sensitivity from 67.7% (95% CI 50.1–81.4%) to 83.9% (95% CI 67.4–92.9%). The combination of Reader 2 with Reader 3 led to the detection of three additional cases, raising sensitivity to 80.6% (CI 95% 63.7–90.8%).
Despite the improved sensitivity, the PPV remained consistently high across all combinations. Each double-reading setup resulted in only one false positive, with PPVs of 96.2% (CI 95% 81.7–99.3%) for Reader 1 + Reader 2, 96.1% (CI 95% 81.7–99.3%) for Reader 1 + Reader 3, and 96.0% (95% CI 80.4–99.2%) for Reader 2 + Reader 3.
Differences in sensitivity and PPV between single and double reading were not statistically significant (p = 1.0).
Sensitivity and PPV for each individual reader and for the three double-reading combinations are reported in Table 3.

Discussion
In this exploratory analysis, double-reading mammography showed a promising ability to detect loco-regional recurrences, achieving a sensitivity of up to 83.9% and a PPV of 96.2%, independent of the reader’s level of experience. The obtained performance is comparable to that historically reported in population-based screening programs for asymptomatic women without a history of breast cancer, where double reading achieves a sensitivity of 76.5% [10].
When considering women with a personal history of breast cancer, the available evidence highlights a lower performance of mammography. According to Camps-Herrero et al. [6], the overall sensitivity of mammography in this population is approximately 65.4%, and as low as 60.2% within the first five years after treatment.
In contrast, even a single reading in our cohort yielded higher sensitivities (67.7–74.2%) with a PPV ranging from 95.4 to 100%, reflecting the very low number of false positives. These findings suggest that mammography may retain a higher diagnostic value in selected post-treatment populations when performed within a structured, population-based double-reading setting.
The overall inter-reader agreement in our study was moderate, reflecting the interpretative complexity of postsurgical mammograms due to anatomical changes and ambiguous lesion appearance. This level of agreement highlights that diagnostic variability among readers may occur and supports the clinical rationale for structured double reading. In this context, double reading may act as a compensatory strategy, improving detection through complementary interpretation and reducing the risk of missed recurrences.
Currently, postsurgical follow-up is typically managed within clinical settings using an ‘intensive’ model, where mammography is routinely complemented by CBE and US, regardless of individual risk or imaging findings [6].
Our findings are consistent with the long-recognized influence of tumor biology on recurrence patterns, as reported in prior literature [15]. In our cohort, TN and HER2+ tumors were associated with higher recurrence rates and earlier relapses, with a median time of approximately 5 years post-surgery. Conversely, luminal tumors showed lower recurrence rates and a longer median time to relapse, around 8 years post-treatment.
These findings support a risk-adapted follow-up strategy, in line with recent literature advocating for a personalized surveillance model based on individual risk of recurrence, biological subtype, and time to relapse [16]. In the early post-treatment years, particularly in patients with aggressive subtypes such as TN and HER2+, a more intensive imaging approach may be warranted. Conversely, patients with luminal tumors, given the delayed recurrence profile, could benefit from a longer duration follow-up, such as continuing annual mammography beyond the standard five to ten years [7, 17]. Embedding this extended surveillance within an organized screening framework may help to ensure that these women remain engaged in follow-up, reducing the risk of non-adherence once they exit specialist care.
Beyond the role of tumor subtype, our results also highlight imaging modality-related differences in recurrence detection. As the individual risk window narrows over time, follow-up may progressively rely on mammography alone. This transition to mammography-only surveillance, ideally with independent double reading, aligns with current recommendations from major breast imaging societies [2, 7, 17, 18], which advocate for more intensive monitoring in the early years post-treatment, especially for high-risk subtypes. In Italy, recently published Lombardy region guidelines (DRG XII-3458) [19] similarly recommend that follow-up be organized within breast units during this critical period or until the end of adjuvant therapy.
A particularly unexpected finding was the high recurrence rate of DCIS observed in our cohort (16.9%), which substantially exceeds the 10-year ipsilateral breast tumor relapse rate of approximately 4.5% reported in large observational studies and clinical trials [20, 21]. Although the majority of DCIS patients in our cohort received adjuvant radiotherapy, the recurrence rate remained higher than reported in the literature, especially in trials with strict inclusion criteria or more homogeneous populations [22, 23]. This discrepancy may reflect the presence of aggressive histopathological features, such as comedonecrosis and high-grade lesions, underscoring the biological heterogeneity of DCIS. In addition, 10 out of 11 patients did not receive adjuvant endocrine therapy, which has been associated with a reduced risk of recurrence [21, 23]. These factors may have contributed to the elevated recurrence rate observed in our cohort.
Despite overall good diagnostic performance, mammography alone did not detect all recurrences in our cohort. Five cases (16.1%) were detected through CBE (three cutaneous, one axillary, and one supraclavicular), all occurring in women previously treated with breast-conserving surgery.
While our findings support the value of CBE in detecting recurrences not visible on imaging, it is important to acknowledge the ongoing debate regarding its role in breast cancer surveillance. Guidelines from the US Preventive Services Task Force [24] have discouraged routine CBE in average-risk women, citing the limited benefit on mortality and concerns about false positive findings and overdiagnosis. However, these recommendations are primarily based on screening in asymptomatic, average-risk populations. Breast cancer survivors are characterized by altered breast anatomy and reduced imaging sensitivity, particularly in the axillary region, which is often inadequately evaluated by mammography alone. This is consistent with findings from Godding et al. [25] and the systematic review by Montgomery et al. [26], which reported that CBE may detect up to one-third of loco-regional relapses following breast surgery.
This is in line with the recommendations of Bucchi et al. [17], who advocate for follow-up protocols that include not only annual mammography but also regular CBE. This approach acknowledges the persistent risk of recurrence, estimated at 1.0–1.5% per year over 15–20 years, placing these women in an intermediate-risk category, between the general population (approximately 0.1–0.2% annual risk) and high-risk genetic mutation carriers, such as BRCA1/2, who face an annual risk of 1.5–2%. For this reason, reconsidering the role of CBE in long-term follow-up may be warranted.
The optimal interval for mammographic surveillance in breast cancer survivors remains under discussion. Current evidence supports extending annual imaging beyond the five-year post-treatment period, particularly for subtypes prone to late recurrences. The transition from hospital-based follow-up to population-based organized screening may improve care continuity and optimize resource use. However, this shift must also account for the persistent risk in this population. Breast cancer survivors differ from average-risk women, presenting both a higher likelihood of recurrence and lower mammographic sensitivity due to postsurgical changes [9].
For these reasons, the standard biennial schedule adopted in average-risk screening appears inadequate. Instead, maintaining an annual interval ensures timely detection of both mammographic and clinically evident recurrences, even in a mammography-only setting. This approach is consistent with current international guidelines, which recommend annual mammography as the sole imaging modality for surveillance in breast cancer survivors (ASCO, NCCN, ESMO) [8, 27, 28].
In Italy, the integration of breast cancer survivors into organized screening programs is still marginal. A 2022 national survey conducted among Senonetwork-affiliated breast centers revealed that follow-up is predominantly managed within breast units during the first 10 years after diagnosis, and only a minority of centers report active involvement of screening programs in this phase [29]. While preliminary, our results partly support the potential feasibility of transitioning selected groups of breast cancer survivors to organized screening-based follow-up. Structured systems that combine hospital-based monitoring during the high-risk phase with a gradual shift to organized screening programs may be effective.
This study has several important limitations that must be acknowledged. First, the single-center, retrospective design may restrict the generalizability of our findings across diverse healthcare settings. In addition, the small sample size, and particularly the limited number of mammograms available for radiological review, may have reduced the statistical power of our analyses and limited the robustness of the diagnostic performance metrics for both single and double reading.
Moreover, the use of a broad definition of ‘recurrence,’ including both true local recurrences and new contralateral tumors, while appropriate from a diagnostic performance perspective, may have obscured differences in biological ehavior and survival implications. Future research could benefit from analyzing them separately to refine risk stratification and surveillance strategies. Lastly, our study population included only women with diagnosed recurrence, and readers’ interpretations could have been influenced by the awareness that the reviewed cases were from a recurrence cohort, even though they were blinded to clinical information and outcomes. However, the 16.1% false negative rate reduces the extent to which the informed reading may have impacted the readers’ assessments.
In conclusion, despite the promising diagnostic performance observed in our cohort with independent double reading, several system-level challenges, such as increased radiologist workload, limited resource allocation, and the lack of structured integration between hospital-based follow-up and population-based screening programs, may hinder its widespread implementation in clinical practice. Moreover, the heterogeneity of surveillance models across institutions remains a barrier to protocol standardization. Therefore, while our findings suggest that double-reading mammography could support a more structured and resource-conscious surveillance strategy for breast cancer survivors, they should be interpreted as exploratory and hypothesis-generating. Further validation through larger, prospective, multicenter studies conducted within real-world settings is essential to confirm the potential of this approach and assess its feasibility in routine practice.