Comprehensive evaluation of AI in a large regional mammography screening program for breast cancer: detection, interval cancers, and radiologist workload

Comprehensive evaluation of AI in a large regional mammography screening program for breast cancer: detection, interval cancers, and radiologist workload | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Comprehensive evaluation of AI in a large regional mammography screening program for breast cancer: detection, interval cancers, and radiologist workload Andreas Lauritzen, Martin Lillholm, Mads Nielsen, Nico Karssemeijer, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7487727/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Artificial intelligence (AI) has been proposed to enhance mammography screening. Published literature suggested that AI increases detection and reduces workload. No published study has investigated the prospective impact of AI on sensitivity, specificity, and interval cancers. This was a population-based study on screening quality before and after implementation of AI in screening for triage and decision support. Among 270,974 screened women (156,151 with AI-support) AI-assistance improved sensitivity (73.3% with AI vs. 69.7% before AI, P=0.049), specificity (98.3% vs. 97.9%, P<0.001), and detection rates (7.8 vs. 6.5 per 1,000, P<0.001), while reducing recall and false-positive rates. Importantly, the 2-year interval cancer ratio decreased (26.7% vs. 30.3%, P=0.049). The reading workload was reduced by 36%. Cancers detected with AI-support presented with more favorable subtypes, including higher rates of ER-positive and luminal A tumors. These results provide evidence that AI can safely and effectively improve population-based screening by earlier detection of breast cancer. Health sciences/Diseases/Cancer/Breast cancer Health sciences/Health care/Medical imaging Health sciences/Medical research/Outcomes research Figures Figure 1 Figure 2 Introduction Population-based mammography screening for breast cancer using artificial intelligence (AI) has been shown to substantially reduce radiologists' reading workload by triaging screenings for single reading instead of double reading 1–3 . More recently, published studies have suggested that using AI as decision support, in addition to triaging, increases cancer detection without increasing false-positive rates 1,4 . While implementing AI into mammography screening has been effective in reducing workload and improving detection there is still a need to comprehensively evaluate the overall screening quality when using AI. Interval cancers are breast cancers diagnosed outside screening regime between screening rounds either due to being missed or deemed benign at screening, an aggressive/fast-growing subtype, or masked by dense breast tissue 5 . Therefore, the interval cancer ratio is an essential parameter to assess the true screening quality of a screening program 5,6 . This parameter has not been evaluated yet in published literature on mammography screening with AI in prospective settings. However, retrospective studies have suggested AI could be safely used maintaining or even decreasing interval cancer rates 7,8 . Additionally, breast cancer prognosis and mortality are strongly influenced by the cancer subtype. Breast cancer, being a heterogeneous disease, is characterized and treated based on multiple features such as receptor status, molecular subtypes, grading, lesion size, metastases, and other pathological and imaging properties 9 . Understanding whether the use of AI alters the characteristics of breast cancers at the time of diagnosis is essential. The optimal effect of AI-supported screening would be detection of breast cancers with features indicative of earlier-stage disease, and a shift towards fewer and less aggressive characteristics among interval cancers, indicating improved prognosis. In November 2021, AI was fully implemented into the largest regional mammography screening program in Denmark based on evidence from a prior simulation study 8 . All screenings have since been triaged using AI for either AI-assisted single reading or double reading with AI decision support. Early indicators of screening quality following implementation showed a significant increase in cancer detection rate with AI 1 . Now, with more than three years of AI-supported screening and sufficient follow-up, we evaluated whether this increase in detection was accompanied by a reduction in the two-year interval cancer ratio and improvements in screening sensitivity and specificity. Specifically, this is a prospective study with a comprehensive historical comparison of screening quality before and after AI implementation. Additionally, the study assessed changes in tumor characteristics for both screen-detected and interval cancers between the two periods. Results This study concerned women screened for breast cancer in the Capital Region of Denmark. The screening program aims to invite asymptomatic women every two years (±3 months), although the actual interval between invitations may vary slightly. Women are screened using full-field digital mammography (Siemens Revelation or Inspiration). Radiographers capture at least a mediolateral oblique and a craniocaudal view for each breast. Before AI implementation, every screening was read by two specialized full-time breast radiologists, at least one of them being an experienced senior breast radiologist. The AI system in this study was Transpara (v1.7.1, ScreenPoint Medical) 10–13 . The AI provides an exam score from 1 to 10 for each screening. Exam score 10 indicates a high probability of breast cancer. Additionally, the AI provides decision support where suspicious findings are highlighted to the reader. After AI implementation on November 18, 2021, all screenings are analyzed by AI and receives an exam score after successful analysis. If the exam is initially triaged as low risk of a present breast cancer by the AI, which is defined as exam score 5 or below, the screening is read by a single senior experienced breast radiologist only. From May 2022, this threshold was changed to exam score 7 or below. All screenings not triaged for single reading, are read by two radiologists, one of them senior, both with access to decision support by the AI. Further information on reading protocols 1 and the AI system 1,8,10–17 can be found in previously published literature. In case that radiologists see suspicious findings, the woman is recalled for diagnostic assessment. All breast cancers in this study were diagnosed based on clinical examination, supplemental mammographic imaging, ultrasound examination and needle biopsy, and include both invasive and in situ carcinomas all verified by histology. Study populations and characteristics This study comprised three cohorts (see Figure 1). The cohorts were women 1) screened before AI, 2) with AI with long follow-up, and 3) with AI with short follow-up. The cohorts of women screened with AI have been divided into two overlapping samples with different inclusions periods and lengths of follow-up; one for quality indicators that require two years of follow-up (e.g. interval cancers) and another for quality indicators requiring 180 days of follow-up (e.g. screen-detected cancers). This maximized the number of included women while allowing sufficient follow-up for each quality indicators. The inclusion period for the cohort before AI was November 17, 2019, to November 17, 2021, and initially included 118,631 screenings. After excluding women older than 70 years and 3 months at screening, screenings assessed by AI during the implementation period, and duplicated screenings, the sample included 114,823 women with one screening each. The inclusion period for the cohort with AI with long follow-up was November 18, 2021, to November 18, 2022, and initially included 68,951 screenings. After exclusions, the sample included 66,492 women with one screening each. This sample was used to calculate quality indicators related to sensitivity, specificity, and interval cancer ratios for which at least two years of follow-up was required. The inclusion period for the cohort with AI with short follow-up was November 18, 2021, to November 18, 2023, and initially included 161,655 screenings. After exclusions the sample included 156,151 women with one screening each. This sample was used for quality indicators related to recall and detection for which at least 180 days of follow-up was required. The AI system successfully analyzed 99.13% (154.798/156.151) of the screenings in this sample. Table 1 presents the characteristics of the screened women in the three samples. In the sample before AI the proportion of women with a prevalent screening (first screening) was lower than either of the samples with AI (16.71% vs. 19.30%, P<0.001 and 16.71% vs. 18.40%, P<0.001). In the sample before AI, the average number of days since last screening was within 822 days and therefore within the aim of 2 years ±3 months. In the samples with AI were 988 and 955 days and both were higher than before AI (P<0.001 and P<0.001). There was no difference in the number of women with a prior breast cancer surgery in the samples before AI and with AI (long-follow up) (P=0.95). There were less women with a prior breast cancer surgery in the sample with AI (short follow-up) compared to before AI (P<0.001). Table 1: Characteristics of study samples Characteristic Before AI (N=114,823) With AI, long follow-up (N=66,492) With AI, short follow-up (N=156,151) Average age 58.76 58.81 58.94 50-54 (%) 33,867 (29.49%) 19,267 (28.98%) 43,256 (27.70%) 55-59 (%) 31,038 (27.03%) 17,999 (27.07%) 43,390 (27.79%) 60-64 (%) 25,337 (22.07%) 14,494 (21.80%) 34,968 (22.39%) 65-70 (%) 24,581 (21.41%) 14,732 (22.16%) 34,537 (22.12%) Average BI-RADS density* 1.84 1.82 - 1 (%) 47,674 (41.52%) 28,000 (42.11%) - 2 (%) 41,783 (36.39%) 24,759 (37.24%) - 3 (%) 21,285 (18.54%) 11,128 (16.74%) - 4 (%) 3,951 (3.44%) 2,539 (3.82%) - Unknown (%) 130 (0.11%) 66 (0.10%) - Prevalent screenings (%) 19,183 (16.71%) 12,832 (19.30%) 28,735 (18.40%) Average number of days since last screening ‡ 821 986 953 Women with prior breast cancer surgery 5,044 (4.39%) 2,925 (4.40%) 6,048 (3.87%) * Based only on second reader BI-RADS density assignments ‡ Based only on women screened in the previous round Note, Radiologist's BI-RADS density assignments were not available for the full period of screening with AI with short follow-up. Screening quality indicators Table 2 presents key quality indicators comparing screening performances before and after AI implementation over the long follow-up period (2 years). The sensitivity (screen-detected cancers out of all cancers within two years after screening) increased from 69.97% (751/1,078) to 73.34% (553/754) with a one-sided Fisher's exact test (P=0.049). The specificity also increased from 97.86% (111,349/113,745) to 98.34% (64,648/65,738), with P<0.001. The interval cancer ratio (interval cancers diagnosed within two years after screening) decreased from 30.33% (327/1,078) to 26.66% (201/754), with P=0.049. The first year interval cancer ratio decreased from 13.28% (115/866) to 10.37% (64/617), a 22% relative decrease, however non-significant (P=0.053). Table 2: Screening quality indicators related to sensitivity, specificity, and interval cancers based on long follow-up (2 years) Indicator Before AI (N=114,823) With AI, long follow-up (N=66,492) P-value * Sensitivity (%) 69.67 (751/1,078, [66.86, 72.48]) 73.34 (553/754, [70.07, 76.61]) 0.049 Specificity (%) 97.89 (111,349/113,745, [97.81, 97.98]) 98.34 (64,648/65,738, [98.24, 98.44]) < 0.001 Interval cancer ratio (%) 30.33 (327/1,078, [27.66, 33.00]) 26.66 (201/754, [23.63, 29.69]) 0.049 First year interval cancer ratio (%) 13.28 (115/866, [11.18, 15.38]) 10.37 (64/617, [8.21, 12.54]) 0.053 Note, data in parentheses are numerator and denominator values, and the 95% confidence interval is in brackets. P-values have been calculated using the Fishers exact test. For all indicators, breast cancers include both invasive and in situ cancers. * One-sided Fishers exact test directions are predefined: Alternative hypotheses for sensitivity and specificity are that rates with AI are higher, and for interval cancer ratios the rates with AI are lower. As presented in Figure 2, when stratified by BIRADS density, the sensitivity increased across all density groups when screening with AI. The largest increase was observed for the group with density 4 where the sensitivity changed from 41.3% before AI to 57.1% with AI which corresponds to a 38.3% relative increase. However, none of the differences were significant (P≥0.12). The specificity increased for density groups 1, 2 and 3 (density 1: P<0.01, density 2: P<0.001, density 3: P<0.001), while the change in density 4 was not significant (P=0.21). The overall interval cancer ratio decreased across all density groups, corresponding to the increase in the sensitivity, but without reaching significance (P≥0.12). However, in the density 4 group, the interval cancer ratio decreased from 58.7% to 42.9% (P=0.12), which corresponds to a 37.0% decrease. First year interval cancer ratios also decreased across all density groups but none of the changes were statistically significant (P≥0.08). The largest reduction was observed in density 4 group where the first year interval cancer ratio decreased from 32.1% to 20.0% (P=0.25). The statistical tests for sensitivity and interval cancer ratios were likely underpowered to detect differences in the stratified samples. Screening sensitivity was modelled using logistic regression, with Method (with AI vs. before AI) and BIRADS density as predictors, which provides the relative benefit of AI varied with BIRADS density. A positive trend was observed, showing that the sensitivity gains with AI increased with higher breast density (OR for trend per BIRADS density step = 1.18, 95% CI 1.06, 1.31, P=0.004). Table 3 presents key quality indicators comparing screening performances before and after AI implementation over the short follow-up period (180 days). When screening with AI (short follow-up), the recall rate decreased from 2.75% (3,160/114,823) to 2.37% (3,696/156,151), with P<0.001, compared to before AI implementation. The overall cancer detection rate increased from 6.54 per 1,000 (751/114,823) to 7.80 per 1,000 (1,218/156,151), with P<0.001. For prevalent screenings, the cancer detection rate increased but non-significantly from 6.62 per 1,000 (127/19,183) to 7.52 per 1,000 (216/28,735) (P=0.25). The cancer detection rate for incident screenings increased from 6.52 per 1,000 (624/95,640) to 7.86 per 1,000 (1,002/127,416), with P<0.001. The false-positive rate decreased from 2.10% (2,409/114,823) to 1.59% (2,478/156,151), with P<0.001. The positive predictive value of recall increased from 23.77% (751/3,160) to 32.95% (1,218/3,696), with P<0.001. The consensus meeting/arbitration rate changed from 3.79% (4,351/114,823) to 3.86% (6,022/156,151) which was not significant (P=0.37). Radiologists' reading workload was reduced by 35.78% (111,752/312,302 readings saved) following AI implementation. Before the threshold was increased to 7, the reading workload reduction was 27.96% (14,616/52,282 readings saved), and 37.36% (97,136/260,020 readings saved) after the increase. Table 3: Screening quality indicators related to recall, detection, and workload change based on short follow-up (180 days) Indicator Before AI (N=114,823) With AI, short follow-up (N=156,151) P-value * Recall rate (%) 2.75 (3160/114823, [2.66, 2.85]) 2.37 (3696/156151, [2.29, 2.44]) < 0.001 Cancer detection rate (per 1000) 6.54 (751/114823, [6.09, 6.99]) 7.80 (1218/156151, [7.38, 8.22]) < 0.001 Cancer detection rate (prevalent screenings, per 1000) 6.62 (127/19183, [5.57, 7.67]) 7.52 (216/28735, [6.58, 8.45]) 0.12 Cancer detection rate (incident screenings, per 1000) 6.52 (624/95640, [6.03, 7.02]) 7.86 (1002/127416, [7.39, 8.33]) < 0.001 False-positive rate (%) 2.10 (2409/114823, [2.02, 2.18]) 1.59 (2478/156151, [1.53, 1.65]) < 0.001 Positive predictive value of recall (%) 23.77 (751/3160, [22.31, 25.22]) 32.95 (1218/3696, [31.46, 34.45]) < 0.001 Consensus meeting/arbitration rate (%) 3.79 (4351/114823, [3.68, 3.90]) 3.86 (6022/156151, [3.76, 3.95]) 0.19 Reading workload reduction (%) - 35.78 - Note, data in parentheses are numerator and denominator values, and the 95% confidence interval is in brackets. P-values have been calculated using the Fishers exact test. For all indicators, breast cancers include both invasive and in situ cancers. * One-sided Fishers exact test directions are predefined: Alternative hypotheses for cancer detection rates and positive predictive value are that rates with AI are higher, and for recall, false-positives, consensus meeting/arbitration the rates with AI are lower. Cancer characteristics Table 5 present characteristics of screen-detected cancers before and with AI (short follow-up). The rate of screen-detected in situ carcinomas increased from 16.78% to 19.13% but not significantly (P=0.19) after AI implementation. Rate of estrogen receptor (ER) positive increased and the rate of human epidermal growth factor receptor 2 (HER2) positive invasive cancers decreased (P=0.03 and P=0.04, respectively). The rate of invasive cancers characterized as luminal A (ER positive + HER2 negative) increased from 85.03% to 89.29% (P=0.02). Conversely, the rate non-luminal A invasive cancers decreased from 14.05% to 10.25% (P=0.03). Neither the rate of invasive cancers characterized as luminal B (ER positive + HER2 positive) nor the rate of double negative invasive cancers changed significantly (P=0.31 and P=0.46, respectively). There were no significant changes regarding invasive cancer malignancy grades (grade 1: P=0.25, grade 2: P=0.97, grade 3: P=0.29), rate of small invasive tumors (≤10mm, P=0.9), or lymph-node involvement for invasive cancers (P=0.82). The rate of women who received neoadjuvant chemotherapy (NACT) decreased from 7.06% to 5.91%, a 16.3% relative decrease, which was not significant (P=0.31). The rate of DCIS classified as Van Nuys prognostic index 3 increased from 47.41% before AI to 55.15% with AI, which corresponds to a 16.3% increase, but the change was not significant (P=0.19). Table 5: Characteristics of screen-detected cancers before AI vs. with AI based on short follow-up (180 days) Cancer characteristic Before AI (N=114,823) With AI (N=156,151) P-value § Number of all cancers 751 1218 - Number of invasive cancers 625 985 - Number of valid* invasive cancers 541 878 - Number of in situ cancers 126 233 - Number of valid* in situ cancers 116 194 - Invasive rate (%) 83.22 (625/751, [80.38, 86.06]) 80.87 (985/1218, [78.57, 83.17]) 0.19 In situ rate (%) 16.78 (126/751, [14.28, 19.28]) 19.13 (233/1218, [17.02, 21.24]) 0.19 ER positive rate (%) 92.05 (498/541, [89.46, 94.64]) 94.99 (834/878, [93.34, 96.64]) 0.03 ER unknown rate (%) 0.37 (2/541, [0.10, 0.64]) 0.11 (1/878, [0.02, 0.21]) 0.31 HER2 positive rate (%) 9.61 (52/541, [7.41, 11.82]) 6.61 (58/878, [5.14, 8.07]) 0.04 HER2 unknown rate (%) 0.74 (4/541, [0.29, 1.19]) 0.34 (3/878, [0.12, 0.57]) 0.30 Luminal A rate (%) 85.03 (460/541, [81.77, 88.28]) 89.29 (784/878, [87.07, 91.51]) 0.02 Non-luminal A rate (%) 14.05 (76/541, [11.37, 16.72]) 10.25 (90/878, [8.41, 12.09]) 0.03 Luminal B rate (%) 6.65 (36/541, [4.85, 8.46]) 5.35 (47/878, [4.05, 6.66]) 0.31 Double negative rate (%) 4.44 (24/541, [3.00, 5.87]) 3.64 (32/878, [2.59, 4.70]) 0.46 ER negative & HER2 positive rate (%) 2.96 (16/541, [1.83, 4.09]) 1.25 (11/878, [0.70, 1.80]) 0.02 Luminal unknown rate (%) 0.92 (5/541, [0.40, 1.45]) 0.46 (4/878, [0.18, 0.73]) 0.28 Grade 1 rate (%) 43.81 (237/541, [39.69, 47.93]) 46.92 (412/878, [43.64, 50.21]) 0.25 Grade 2 rate (%) 43.62 (236/541, [39.50, 47.74]) 43.74 (384/878, [40.49, 46.98]) 0.97 Grade 3 rate (%) 9.98 (54/541, [7.73, 12.23]) 8.31 (73/878, [6.66, 9.96]) 0.29 Grade unknown rate (%) 2.59 (14/541, [1.55, 3.63]) 1.03 (9/878, [0.54, 1.51]) 0.02 Small cancer rate (%) 43.07 (233/541, [38.96, 47.18]) 43.39 (381/878, [40.15, 46.64]) 0.90 Small cancer rate (including NACT-patients ‡ ) (%) 39.09 (233/596, [35.26, 42.93]) 40.15 (381/949, [37.07, 43.22]) 0.68 Size unknown rate (%) 0.18 (1/541, [0.03, 0.34]) 0.11 (1/878, [0.02, 0.21]) 0.73 Lymph node-negative rate (%) 77.63 (420/541, [73.93, 81.33]) 77.11 (677/878, [74.21, 80.00]) 0.82 Node unknown rate (%) 1.11 (6/541, [0.51, 1.71]) 1.03 (9/878, [0.54, 1.51]) 0.88 NACT-patient rate (%) 7.06 (53/751, [5.44, 8.68]) 5.91 (72/1218, [4.72, 7.10]) 0.31 DCIS Van Nuys Grade 1 rate (%) 12.93 (15/116, [8.00, 17.87]) 10.82 (21/194, [7.19, 14.46]) 0.58 Grade 2 rate (%) 38.79 (45/116, [30.42, 47.17]) 32.47 (63/194, [26.28, 38.67]) 0.26 Grade 3 rate (%) 47.41 (55/116, [38.56, 56.27]) 55.15 (107/194, [48.12, 62.19]) 0.19 Grade unknown rate (%) 0.86 (1/116, [0.15, 1.57]) 1.55 (3/194, [0.53, 2.57]) 0.61 Note, data in parentheses are numerator and denominator values, and the 95% confidence interval is in brackets. DCIS = ductal carcinoma in situ. NACT = Neoadjuvant chemotherapy. ER = Estrogen receptor. HER2 = Human Epidermal Growth Factor Receptor 2. * Valid cancers refer to women with diagnosed breast carcinomas that are not NACT-patients, having had surgery, and for which the surgical specimen contains residual of the cancer diagnosed at the preceding biopsy. See Supplemental Table 1 for further details. ‡ NACT-patients count as having a tumor size > 10mm. § P-values have been calculated using the chi-square test. Table 6 present characteristics of interval cancers before and with AI (long follow-up). None of the measured cancer characteristics changes was concluded to be statistically significant. However, the rate of HER2 positive invasive cancers changed from 9.52% before AI to 5.34% with AI (P=0.16) which corresponds to a 43.9% relative decrease. The rate of luminal B invasive cancer decreased from 6.67% to 3.82% (P=0.26) which is a 42.7% relative decrease. The rate of small invasive cancers (≤ 10mm) decreased from 27.14% to 21.37% (P=0.23) which was a relative decrease of 21.3% The rate of DCIS classified as Van Nuys prognostic index 3 increased from 29.41% before AI to 42.86% with AI, which corresponds to a 45.7% relative increase, but the change was not significant (P=0.44). Table 6: Characteristics of interval cancers before AI vs. with AI based on long follow-up (2 years) Cancer characteristic Before AI (N=114,823) With AI (N=66,492) P-value § Number of all cancers 327 201 - Number of invasive cancers 309 185 - Number of valid* invasive cancers 210 131 - Number of in situ cancers 18 16 - Number of valid* in situ cancers 17 14 - Invasive rate (%) 94.50 (309/327, [91.47, 97.52]) 92.04 (185/201, [87.46, 96.62]) 0.26 In situ rate (%) 5.50 (18/327, [3.51, 7.50]) 7.96 (16/201, [4.96, 10.96]) 0.26 ER positive rate (%) 89.52 (188/210, [84.65, 94.40]) 87.79 (115/131, [81.08, 94.49]) 0.62 ER unknown rate (%) 0.00 (0/210, [0.00, 0.00]) 0.00 (0/131, [0.00, 0.00]) - HER2 positive rate (%) 9.52 (20/210, [6.25, 12.80]) 5.34 (7/131, [2.61, 8.07]) 0.16 HER2 unknown rate (%) 0.00 (0/210, [0.00, 0.00]) 0.00 (0/131, [0.00, 0.00]) - Luminal A rate (%) 82.86 (174/210, [77.18, 88.53]) 83.21 (109/131, [75.88, 90.53]) 0.93 Non-luminal A rate (%) 17.14 (36/210, [12.65, 21.64]) 16.03 (21/131, [10.73, 21.33]) 0.79 Luminal B rate (%) 6.67 (14/210, [4.01, 9.32]) 3.82 (5/131, [1.64, 5.99]) 0.26 Double negative rate (%) 7.62 (16/210, [4.74, 10.49]) 10.69 (14/131, [6.47, 14.90]) 0.33 ER negative & HER2 positive rate (%) 2.86 (6/210, [1.32, 4.40]) 1.53 (2/131, [0.42, 2.63]) 0.43 Luminal unknown rate (%) 0.00 (0/210, [0.00, 0.00]) 0.76 (1/131, [0.13, 1.39]) - Grade 1 rate (%) 34.76 (73/210, [28.65, 40.88]) 35.11 (46/131, [27.47, 42.76]) 0.95 Grade 2 rate (%) 48.10 (101/210, [41.43, 54.76]) 48.09 (63/131, [39.71, 56.47]) > 0.99 Grade 3 rate (%) 16.19 (34/210, [11.82, 20.56]) 16.79 (22/131, [11.36, 22.23]) 0.88 Grade unknown rate (%) 0.95 (2/210, [0.26, 1.64]) 0.00 (0/131, [0.00, 0.00]) - Small cancer rate (%) 27.14 (57/210, [21.58, 32.71]) 21.37 (28/131, [15.22, 27.53]) 0.23 Small cancer rate (including NACT-patients ‡ ) (%) 21.19 (57/269, [16.73, 25.65]) 16.87 (28/166, [11.94, 21.80]) 0.27 Size unknown rate (%) 0.48 (1/210, [0.08, 0.87]) 0.00 (0/131, [0.00, 0.00]) - Node-negative rate (%) 63.33 (133/210, [56.63, 70.04]) 63.36 (83/131, [54.84, 71.88]) > 0.99 Node unknown rate (%) 1.43 (3/210, [0.49, 2.37]) 1.53 (2/131, [0.42, 2.63]) 0.94 NACT-patient rate (%) 15.90 (52/327, [12.34, 19.47]) 16.42 (33/201, [11.94, 20.90]) 0.88 DCIS Van Nuys Grade 1 rate (%) 29.4 (5/17, [13.28, 45.54]) 28.57 (4/14, [11.72, 45.42]) 0.96 Grade 2 rate (%) 35.29 (6/17, [17.31, 53.28]) 21.43 (3/14, [7.57, 35.29]) 0.40 Grade 3 rate (%) 29.41 (5/17, [13.28, 45.54]) 42.86 (6/14, [21.38, 64.33]) 0.44 Grade unknown rate (%) 5.88 (1/17, [1.05, 10.72]) 7.14 (1/14, [1.27, 13.01]) 0.89 Note, data in parentheses are numerator and denominator values, and the 95% confidence interval is in brackets. DCIS = ductal carcinoma in situ. NACT = Neoadjuvant chemotherapy. ER = Estrogen receptor. HER2 = Human Epidermal Growth Factor Receptor 2. * Valid cancers refer to women with diagnosed breast carcinomas that are not NACT-patients, having had surgery, and for which the surgical specimen contains residual of the cancer diagnosed at the preceding biopsy. See Supplemental Table 1 for further details. ‡ NACT-patients count as having a tumor size > 10mm. § P-values have been calculated using the chi-square test. Discussion In this study, we presented a comprehensive analysis of the impact of implementing AI in mammography screening across a large regional program with two years of follow-up. This study demonstrates that introducing AI as a tool for both triage and decision-support reduced radiologists' workload by 36% while enhancing screening performance. This is made evident by improvements in sensitivity (69.7% to 73.3%, P=0.049), cancer detection rates (6.54 to 7.80 per 1,000, P<0.001), interval cancer ratios (30.3% to 26.7%, P=0.049), recall rates (2.75% to 2.37%, P<0.001), and false-positive rates (2.10% to 1.59%, P<0.001). The results of this study largely agrees with recent large-scale trials. In the MASAI trial, AI-supported workflow decreased reading workload by 44% while increasing cancer detection by 29% (5.0 to 6.4 per 1,000 screenings) without increasing the number of false-positives or recalls 2,4 . Likewise, our study observed that cancer detection rates increased from 6.54 to 7.80 per 1,000 (P<0.001) along with lower recall and false-positive rates. In a large Swedish prospective trial (ScreenTrustCAD), showed that replacing one radiologist with AI, compared to double reading, achieved non-inferior cancer detection rates (relative ratio 1.04, 95% CI: 1.00, 1.09) 3 . These two Scandinavian studies along with a number of other international prospective and retrospective studies supports our results that AI can be used for triage and act as a substitute reader to reduce workload, and provide decision support to maintain screening quality or even improve it 18–24 . In this study, analysis of tumor subtypes in invasive screen-detected cancers showed a shift toward more favorable biological profiles where the ER-positive rate increased from 92% to 95% (P=0.03) and the rate luminal A positives increased from 85% to 89% (P=0.02). Results suggest that for interval cancers, there was a trend towards fewer HER2-positive and luminal B subtypes. These changes were not significant, so larger samples sizes are needed to assess AI’s impact on aggressive tumor profiles. Across the BIRADS density categories, interval cancer ratios were consistently equal or lower when screening with AI compared to before AI, though statistical power was limited. These trends may suggest that AI can support radiologists in detecting cancers during screening, as opposed to the cancers being diagnosed as interval cancers, especially for women with high-density breasts. This study’s strengths include its large-scale regional coverage, but most importantly, to our knowledge, the first study to present 2-year interval cancer ratios when screening with AI. Limitations include, firstly, that this is single institution (but multi-site) screening program and observations may not correspond to other screening workflows or populations. However, the results align with the recent MASAI and ScreenTrustCAD trials, and the concurrent findings of the studies enhances the evidence that AI can improve screening across different health care systems, workflows, and populations. Secondly, this is a historical comparison of two cohorts (before and with AI) and these cohort differ in terms of a longer screening interval which may increase both cancer detection and interval cancer rates as tumors have more time to develop. There are also a larger proportion of prevalent screenings in the samples with AI which could be a source of increased detection, however, the cancer detection rate also increased for incident screenings alone. Lastly, there were less women with a prior breast cancer in the sample with AI. However, as these women have a higher overall risk this should decrease cancer detection which was not observed in this study. On the contrary, the cancer detection rate increased (P < 0.001). Additionally, this study showed a slight and non-significant increase in DCIS detection (17% to 19%, P=0.19). This trend should be carefully monitored to prevent overdiagnosis. This study provides additional evidence that AI-supported mammography screening workflows can enhance detection and shift detected breast cancer subtypes towards more favorable subtypes while considerably reducing radiologist reading workload. Most importantly, we provide novel evidence that AI can further reduce the number of interval cancer which indicates earlier detection and thereby improve prognosis. Future work should focus on long-term clinical outcomes for women in AI-supported population-based breast cancer screening. Methods Study Approvals This study received approval from The Danish Patient Safety Authority and Danish Data Protection Agency, who also waived the need for informed consent from women participating in mammography screening in the Capital Region of Denmark (reference 3–3013–2118, addendum 2019/2023). Outcomes The primary outcomes of this study were the screening sensitivity, specificity, and interval cancer ratio in the samples before AI and with AI with long follow-up (two years). Note that breast cancers include both invasive and in situ carcinomas. The sensitivity is calculated as the percentage of breast cancers detected at screening out of all breast cancers diagnosed within two years from the screening date. The specificity is calculated as the percentage of true negative exams (not recalled and no cancers) out of all negative exams within the 2-year follow-up period. The interval cancer ratio is calculated as the percentage of interval cancers (cancers diagnosed outside screening regime) out of all breast cancers diagnosed within the 2-year follow-up period. The secondary outcomes are the recall rate, cancer detection rate, false-positive rate, positive predictive value of recall, consensus meeting/arbitration rate, and radiologists reading workload in the samples before AI and with AI with short follow-up (180 days). The recall rate is calculated as the percentage of recalled women out of all women in the sample. The cancer detection rate is the number of breast cancers diagnosed at screening out of all screenings in the sample per 1,000 screenings. The false-positive rate is calculated as the percentage of women who were recalled but did not receive a breast cancer diagnosis within 180 days from the screening date out of all screenings in the sample. The positive predictive value is calculated as the percentage of women who were recalled and received a breast cancer diagnosis within 180 days out of all women who were recalled. The consensus meeting/arbitration rate were calculated as the number of screenings for which a consensus meeting was held, or an additional radiologist made the final decision on recall out of all screenings in the sample. Before AI, a consensus meeting/arbitration triggers when the two readers disagree on recall. with AI a consensus meeting/arbitration triggers when either the single reader decides to recall (AI triaged the screening to single reading as likely normal), or when the two readers disagreed for screenings for which the AI selected the exam for AI-assisted double reading. The reading workload reduction is calculated as the percentage of reads that were saved compared to standard double reading (e.g. a 36% reduction means that for 50 screenings that would normally be double read, the AI triaged 36 screenings for single reading. Thus, with AI, the total number of independent reads would be 36*1+14*2=64 reads which correspond to a 36% reduction of the 100 independent reads before AI. Statistical Analysis The samples used in this study along with the inclusion periods, exclusion criteria, and follow-up times were all predefined prior to analysis. All the quality indicators were defined based on national guidelines 25 . Primary and secondary screening quality indicators were compared between study samples using the one-sided Fisher’s exact tests 26 . A previous publication on screening with AI in the Capital Region of Denmark, suggested that cancer detection increased as a results of implementing AI 1 . Based on those results, the test directions were predefined before analysis such that the alternative hypotheses for sensitivity, specificity, cancer detection rate, and positive predictive value of recall were that rates in the samples with AI were higher than the sample before AI. The alternative hypotheses for interval cancer ratios, recall rates, false-positive rates, and consensus meeting/arbitration were that rates were lower. For subgroup analyses (e.g., BI-RADS density strata), Fisher’s exact tests were used given smaller sample sizes. In the analyses of breast cancer characteristics, only the chi-squared test was used to test for any differences in values regardless of direction of change. Statistical significance was set at P<0.05 for key metrics. Analyses were performed in Python (version 3.11). References Lauritzen, A. D. et al. Early Indicators of the Impact of Using AI in Mammography Screening for Breast Cancer. Radiology 311 , (2024). Lång, K. et al. Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI): a clinical safety analysis of a randomised, controlled, non-inferiority, single-blinded, screening accuracy study. Lancet Oncol 24 , 936–944 (2023). Dembrower, K., Crippa, A., Colón, E., Eklund, M. & Strand, F. Artificial intelligence for breast cancer detection in screening mammography in Sweden: a prospective, population-based, paired-reader, non-inferiority study. Lancet Digit Health 5 , e703–e711 (2023). Hernström, V. et al. Screening performance and characteristics of breast cancer detected in the Mammography Screening with Artificial Intelligence trial (MASAI): a randomised, controlled, parallel-group, non-inferiority, single-blinded, screening accuracy study. Lancet Digit Health 7 , e175–e183 (2025). Perry, N. et al. European Guidelines for Quality Assurance in Breast Cancer Screening and Diagnosis. Fourth Edition - Summary Document . Annals of Oncology vol. 19 (2008). Radswiki, T., Campos, A. & Murphy, A. Interval breast cancer. in Radiopaedia.org (Radiopaedia.org, 2011). doi:10.53347/rID-15319. Lång, K., Hofvind, S., Rodríguez-Ruiz, A. & Andersson, I. Can artificial intelligence reduce the interval cancer rate in mammography screening? doi:10.1007/s00330-021-07686-3/Published. Lauritzen, A. D. et al. An artificial intelligence–based mammography screening protocol for breast cancer: outcome and radiologist workload. Radiology 304 , 41–49 (2022). Danish Breast Cancer Group. Systemisk Behandling Af Brystkræft (I) - Hvem Skal Anbefales Adjuverende Systemisk Behandling? https://www.dmcg.dk/siteassets/kliniske-retningslinjer---skabeloner-og-vejledninger/kliniske-retningslinjer-opdelt-pa-dmcg/dbcg/dbcg_adjuve-systemisk-bh-1_v1.3_admgodk070722.pdf (2024). Rodriguez-Ruiz, A. et al. Stand-Alone Artificial Intelligence for Breast Cancer Detection in Mammography: Comparison With 101 Radiologists. J Natl Cancer Inst 111 , 916–922 (2019). Rodríguez-Ruiz, A. et al. Detection of breast cancer with mammography: Effect of an artificial intelligence support system. Radiology 290 , 305–314 (2019). Raya-Povedano, J. L. et al. AI-based Strategies to Reduce Workload in Breast Cancer Screening with Mammography and Tomosynthesis: A Retrospective Evaluation. Radiology 300 , 57–65 (2021). Romero-Martín, S. et al. Stand-Alone Use of Artificial Intelligence for Digital Mammography and Digital Breast Tomosynthesis Screening: A Retrospective Evaluation. Radiology 302 , 535–542 (2022). Pinto, M. C. et al. Impact of artificial intelligence decision support using deep learning on breast cancer screening interpretation with single-view wide-angle digital breast tomosynthesis. Radiology 300 , 529–536 (2021). Rodríguez-Ruiz, A. et al. Detection of breast cancer with mammography: Effect of an artificial intelligence support system. Radiology 290 , 305–314 (2019). Rodriguez-Ruiz, A. et al. Stand-Alone Artificial Intelligence for Breast Cancer Detection in Mammography: Comparison With 101 Radiologists. J Natl Cancer Inst 111 , 916–922 (2019). Wanders, J. O. P. et al. The combined effect of mammographic texture and density on breast cancer risk: A cohort study. Breast Cancer Research 20 , 1–10 (2018). Eisemann, N. et al. Nationwide real-world implementation of AI for cancer detection in population-based mammography screening. Nat Med 31 , 917–924 (2025). Dembrower, K. et al. Effect of artificial intelligence-based triaging of breast cancer screening mammograms on cancer detection and radiologist workload: a retrospective simulation study. Lancet Digit Health 2 , e468–e474 (2020). Larsen, M., Aglen, C. F., Hoff, S. R., Lund-Hanssen, H. & Hofvind, S. Possible strategies for use of artificial intelligence in screen-reading of mammograms, based on retrospective data from 122,969 screening examinations. Eur Radiol (2022) doi:10.1007/s00330-022-08909-x. Ng, A. Y. et al. Prospective implementation of AI-assisted screen reading to improve early detection of breast cancer. Nat Med 29 , 3044–3049 (2023). Pacilè, S. et al. Improving Breast Cancer Detection Accuracy of Mammography with the Concurrent Use of an Artificial Intelligence Tool. Radiol Artif Intell 2 , e190208 (2020). Rodriguez-Ruiz, A. et al. Can we reduce the workload of mammographic screening by automatic identification of normal exams with artificial intelligence? A feasibility study. Eur Radiol 29 , 4825–4832 (2019). Leibig, C. et al. Combining the strengths of radiologists and AI for breast cancer screening: a retrospective analysis. Lancet Digit Health 4 , e507–e519 (2022). Vejborg, I. Mammografiscreening i Danmark - Kliniske Retningslinjer . rkkp.dk/siteassets/de-kliniske-kvalitetsdatabaser/databaser/mammografiscreening/kliniske-retningslinjer-for-mammografiscreening-i-danmark.-revideret-2.maj-2022_._..__.pdf (2022). Author, P. & Fisher, R. A. On the Interpretation of χ 2 from Contingency Tables, and the Calculation Of . Source: Journal of the Royal Statistical Society vol. 85 (1922). Additional Declarations Yes there is potential Competing Interest. Author Andreas D. Lauritzen is employed as AI Research Engineer in ScreenPoint Medical, and holds shares in a company that holds depositary receipts in ScreenPoint Medical, the provider of the AI system evaluated in this study. Author Martin Lillholm holds shares in a company that holds depositary receipts in ScreenPoint Medical. Author Mads Nielsen holds shares in a company that holds depositary receipts in ScreenPoint Medical. Author Nico Karssemeijer is employed as Chief Science Officer in ScreenPoint Medical, and is a shareholder of ScreenPoint Medical. Author Ilse Vejborg has no competing interests. Supplementary Files Lauritzenetalsupplementalmaterials280825.docx Supplemental Material Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7487727","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":509730839,"identity":"71df127d-9e9b-4316-974c-17ed11a73ac8","order_by":0,"name":"Andreas Lauritzen","email":"data:image/png;base64,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","orcid":"https://orcid.org/0000-0002-0638-0126","institution":"Gentofte Hospital, Capital Region of Denmark","correspondingAuthor":true,"prefix":"","firstName":"Andreas","middleName":"","lastName":"Lauritzen","suffix":""},{"id":509730840,"identity":"948b4b9c-e16b-48d2-8bf3-860a8e78b6c0","order_by":1,"name":"Martin Lillholm","email":"","orcid":"","institution":"Department of Computer Science, University of Copenhagen","correspondingAuthor":false,"prefix":"","firstName":"Martin","middleName":"","lastName":"Lillholm","suffix":""},{"id":509730841,"identity":"0c1fb80c-2642-4e79-a1fc-b771fcc845dc","order_by":2,"name":"Mads Nielsen","email":"","orcid":"","institution":"Department of Computer Science, University of Copenhagen","correspondingAuthor":false,"prefix":"","firstName":"Mads","middleName":"","lastName":"Nielsen","suffix":""},{"id":509730842,"identity":"68caec87-dd33-4e3f-acd2-b43f85203aa1","order_by":3,"name":"Nico Karssemeijer","email":"","orcid":"","institution":"ScreenPoint Medical","correspondingAuthor":false,"prefix":"","firstName":"Nico","middleName":"","lastName":"Karssemeijer","suffix":""},{"id":509730843,"identity":"e4a61b83-fe46-45b9-a7cb-b324053d70d5","order_by":4,"name":"Ilse Vejborg","email":"","orcid":"https://orcid.org/0000-0002-2329-203X","institution":"Gentofte Hospital, Capital Region of Denmark","correspondingAuthor":false,"prefix":"","firstName":"Ilse","middleName":"","lastName":"Vejborg","suffix":""}],"badges":[],"createdAt":"2025-08-29 10:25:41","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7487727/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7487727/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":90648644,"identity":"9a6a7cf0-6ebe-46e4-9d7b-af0c49a8d541","added_by":"auto","created_at":"2025-09-05 08:19:49","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":96538,"visible":true,"origin":"","legend":"\u003cp\u003eTimeline of inclusion periods (a) and Flow diagram of the inclusion and exclusion process for the cohort of women screened before AI (b), with AI (long follow-up) (c), with AI (short follow-up) (d)\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7487727/v1/0958c921d9e5822cd1bf8b98.png"},{"id":90648643,"identity":"652399ea-cf2d-4ce8-bf74-40809ab82243","added_by":"auto","created_at":"2025-09-05 08:19:49","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":158978,"visible":true,"origin":"","legend":"\u003cp\u003eBarplots of (a) screening sensitivity, (b) screening specificity, (c) 2-year interval cancer ratio, and (d) first year interval cancer ratio over the second reader BIRADS density assignment. The black bars are 95% confidence intervals. P-values have been calculated using a one-sided Fisher's exact test with predefined directions (alternative hypotheses: Before AI rate \u0026lt; with AI rate for sensitivity/specificity, and before AI rate \u0026gt; with AI rate for interval cancer ratios).\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7487727/v1/212c0bdd991eab9acacd679f.png"},{"id":90649703,"identity":"daa3e3e0-00d3-45bb-8f65-0b7d64fd351d","added_by":"auto","created_at":"2025-09-05 08:35:52","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1270928,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7487727/v1/20bbe620-281b-442c-923c-03d913001c16.pdf"},{"id":90648645,"identity":"d64b85bd-cd00-4a4c-bdee-e1117fb9a081","added_by":"auto","created_at":"2025-09-05 08:19:49","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":137422,"visible":true,"origin":"","legend":"Supplemental Material","description":"","filename":"Lauritzenetalsupplementalmaterials280825.docx","url":"https://assets-eu.researchsquare.com/files/rs-7487727/v1/c80e9f82e89649ebc5e3aafd.docx"}],"financialInterests":"\u003cb\u003eYes\u003c/b\u003e there is potential Competing Interest.\nAuthor Andreas D. Lauritzen is employed as AI Research Engineer in ScreenPoint Medical, and holds shares in a company that holds depositary receipts in ScreenPoint Medical, the provider of the AI system evaluated in this study.\r\n\r\nAuthor Martin Lillholm holds shares in a company that holds depositary receipts in ScreenPoint Medical.\r\n\r\nAuthor Mads Nielsen holds shares in a company that holds depositary receipts in ScreenPoint Medical.\r\n\r\nAuthor Nico Karssemeijer is employed as Chief Science Officer in ScreenPoint Medical, and is a shareholder of ScreenPoint Medical.\r\n\r\nAuthor Ilse Vejborg has no competing interests.","formattedTitle":"Comprehensive evaluation of AI in a large regional mammography screening program for breast cancer: detection, interval cancers, and radiologist workload","fulltext":[{"header":"Introduction","content":"\u003cp\u003ePopulation-based mammography screening for breast cancer using artificial intelligence (AI) has been shown to substantially reduce radiologists' reading workload by triaging screenings for single reading instead of double reading\u003csup\u003e1–3\u003c/sup\u003e. More recently, published studies have suggested that using AI as decision support, in addition to triaging, increases cancer detection without increasing false-positive rates\u003csup\u003e1,4\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eWhile implementing AI into mammography screening has been effective in reducing workload and improving detection there is still a need to comprehensively evaluate the overall screening quality when using AI. Interval cancers are breast cancers diagnosed outside screening regime between screening rounds either due to being missed or deemed benign at screening, an aggressive/fast-growing subtype, or masked by dense breast tissue\u003csup\u003e5\u003c/sup\u003e. Therefore, the interval cancer ratio is an essential parameter to assess the true screening quality of a screening program\u003csup\u003e5,6\u003c/sup\u003e. This parameter has not been evaluated yet in published literature on mammography screening with AI in prospective settings. However, retrospective studies have suggested AI could be safely used maintaining or even decreasing interval cancer rates\u003csup\u003e7,8\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eAdditionally, breast cancer prognosis and mortality are strongly influenced by the cancer subtype. Breast cancer, being a heterogeneous disease, is characterized and treated based on multiple features such as receptor status, molecular subtypes, grading, lesion size, metastases, and other pathological and imaging properties\u003csup\u003e9\u003c/sup\u003e. Understanding whether the use of AI alters the characteristics of breast cancers at the time of diagnosis is essential. The optimal effect of AI-supported screening would be detection of breast cancers with features indicative of earlier-stage disease, and a shift towards fewer and less aggressive characteristics among interval cancers, indicating improved prognosis.\u003c/p\u003e\n\u003cp\u003eIn November 2021, AI was fully implemented into the largest regional mammography screening program in Denmark based on evidence from a prior simulation study\u003csup\u003e8\u003c/sup\u003e. All screenings have since been triaged using AI for either AI-assisted single reading or double reading with AI decision support. Early indicators of screening quality following implementation showed a significant increase in cancer detection rate with AI\u003csup\u003e1\u003c/sup\u003e. Now, with more than three years of AI-supported screening and sufficient follow-up, we evaluated whether this increase in detection was accompanied by a reduction in the two-year interval cancer ratio and improvements in screening sensitivity and specificity. Specifically, this is a prospective study with a comprehensive historical comparison of screening quality before and after AI implementation. Additionally, the study assessed changes in tumor characteristics for both screen-detected and interval cancers between the two periods.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eThis study concerned women screened for breast cancer in the Capital Region of Denmark. The screening program aims to invite asymptomatic women every two years (\u0026plusmn;3 months), although the actual interval between invitations may vary slightly. Women are screened using full-field digital mammography (Siemens Revelation or Inspiration). Radiographers capture at least a mediolateral oblique and a craniocaudal view for each breast.\u003c/p\u003e\n\u003cp\u003eBefore AI implementation, every screening was read by two specialized full-time breast radiologists, at least one of them being an experienced senior breast radiologist.\u003cbr\u003eThe AI system in this study was Transpara (v1.7.1, ScreenPoint Medical)\u003csup\u003e10\u0026ndash;13\u003c/sup\u003e. The AI provides an exam score from 1 to 10 for each screening. Exam score 10 indicates a high probability of breast cancer. Additionally, the AI provides decision support where suspicious findings are highlighted to the reader. After AI implementation on November 18, 2021, all screenings are analyzed by AI and receives an exam score after successful analysis. If the exam is initially triaged as low risk of a present breast cancer by the AI, which is defined as exam score 5 or below, the screening is read by a single senior experienced breast radiologist only. From May 2022, this threshold was changed to exam score 7 or below. All screenings not triaged for single reading, are read by two radiologists, one of them senior, both with access to decision support by the AI. Further information on reading protocols\u003csup\u003e1\u003c/sup\u003e and the AI system\u003csup\u003e1,8,10\u0026ndash;17\u003c/sup\u003e can be found in previously published literature.\u003c/p\u003e\n\u003cp\u003eIn case that radiologists see suspicious findings, the woman is recalled for diagnostic assessment. All breast cancers in this study were diagnosed based on clinical examination, supplemental mammographic imaging, ultrasound examination and needle biopsy, and include both invasive and in situ carcinomas all verified by histology.\u003c/p\u003e\n\u003ch2\u003eStudy populations and characteristics\u003c/h2\u003e\n\u003cp\u003eThis study comprised three cohorts (see Figure 1). The cohorts were women 1) screened before AI, 2) with AI with long follow-up, and 3) with AI with short follow-up. The cohorts of women screened with AI have been divided into two overlapping samples with different inclusions periods and lengths of follow-up; one for quality indicators that require two years of follow-up (e.g. interval cancers) and another for quality indicators requiring 180 days of follow-up (e.g. screen-detected cancers). This maximized the number of included women while allowing sufficient follow-up for each quality indicators.\u003c/p\u003e\n\u003cp\u003eThe inclusion period for the cohort before AI was November 17, 2019, to November 17, 2021, and initially included 118,631 screenings. After excluding women older than 70 years and 3 months at screening, screenings assessed by AI during the implementation period, and duplicated screenings, the sample included 114,823 women with one screening each.\u003c/p\u003e\n\u003cp\u003eThe inclusion period for the cohort with AI with long follow-up was November 18, 2021, to November 18, 2022, and initially included 68,951 screenings. After exclusions, the sample included 66,492 women with one screening each. This sample was used to calculate quality indicators related to sensitivity, specificity, and interval cancer ratios for which at least two years of follow-up was required.\u003cbr\u003e\u0026nbsp;\u003cbr\u003e\u0026nbsp;The inclusion period for the cohort with AI with short follow-up was November 18, 2021, to November 18, 2023, and initially included 161,655 screenings. After exclusions the sample included 156,151 women with one screening each. This sample was used for quality indicators related to recall and detection for which at least 180 days of follow-up was required. The AI system successfully analyzed 99.13% (154.798/156.151) of the screenings in this sample.\u003c/p\u003e\n\u003cp\u003eTable 1 presents the characteristics of the screened women in the three samples. In the sample before AI the proportion of women with a prevalent screening (first screening) was lower than either of the samples with AI (16.71% vs. 19.30%, P\u0026lt;0.001 and 16.71% vs. 18.40%, P\u0026lt;0.001). In the sample before AI, the average number of days since last screening was within 822 days and therefore within the aim of 2 years \u0026plusmn;3 months. In the samples with AI were 988 and 955 days and both were higher than before AI (P\u0026lt;0.001 and P\u0026lt;0.001). There was no difference in the number of women with a prior breast cancer surgery in the samples before AI and with AI (long-follow up) (P=0.95). There were less women with a prior breast cancer surgery in the sample with AI (short follow-up) compared to before AI (P\u0026lt;0.001).\u003cbr\u003e\u0026nbsp;\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 664px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTable 1: Characteristics of study samples\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCharacteristic\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBefore AI\u0026nbsp;\u003cbr\u003e\u0026nbsp;(N=114,823)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eWith AI, long follow-up (N=66,492)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 163px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eWith AI, short follow-up (N=156,151)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eAverage age\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e58.76\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e58.81\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 163px;\"\u003e\n \u003cp\u003e58.94\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; 50-54 (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e33,867 (29.49%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e19,267 (28.98%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 163px;\"\u003e\n \u003cp\u003e43,256 (27.70%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; 55-59 (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e31,038 (27.03%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e17,999 (27.07%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 163px;\"\u003e\n \u003cp\u003e43,390 (27.79%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; 60-64 (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e25,337 (22.07%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e14,494 (21.80%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 163px;\"\u003e\n \u003cp\u003e34,968 (22.39%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; 65-70 (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e24,581 (21.41%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e14,732 (22.16%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 163px;\"\u003e\n \u003cp\u003e34,537 (22.12%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eAverage BI-RADS density*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e1.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e1.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 163px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; 1 (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e47,674 (41.52%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e28,000 (42.11%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 163px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; 2 (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e41,783 (36.39%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e24,759 (37.24%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 163px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; 3 (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e21,285 (18.54%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e11,128 (16.74%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 163px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; 4 (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e3,951 (3.44%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e2,539 (3.82%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 163px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Unknown (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e130 (0.11%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e66 (0.10%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 163px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003ePrevalent screenings (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e19,183 (16.71%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e12,832 (19.30%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 163px;\"\u003e\n \u003cp\u003e28,735 (18.40%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eAverage number of days since last screening\u003csup\u003e\u0026Dagger;\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e821\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e986\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 163px;\"\u003e\n \u003cp\u003e953\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eWomen with prior breast cancer surgery\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e5,044 (4.39%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e2,925 (4.40%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 163px;\"\u003e\n \u003cp\u003e6,048 (3.87%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 664px;\"\u003e\n \u003cp\u003e* Based only on second reader BI-RADS density assignments\u003c/p\u003e\n \u003cp\u003e\u003csup\u003e\u0026Dagger;\u003c/sup\u003e Based only on women screened in the previous round\u003cbr\u003e\u0026nbsp;Note, Radiologist\u0026apos;s BI-RADS density assignments were not available for the full period of screening with AI with short follow-up.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003ch2\u003eScreening quality indicators\u003c/h2\u003e\n\u003cp\u003eTable 2 presents key quality indicators comparing screening performances before and after AI implementation over the long follow-up period (2 years). The sensitivity (screen-detected cancers out of all cancers within two years after screening) increased from 69.97% (751/1,078) to 73.34% (553/754) with a one-sided Fisher\u0026apos;s exact test (P=0.049). The specificity also increased from 97.86% (111,349/113,745) to 98.34% (64,648/65,738), with P\u0026lt;0.001.\u003cbr\u003e\u0026nbsp;The interval cancer ratio (interval cancers diagnosed within two years after screening) decreased from 30.33% (327/1,078) to 26.66% (201/754), with P=0.049. The first year interval cancer ratio decreased from 13.28% (115/866) to 10.37% (64/617), a 22% relative decrease, however non-significant (P=0.053).\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 661px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTable 2: Screening quality indicators related to sensitivity, specificity, and interval cancers based on long follow-up (2 years)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 188px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIndicator\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBefore AI\u003cbr\u003e\u0026nbsp;(N=114,823)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eWith AI, long follow-up\u003cbr\u003e\u0026nbsp;(N=66,492)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 95px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eP-value *\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 188px;\"\u003e\n \u003cp\u003eSensitivity (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e69.67\u003c/p\u003e\n \u003cp\u003e(751/1,078, [66.86, 72.48])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e73.34\u003c/p\u003e\n \u003cp\u003e(553/754, [70.07, 76.61])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 95px;\"\u003e\n \u003cp\u003e0.049\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 188px;\"\u003e\n \u003cp\u003eSpecificity (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e97.89\u003c/p\u003e\n \u003cp\u003e(111,349/113,745, [97.81, 97.98])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e98.34\u003c/p\u003e\n \u003cp\u003e(64,648/65,738, [98.24, 98.44])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 95px;\"\u003e\n \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 188px;\"\u003e\n \u003cp\u003eInterval cancer ratio (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e30.33\u003c/p\u003e\n \u003cp\u003e(327/1,078, [27.66, 33.00])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e26.66\u003c/p\u003e\n \u003cp\u003e(201/754, [23.63, 29.69])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 95px;\"\u003e\n \u003cp\u003e0.049\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 188px;\"\u003e\n \u003cp\u003eFirst year interval cancer ratio (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e13.28\u003c/p\u003e\n \u003cp\u003e(115/866, [11.18, 15.38])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e10.37\u003c/p\u003e\n \u003cp\u003e(64/617, [8.21, 12.54])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 95px;\"\u003e\n \u003cp\u003e0.053\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 661px;\"\u003e\n \u003cp\u003eNote, data in parentheses are numerator and denominator values, and the 95% confidence interval is in brackets. P-values have been calculated using the Fishers exact test. For all indicators, breast cancers include both invasive and in situ cancers.\u003c/p\u003e\n \u003cp\u003e* One-sided Fishers exact test directions are predefined: Alternative hypotheses for sensitivity and specificity are that rates with AI are higher, and for interval cancer ratios the rates with AI are lower.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eAs presented in Figure 2, when stratified by BIRADS density, the sensitivity increased across all density groups when screening with AI. The largest increase was observed for the group with density 4 where the sensitivity changed from 41.3% before AI to 57.1% with AI which corresponds to a 38.3% relative increase. However, none of the differences were significant (P\u0026ge;0.12). \u0026nbsp;The specificity increased for density groups 1, 2 and 3 (density 1: P\u0026lt;0.01, density 2: P\u0026lt;0.001, density 3: P\u0026lt;0.001), while the change in density 4 was not significant (P=0.21). The overall interval cancer ratio decreased across all density groups, corresponding to the increase in the sensitivity, but without reaching significance (P\u0026ge;0.12). However, in the density 4 group, the interval cancer ratio decreased from 58.7% to 42.9% (P=0.12), which corresponds to a 37.0% decrease. First year interval cancer ratios also decreased across all density groups but none of the changes were statistically significant (P\u0026ge;0.08). The largest reduction was observed in density 4 group where the first year interval cancer ratio decreased from 32.1% to 20.0% (P=0.25). The statistical tests for sensitivity and interval cancer ratios were likely underpowered to detect differences in the stratified samples.\u003cbr\u003e\u0026nbsp;Screening sensitivity was modelled using logistic regression, with Method (with AI vs. before AI) and BIRADS density as predictors, which provides the relative benefit of AI varied with BIRADS density. A positive trend was observed, showing that the sensitivity gains with AI increased with higher breast density (OR for trend per BIRADS density step = 1.18, 95% CI 1.06, 1.31, P=0.004).\u003c/p\u003e\n\u003cp\u003eTable 3 presents key quality indicators comparing screening performances before and after AI implementation over the short follow-up period (180 days). When screening with AI (short follow-up), the recall rate decreased from 2.75% (3,160/114,823) to 2.37% (3,696/156,151), with P\u0026lt;0.001, compared to before AI implementation. The overall cancer detection rate increased from 6.54 per 1,000 (751/114,823) to 7.80 per 1,000 (1,218/156,151), with P\u0026lt;0.001. For prevalent screenings, the cancer detection rate increased but non-significantly from 6.62 per 1,000 (127/19,183) to 7.52 per 1,000 (216/28,735) (P=0.25). The cancer detection rate for incident screenings increased from 6.52 per 1,000 (624/95,640) to 7.86 per 1,000 (1,002/127,416), with P\u0026lt;0.001. The false-positive rate decreased from 2.10% (2,409/114,823) to 1.59% (2,478/156,151), with P\u0026lt;0.001. The positive predictive value of recall increased from 23.77% (751/3,160) to 32.95% (1,218/3,696), with P\u0026lt;0.001. The consensus meeting/arbitration rate changed from 3.79% (4,351/114,823) to 3.86% (6,022/156,151) which was not significant (P=0.37).\u003c/p\u003e\n\u003cp\u003eRadiologists\u0026apos; reading workload was reduced by 35.78% (111,752/312,302 readings saved) following AI implementation. Before the threshold was increased to 7, the reading workload reduction was 27.96% (14,616/52,282 readings saved), and 37.36% (97,136/260,020 readings saved) after the increase.\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 661px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTable 3: Screening quality indicators related to recall, detection, and workload change based on short follow-up (180 days)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 217px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIndicator\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBefore AI\u003cbr\u003e\u0026nbsp;(N=114,823)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eWith AI, short follow-up\u003cbr\u003e\u0026nbsp;(N=156,151)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eP-value *\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 217px;\"\u003e\n \u003cp\u003eRecall rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e2.75\u003c/p\u003e\n \u003cp\u003e(3160/114823, [2.66, 2.85])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e2.37\u003c/p\u003e\n \u003cp\u003e(3696/156151, [2.29, 2.44])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 113px;\"\u003e\n \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 217px;\"\u003e\n \u003cp\u003eCancer detection rate\u003cbr\u003e(per 1000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e6.54\u003c/p\u003e\n \u003cp\u003e(751/114823, [6.09, 6.99])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e7.80\u003c/p\u003e\n \u003cp\u003e(1218/156151, [7.38, 8.22])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 113px;\"\u003e\n \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 217px;\"\u003e\n \u003cp\u003eCancer detection rate\u003cbr\u003e\u0026nbsp;(prevalent screenings, per 1000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e6.62\u003c/p\u003e\n \u003cp\u003e(127/19183, [5.57, 7.67])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e7.52\u003c/p\u003e\n \u003cp\u003e(216/28735, [6.58, 8.45])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 113px;\"\u003e\n \u003cp\u003e0.12\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 217px;\"\u003e\n \u003cp\u003eCancer detection rate\u003cbr\u003e\u0026nbsp;(incident screenings, per 1000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e6.52\u003c/p\u003e\n \u003cp\u003e(624/95640, [6.03, 7.02])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e7.86\u003c/p\u003e\n \u003cp\u003e(1002/127416, [7.39, 8.33])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 113px;\"\u003e\n \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 217px;\"\u003e\n \u003cp\u003eFalse-positive rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e2.10\u003c/p\u003e\n \u003cp\u003e(2409/114823, [2.02, 2.18])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e1.59\u003c/p\u003e\n \u003cp\u003e(2478/156151, [1.53, 1.65])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 113px;\"\u003e\n \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 217px;\"\u003e\n \u003cp\u003ePositive predictive value of recall (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e23.77\u003c/p\u003e\n \u003cp\u003e(751/3160, [22.31, 25.22])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e32.95\u003c/p\u003e\n \u003cp\u003e(1218/3696, [31.46, 34.45])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 113px;\"\u003e\n \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 217px;\"\u003e\n \u003cp\u003eConsensus meeting/arbitration rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e3.79\u003c/p\u003e\n \u003cp\u003e(4351/114823, [3.68, 3.90])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e3.86\u003c/p\u003e\n \u003cp\u003e(6022/156151, [3.76, 3.95])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 113px;\"\u003e\n \u003cp\u003e0.19\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 217px;\"\u003e\n \u003cp\u003eReading workload reduction (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e35.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 113px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 661px;\"\u003e\n \u003cp\u003eNote, data in parentheses are numerator and denominator values, and the 95% confidence interval is in brackets. P-values have been calculated using the Fishers exact test. For all indicators, breast cancers include both invasive and in situ cancers.\u003c/p\u003e\n \u003cp\u003e* One-sided Fishers exact test directions are predefined: Alternative hypotheses for cancer detection rates and positive predictive value are that rates with AI are higher, and for recall, false-positives, consensus meeting/arbitration the rates with AI are lower.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003ch2\u003eCancer characteristics\u003c/h2\u003e\n\u003cp\u003eTable 5 present characteristics of screen-detected cancers before and with AI (short follow-up). The rate of screen-detected in situ carcinomas increased from 16.78% to 19.13% but not significantly (P=0.19) after AI implementation. Rate of estrogen receptor (ER) positive increased and the rate of human epidermal growth factor receptor 2 (HER2) positive invasive cancers decreased (P=0.03 and P=0.04, respectively). The rate of invasive cancers characterized as luminal A (ER positive + HER2 negative) increased from 85.03% to 89.29% (P=0.02). Conversely, the rate non-luminal A invasive cancers decreased from 14.05% to 10.25% (P=0.03). Neither the rate of invasive cancers characterized as luminal B (ER positive + HER2 positive) nor the rate of double negative invasive cancers changed significantly (P=0.31 and P=0.46, respectively). There were no significant changes regarding invasive cancer malignancy grades (grade 1: P=0.25, grade 2: P=0.97, grade 3: P=0.29), rate of small invasive tumors (\u0026le;10mm, P=0.9), or lymph-node involvement for invasive cancers (P=0.82). The rate of women who received neoadjuvant chemotherapy (NACT) decreased from 7.06% to 5.91%, a 16.3% relative decrease, which was not significant (P=0.31). The rate of DCIS classified as Van Nuys prognostic index 3 increased from 47.41% before AI to 55.15% with AI, which corresponds to a 16.3% increase, but the change was not significant (P=0.19).\u0026nbsp;\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 664px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTable 5: Characteristics of screen-detected cancers before AI vs. with AI based on short follow-up (180 days)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCancer characteristic\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBefore AI (N=114,823)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eWith AI (N=156,151)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eP-value\u0026nbsp;\u003c/strong\u003e\u003csup\u003e\u0026sect;\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eNumber of all cancers\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e751\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e1218\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eNumber of invasive cancers\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e625\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e985\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eNumber of valid* invasive cancers\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e541\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e878\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eNumber of in situ cancers\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e126\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e233\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eNumber of valid* in situ cancers\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e116\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e194\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eInvasive rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e83.22\u003c/p\u003e\n \u003cp\u003e(625/751, [80.38, 86.06])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e80.87\u003c/p\u003e\n \u003cp\u003e(985/1218, [78.57, 83.17])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.19\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eIn situ rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e16.78\u003c/p\u003e\n \u003cp\u003e(126/751, [14.28, 19.28])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e19.13\u003c/p\u003e\n \u003cp\u003e(233/1218, [17.02, 21.24])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.19\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eER positive rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e92.05\u003c/p\u003e\n \u003cp\u003e(498/541, [89.46, 94.64])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e94.99\u003c/p\u003e\n \u003cp\u003e(834/878, [93.34, 96.64])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.03\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eER unknown rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e0.37\u003c/p\u003e\n \u003cp\u003e(2/541, [0.10, 0.64])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e0.11\u003c/p\u003e\n \u003cp\u003e(1/878, [0.02, 0.21])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.31\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eHER2 positive rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e9.61\u003c/p\u003e\n \u003cp\u003e(52/541, [7.41, 11.82])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e6.61\u003c/p\u003e\n \u003cp\u003e(58/878, [5.14, 8.07])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.04\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eHER2 unknown rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e0.74\u003c/p\u003e\n \u003cp\u003e(4/541, [0.29, 1.19])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e0.34\u003c/p\u003e\n \u003cp\u003e(3/878, [0.12, 0.57])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.30\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eLuminal A rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e85.03\u003c/p\u003e\n \u003cp\u003e(460/541, [81.77, 88.28])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e89.29\u003c/p\u003e\n \u003cp\u003e(784/878, [87.07, 91.51])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.02\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eNon-luminal A rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e14.05\u003c/p\u003e\n \u003cp\u003e(76/541, [11.37, 16.72])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e10.25\u003c/p\u003e\n \u003cp\u003e(90/878, [8.41, 12.09])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.03\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Luminal B rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e6.65\u003c/p\u003e\n \u003cp\u003e(36/541, [4.85, 8.46])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e5.35\u003c/p\u003e\n \u003cp\u003e(47/878, [4.05, 6.66])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.31\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Double negative rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e4.44\u003c/p\u003e\n \u003cp\u003e(24/541, [3.00, 5.87])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e3.64\u003c/p\u003e\n \u003cp\u003e(32/878, [2.59, 4.70])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.46\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; ER negative \u0026amp; HER2 positive rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e2.96\u003c/p\u003e\n \u003cp\u003e(16/541, [1.83, 4.09])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e1.25\u003c/p\u003e\n \u003cp\u003e(11/878, [0.70, 1.80])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.02\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eLuminal unknown rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e0.92\u003c/p\u003e\n \u003cp\u003e(5/541, [0.40, 1.45])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e0.46\u003c/p\u003e\n \u003cp\u003e(4/878, [0.18, 0.73])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.28\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eGrade 1 rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e43.81\u003c/p\u003e\n \u003cp\u003e(237/541, [39.69, 47.93])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e46.92\u003c/p\u003e\n \u003cp\u003e(412/878, [43.64, 50.21])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.25\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eGrade 2 rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e43.62\u003c/p\u003e\n \u003cp\u003e(236/541, [39.50, 47.74])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e43.74\u003c/p\u003e\n \u003cp\u003e(384/878, [40.49, 46.98])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eGrade 3 rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e9.98\u003c/p\u003e\n \u003cp\u003e(54/541, [7.73, 12.23])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e8.31\u003c/p\u003e\n \u003cp\u003e(73/878, [6.66, 9.96])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.29\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eGrade unknown rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e2.59\u003c/p\u003e\n \u003cp\u003e(14/541, [1.55, 3.63])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e1.03\u003c/p\u003e\n \u003cp\u003e(9/878, [0.54, 1.51])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.02\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eSmall cancer rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e43.07\u003c/p\u003e\n \u003cp\u003e(233/541, [38.96, 47.18])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e43.39\u003c/p\u003e\n \u003cp\u003e(381/878, [40.15, 46.64])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.90\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Small cancer rate (including NACT-patients\u003csup\u003e\u0026Dagger;\u003c/sup\u003e) (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e39.09\u003c/p\u003e\n \u003cp\u003e(233/596, [35.26, 42.93])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e40.15\u003c/p\u003e\n \u003cp\u003e(381/949, [37.07, 43.22])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.68\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eSize unknown rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e0.18\u003c/p\u003e\n \u003cp\u003e(1/541, [0.03, 0.34])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e0.11\u003c/p\u003e\n \u003cp\u003e(1/878, [0.02, 0.21])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.73\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eLymph node-negative rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e77.63\u003c/p\u003e\n \u003cp\u003e(420/541, [73.93, 81.33])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e77.11\u003c/p\u003e\n \u003cp\u003e(677/878, [74.21, 80.00])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eNode unknown rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e1.11\u003c/p\u003e\n \u003cp\u003e(6/541, [0.51, 1.71])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e1.03\u003c/p\u003e\n \u003cp\u003e(9/878, [0.54, 1.51])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003eNACT-patient rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e7.06\u003c/p\u003e\n \u003cp\u003e(53/751, [5.44, 8.68])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e5.91\u003c/p\u003e\n \u003cp\u003e(72/1218, [4.72, 7.10])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.31\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 664px;\"\u003e\n \u003cp\u003eDCIS Van Nuys\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Grade 1 rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e12.93\u003c/p\u003e\n \u003cp\u003e(15/116, [8.00, 17.87])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e10.82\u003c/p\u003e\n \u003cp\u003e(21/194, [7.19, 14.46])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.58\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Grade 2 rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e38.79\u003c/p\u003e\n \u003cp\u003e(45/116, [30.42, 47.17])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e32.47\u003c/p\u003e\n \u003cp\u003e(63/194, [26.28, 38.67])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.26\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Grade 3 rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e47.41\u003c/p\u003e\n \u003cp\u003e(55/116, [38.56, 56.27])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e55.15\u003c/p\u003e\n \u003cp\u003e(107/194, [48.12, 62.19])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.19\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 289px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Grade unknown rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003e0.86\u003c/p\u003e\n \u003cp\u003e(1/116, [0.15, 1.57])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e1.55\u003c/p\u003e\n \u003cp\u003e(3/194, [0.53, 2.57])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.61\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 664px;\"\u003e\n \u003cp\u003eNote, data in parentheses are numerator and denominator values, and the 95% confidence interval is in brackets. DCIS = ductal carcinoma in situ. NACT = Neoadjuvant chemotherapy. ER = Estrogen receptor. HER2 = Human Epidermal Growth Factor Receptor 2.\u003c/p\u003e\n \u003cp\u003e* Valid cancers refer to women with diagnosed breast carcinomas that are not NACT-patients, having had surgery, and for which the surgical specimen contains residual of the cancer diagnosed at the preceding biopsy. See Supplemental Table 1 for further details.\u003c/p\u003e\n \u003cp\u003e\u003csup\u003e\u0026Dagger;\u0026nbsp;\u003c/sup\u003eNACT-patients count as having a tumor size \u0026gt; 10mm.\u003c/p\u003e\n \u003cp\u003e\u003csup\u003e\u0026sect;\u003c/sup\u003e P-values have been calculated using the chi-square test.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eTable 6 present characteristics of interval cancers before and with AI (long follow-up). None of the measured cancer characteristics changes was concluded to be statistically significant. However, the rate of HER2 positive invasive cancers changed from 9.52% before AI to 5.34% with AI (P=0.16) which corresponds to a 43.9% relative decrease. The rate of luminal B invasive cancer decreased from 6.67% to 3.82% (P=0.26) which is a 42.7% relative decrease. The rate of small invasive cancers (\u0026le; 10mm) decreased from 27.14% to 21.37% (P=0.23) which was a relative decrease of 21.3% The rate of DCIS classified as Van Nuys prognostic index 3 increased from 29.41% before AI to 42.86% with AI, which corresponds to a 45.7% relative increase, but the change was not significant (P=0.44).\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" class=\"fr-table-selection-hover\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 664px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTable 6: Characteristics of interval cancers before AI vs. with AI based on long follow-up (2 years)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCancer characteristic\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBefore AI (N=114,823)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eWith AI (N=66,492)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eP-value \u003csup\u003e\u0026sect;\u003c/sup\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eNumber of all cancers\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e327\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e201\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eNumber of invasive cancers\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e309\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e185\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eNumber of valid* invasive cancers\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e210\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e131\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eNumber of in situ cancers\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eNumber of valid* in situ cancers\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e17\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eInvasive rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e94.50\u003c/p\u003e\n \u003cp\u003e(309/327, [91.47, 97.52])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e92.04\u003c/p\u003e\n \u003cp\u003e(185/201, [87.46, 96.62])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e0.26\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eIn situ rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e5.50\u003c/p\u003e\n \u003cp\u003e(18/327, [3.51, 7.50])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e7.96\u003c/p\u003e\n \u003cp\u003e(16/201, [4.96, 10.96])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e0.26\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eER positive rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e89.52\u003c/p\u003e\n \u003cp\u003e(188/210, [84.65, 94.40])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e87.79\u003c/p\u003e\n \u003cp\u003e(115/131, [81.08, 94.49])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e0.62\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eER unknown rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e0.00\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;(0/210, [0.00, 0.00])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e0.00\u003c/p\u003e\n \u003cp\u003e(0/131, [0.00, 0.00])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eHER2 positive rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e9.52\u003c/p\u003e\n \u003cp\u003e(20/210, [6.25, 12.80])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e5.34\u003c/p\u003e\n \u003cp\u003e(7/131, [2.61, 8.07])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e0.16\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eHER2 unknown rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e0.00\u003c/p\u003e\n \u003cp\u003e(0/210, [0.00, 0.00])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e0.00\u003c/p\u003e\n \u003cp\u003e(0/131, [0.00, 0.00])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eLuminal A rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e82.86\u003c/p\u003e\n \u003cp\u003e(174/210, [77.18, 88.53])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e83.21\u003c/p\u003e\n \u003cp\u003e(109/131, [75.88, 90.53])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eNon-luminal A rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e17.14\u003c/p\u003e\n \u003cp\u003e(36/210, [12.65, 21.64])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e16.03\u003c/p\u003e\n \u003cp\u003e(21/131, [10.73, 21.33])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e0.79\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Luminal B rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e6.67\u003c/p\u003e\n \u003cp\u003e(14/210, [4.01, 9.32])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e3.82\u003c/p\u003e\n \u003cp\u003e(5/131, [1.64, 5.99])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e0.26\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Double negative rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e7.62\u003c/p\u003e\n \u003cp\u003e(16/210, [4.74, 10.49])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e10.69\u003c/p\u003e\n \u003cp\u003e(14/131, [6.47, 14.90])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e0.33\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; ER negative \u0026amp; HER2 positive rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e2.86\u003c/p\u003e\n \u003cp\u003e(6/210, [1.32, 4.40])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e1.53\u003c/p\u003e\n \u003cp\u003e(2/131, [0.42, 2.63])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e0.43\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eLuminal unknown rate\u0026nbsp;(%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e0.00\u003c/p\u003e\n \u003cp\u003e(0/210, [0.00, 0.00])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e0.76\u003c/p\u003e\n \u003cp\u003e(1/131, [0.13, 1.39])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eGrade 1 rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e34.76\u003c/p\u003e\n \u003cp\u003e(73/210, [28.65, 40.88])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e35.11\u003c/p\u003e\n \u003cp\u003e(46/131, [27.47, 42.76])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eGrade 2 rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e48.10\u003c/p\u003e\n \u003cp\u003e(101/210, [41.43, 54.76])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e48.09\u003c/p\u003e\n \u003cp\u003e(63/131, [39.71, 56.47])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e\u0026gt; 0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eGrade 3 rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e16.19\u003c/p\u003e\n \u003cp\u003e(34/210, [11.82, 20.56])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e16.79\u003c/p\u003e\n \u003cp\u003e(22/131, [11.36, 22.23])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eGrade unknown rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003cp\u003e(2/210, [0.26, 1.64])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e0.00\u003c/p\u003e\n \u003cp\u003e(0/131, [0.00, 0.00])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eSmall cancer rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e27.14\u003c/p\u003e\n \u003cp\u003e(57/210, [21.58, 32.71])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e21.37\u003c/p\u003e\n \u003cp\u003e(28/131, [15.22, 27.53])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e0.23\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Small cancer rate (including NACT-patients\u003csup\u003e\u0026Dagger;\u003c/sup\u003e) (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e21.19\u003c/p\u003e\n \u003cp\u003e(57/269, [16.73, 25.65])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e16.87\u003c/p\u003e\n \u003cp\u003e(28/166, [11.94, 21.80])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e0.27\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eSize unknown rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e0.48\u003c/p\u003e\n \u003cp\u003e(1/210, [0.08, 0.87])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e0.00\u003c/p\u003e\n \u003cp\u003e(0/131, [0.00, 0.00])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eNode-negative rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e63.33\u003c/p\u003e\n \u003cp\u003e(133/210, [56.63, 70.04])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e63.36\u003c/p\u003e\n \u003cp\u003e(83/131, [54.84, 71.88])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e\u0026gt; 0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eNode unknown rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e1.43\u003c/p\u003e\n \u003cp\u003e(3/210, [0.49, 2.37])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e1.53\u003c/p\u003e\n \u003cp\u003e(2/131, [0.42, 2.63])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003eNACT-patient rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e15.90\u003c/p\u003e\n \u003cp\u003e(52/327, [12.34, 19.47])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e16.42\u003c/p\u003e\n \u003cp\u003e(33/201, [11.94, 20.90])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 664px;\"\u003e\n \u003cp\u003eDCIS Van Nuys\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Grade 1 rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e29.4\u003c/p\u003e\n \u003cp\u003e(5/17, [13.28, 45.54])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e28.57\u003c/p\u003e\n \u003cp\u003e(4/14, [11.72, 45.42])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Grade 2 rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e35.29\u003c/p\u003e\n \u003cp\u003e(6/17, [17.31, 53.28])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e21.43\u003c/p\u003e\n \u003cp\u003e(3/14, [7.57, 35.29])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e0.40\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Grade 3 rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e29.41\u003c/p\u003e\n \u003cp\u003e(5/17, [13.28, 45.54])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e42.86\u003c/p\u003e\n \u003cp\u003e(6/14, [21.38, 64.33])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e0.44\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 291px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Grade unknown rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e5.88\u003c/p\u003e\n \u003cp\u003e(1/17, [1.05, 10.72])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 151px;\"\u003e\n \u003cp\u003e7.14\u003c/p\u003e\n \u003cp\u003e(1/14, [1.27, 13.01])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 61px;\"\u003e\n \u003cp\u003e0.89\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 664px;\"\u003e\n \u003cp\u003eNote, data in parentheses are numerator and denominator values, and the 95% confidence interval is in brackets. DCIS = ductal carcinoma in situ. NACT = Neoadjuvant chemotherapy. ER = Estrogen receptor. HER2 = Human Epidermal Growth Factor Receptor 2.\u003c/p\u003e\n \u003cp\u003e* Valid cancers refer to women with diagnosed breast carcinomas that are not NACT-patients, having had surgery, and for which the surgical specimen contains residual of the cancer diagnosed at the preceding biopsy. See Supplemental Table 1 for further details.\u003c/p\u003e\n \u003cp\u003e\u003csup\u003e\u0026Dagger;\u0026nbsp;\u003c/sup\u003eNACT-patients count as having a tumor size \u0026gt; 10mm.\u003c/p\u003e\n \u003cp\u003e\u003csup\u003e\u0026sect;\u003c/sup\u003e P-values have been calculated using the chi-square test.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we presented a comprehensive analysis of the impact of implementing AI in mammography screening across a large regional program with two years of follow-up. This study demonstrates that introducing AI as a tool for both triage and decision-support reduced radiologists' workload by 36% while enhancing screening performance. This is made evident by improvements in sensitivity (69.7% to 73.3%, P=0.049), cancer detection rates (6.54 to 7.80 per 1,000, P\u0026lt;0.001), interval cancer ratios (30.3% to 26.7%, P=0.049), recall rates (2.75% to 2.37%, P\u0026lt;0.001), and false-positive rates (2.10% to 1.59%, P\u0026lt;0.001).\u003cbr\u003e\u0026nbsp;\u003cbr\u003eThe results of this study largely agrees with recent large-scale trials. In the MASAI trial, AI-supported workflow decreased reading workload by 44% while increasing cancer detection by 29% (5.0 to 6.4 per 1,000 screenings) without increasing the number of false-positives or recalls\u003csup\u003e2,4\u003c/sup\u003e. Likewise, our study observed that cancer detection rates increased from 6.54 to 7.80 per 1,000 (P\u0026lt;0.001) along with lower recall and false-positive rates.\u003cbr\u003eIn a large Swedish prospective trial (ScreenTrustCAD), showed that replacing one radiologist with AI, compared to double reading, achieved non-inferior cancer detection rates (relative ratio 1.04, 95% CI: 1.00, 1.09)\u003csup\u003e3\u003c/sup\u003e.\u003cbr\u003eThese two Scandinavian studies along with a number of other international prospective and retrospective studies supports our results that AI can be used for triage and act as a substitute reader to reduce workload, and provide decision support to maintain screening quality or even improve it\u003csup\u003e18–24\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u0026nbsp;In this study, analysis of tumor subtypes in invasive screen-detected cancers showed a shift toward more favorable biological profiles where the ER-positive rate increased from 92% to 95% (P=0.03) and the rate luminal A positives increased from 85% to 89% (P=0.02).\u003cbr\u003e\u0026nbsp;Results suggest that for interval cancers, there was a trend towards fewer HER2-positive and luminal B subtypes. These changes were not significant, so larger samples sizes are needed to assess AI’s impact on aggressive tumor profiles.\u003cbr\u003e\u0026nbsp;\u003cbr\u003e\u0026nbsp;Across the BIRADS density categories, interval cancer ratios were consistently equal or lower when screening with AI compared to before AI, though statistical power was limited. These trends may suggest that AI can support radiologists in detecting cancers during screening, as opposed to the cancers being diagnosed as interval cancers, especially for women with high-density breasts.\u003cbr\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThis study’s strengths include its large-scale regional coverage, but most importantly, to our knowledge, the first study to present 2-year interval cancer ratios when screening with AI.\u003c/p\u003e\n\u003cp\u003eLimitations include, firstly, that this is single institution (but multi-site) screening program and observations may not correspond to other screening workflows or populations. However, the results align with the recent MASAI and ScreenTrustCAD trials, and the concurrent findings of the studies enhances the evidence that AI can improve screening across different health care systems, workflows, and populations. Secondly, this is a historical comparison of two cohorts (before and with AI) and these cohort differ in terms of a longer screening interval which may increase both cancer detection and interval cancer rates as tumors have more time to develop. There are also a larger proportion of prevalent screenings in the samples with AI which could be a source of increased detection, however, the cancer detection rate also increased for incident screenings alone. Lastly, there were less women with a prior breast cancer in the sample with AI. However, as these women have a higher overall risk this should decrease cancer detection\u0026nbsp;which was not observed in this study. On the contrary, the cancer detection rate increased (P \u0026lt; 0.001).\u003cbr\u003e\u0026nbsp;Additionally, this study showed a slight and non-significant increase in DCIS detection (17% to 19%, P=0.19). This trend should be carefully monitored to prevent overdiagnosis.\u0026nbsp;\u003cbr\u003e\u0026nbsp;\u003cbr\u003e\u0026nbsp;This study provides additional evidence that AI-supported mammography screening workflows can enhance detection and shift detected breast cancer subtypes towards more favorable subtypes while considerably reducing radiologist reading workload. Most importantly, we provide novel evidence that AI can further reduce the number of interval cancer which indicates earlier detection and thereby improve prognosis. Future work should focus on long-term clinical outcomes for women in AI-supported population-based breast cancer screening.\u003c/p\u003e"},{"header":"Methods","content":"\u003ch2\u003eStudy Approvals\u003c/h2\u003e\n\u003cp\u003eThis study received approval from The Danish Patient Safety Authority and Danish Data Protection Agency, who also waived the need for informed consent from women participating in mammography screening in the Capital Region of Denmark (reference 3\u0026ndash;3013\u0026ndash;2118, addendum 2019/2023).\u003c/p\u003e\n\u003ch2\u003eOutcomes\u003c/h2\u003e\n\u003cp\u003eThe primary outcomes of this study were the screening sensitivity, specificity, and interval cancer ratio in the samples before AI and with AI with long follow-up (two years). Note that breast cancers include both invasive and in situ carcinomas. The sensitivity is calculated as the percentage of breast cancers detected at screening out of all breast cancers diagnosed within two years from the screening date. The specificity is calculated as the percentage of true negative exams (not recalled and no cancers) out of all negative exams within the 2-year follow-up period. The interval cancer ratio is calculated as the percentage of interval cancers (cancers diagnosed outside screening regime) out of all breast cancers diagnosed within the 2-year follow-up period.\u0026nbsp;\u003cbr\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe secondary outcomes are the recall rate, cancer detection rate, false-positive rate, positive predictive value of recall, consensus meeting/arbitration rate, and radiologists reading workload in the samples before AI and with AI with short follow-up (180 days). The recall rate is calculated as the percentage of recalled women out of all women in the sample. The cancer detection rate is the number of breast cancers diagnosed at screening out of all screenings in the sample per 1,000 screenings. The false-positive rate is calculated as the percentage of women who were recalled but did not receive a breast cancer diagnosis within 180 days from the screening date out of all screenings in the sample. The positive predictive value is calculated as the percentage of women who were recalled and received a breast cancer diagnosis within 180 days out of all women who were recalled. The consensus meeting/arbitration rate were calculated as the number of screenings for which a consensus meeting was held, or an additional radiologist made the final decision on recall out of all screenings in the sample. Before AI, a consensus meeting/arbitration triggers when the two readers disagree on recall. with AI a consensus meeting/arbitration triggers when either the single reader decides to recall (AI triaged the screening to single reading as likely normal), or when the two readers disagreed for screenings for which the AI selected the exam for AI-assisted double reading. The reading workload reduction is calculated as the percentage of reads that were saved compared to standard double reading (e.g. a 36% reduction means that for 50 screenings that would normally be double read, the AI triaged 36 screenings for single reading. Thus, with AI, the total number of independent reads would be 36*1+14*2=64 reads which correspond to a 36% reduction of the 100 independent reads before AI.\u003c/p\u003e\n\u003ch2\u003eStatistical Analysis\u003c/h2\u003e\n\u003cp\u003eThe samples used in this study along with the inclusion periods, exclusion criteria, and follow-up times were all predefined prior to analysis. All the quality indicators were defined based on national guidelines\u003csup\u003e25\u003c/sup\u003e.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003ePrimary and secondary screening quality indicators were compared between study samples using the one-sided Fisher\u0026rsquo;s exact tests\u003csup\u003e26\u003c/sup\u003e. A previous publication on screening with AI in the Capital Region of Denmark, suggested that cancer detection increased as a results of implementing AI\u003csup\u003e1\u003c/sup\u003e. Based on those results, the test directions were predefined before analysis such that the alternative hypotheses for sensitivity, specificity, cancer detection rate, and positive predictive value of recall were that rates in the samples with AI were higher than the sample before AI. The alternative hypotheses for interval cancer ratios, recall rates, false-positive rates, and consensus meeting/arbitration were that rates were lower.\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003eFor subgroup analyses (e.g., BI-RADS density strata), Fisher\u0026rsquo;s exact tests were used given smaller sample sizes.\u003c/p\u003e\n\u003cp\u003eIn the analyses of breast cancer characteristics, only the chi-squared test was used to test for any differences in values regardless of direction of change.\u003c/p\u003e\n\u003cp\u003eStatistical significance was set at P\u0026lt;0.05 for key metrics. Analyses were performed in Python (version 3.11).\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eLauritzen, A. D. \u003cem\u003eet al.\u003c/em\u003e Early Indicators of the Impact of Using AI in Mammography Screening for Breast Cancer. \u003cem\u003eRadiology\u003c/em\u003e \u003cstrong\u003e311\u003c/strong\u003e, (2024).\u003c/li\u003e\n \u003cli\u003eL\u0026aring;ng, K. \u003cem\u003eet al.\u003c/em\u003e Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI): a clinical safety analysis of a randomised, controlled, non-inferiority, single-blinded, screening accuracy study. \u003cem\u003eLancet Oncol\u003c/em\u003e \u003cstrong\u003e24\u003c/strong\u003e, 936\u0026ndash;944 (2023).\u003c/li\u003e\n \u003cli\u003eDembrower, K., Crippa, A., Col\u0026oacute;n, E., Eklund, M. \u0026amp; Strand, F. Artificial intelligence for breast cancer detection in screening mammography in Sweden: a prospective, population-based, paired-reader, non-inferiority study. \u003cem\u003eLancet Digit Health\u003c/em\u003e \u003cstrong\u003e5\u003c/strong\u003e, e703\u0026ndash;e711 (2023).\u003c/li\u003e\n \u003cli\u003eHernstr\u0026ouml;m, V. \u003cem\u003eet al.\u003c/em\u003e Screening performance and characteristics of breast cancer detected in the Mammography Screening with Artificial Intelligence trial (MASAI): a randomised, controlled, parallel-group, non-inferiority, single-blinded, screening accuracy study. \u003cem\u003eLancet Digit Health\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, e175\u0026ndash;e183 (2025).\u003c/li\u003e\n \u003cli\u003ePerry, N. \u003cem\u003eet al.\u003c/em\u003e \u003cem\u003eEuropean Guidelines for Quality Assurance in Breast Cancer Screening and Diagnosis. Fourth Edition - Summary Document\u003c/em\u003e. \u003cem\u003eAnnals of Oncology\u003c/em\u003e vol. 19 (2008).\u003c/li\u003e\n \u003cli\u003eRadswiki, T., Campos, A. \u0026amp; Murphy, A. Interval breast cancer. in \u003cem\u003eRadiopaedia.org\u003c/em\u003e (Radiopaedia.org, 2011). doi:10.53347/rID-15319.\u003c/li\u003e\n \u003cli\u003eL\u0026aring;ng, K., Hofvind, S., Rodr\u0026iacute;guez-Ruiz, A. \u0026amp; Andersson, I. Can artificial intelligence reduce the interval cancer rate in mammography screening? doi:10.1007/s00330-021-07686-3/Published.\u003c/li\u003e\n \u003cli\u003eLauritzen, A. D. \u003cem\u003eet al.\u003c/em\u003e An artificial intelligence\u0026ndash;based mammography screening protocol for breast cancer: outcome and radiologist workload. \u003cem\u003eRadiology\u003c/em\u003e \u003cstrong\u003e304\u003c/strong\u003e, 41\u0026ndash;49 (2022).\u003c/li\u003e\n \u003cli\u003eDanish Breast Cancer Group. \u003cem\u003eSystemisk Behandling Af Brystkr\u0026aelig;ft (I) - Hvem Skal Anbefales Adjuverende Systemisk Behandling?\u003c/em\u003e https://www.dmcg.dk/siteassets/kliniske-retningslinjer---skabeloner-og-vejledninger/kliniske-retningslinjer-opdelt-pa-dmcg/dbcg/dbcg_adjuve-systemisk-bh-1_v1.3_admgodk070722.pdf (2024).\u003c/li\u003e\n \u003cli\u003eRodriguez-Ruiz, A. \u003cem\u003eet al.\u003c/em\u003e Stand-Alone Artificial Intelligence for Breast Cancer Detection in Mammography: Comparison With 101 Radiologists. \u003cem\u003eJ Natl Cancer Inst\u003c/em\u003e \u003cstrong\u003e111\u003c/strong\u003e, 916\u0026ndash;922 (2019).\u003c/li\u003e\n \u003cli\u003eRodr\u0026iacute;guez-Ruiz, A. \u003cem\u003eet al.\u003c/em\u003e Detection of breast cancer with mammography: Effect of an artificial intelligence support system. \u003cem\u003eRadiology\u003c/em\u003e \u003cstrong\u003e290\u003c/strong\u003e, 305\u0026ndash;314 (2019).\u003c/li\u003e\n \u003cli\u003eRaya-Povedano, J. L. \u003cem\u003eet al.\u003c/em\u003e AI-based Strategies to Reduce Workload in Breast Cancer Screening with Mammography and Tomosynthesis: A Retrospective Evaluation. \u003cem\u003eRadiology\u003c/em\u003e \u003cstrong\u003e300\u003c/strong\u003e, 57\u0026ndash;65 (2021).\u003c/li\u003e\n \u003cli\u003eRomero-Mart\u0026iacute;n, S. \u003cem\u003eet al.\u003c/em\u003e Stand-Alone Use of Artificial Intelligence for Digital Mammography and Digital Breast Tomosynthesis Screening: A Retrospective Evaluation. \u003cem\u003eRadiology\u003c/em\u003e \u003cstrong\u003e302\u003c/strong\u003e, 535\u0026ndash;542 (2022).\u003c/li\u003e\n \u003cli\u003ePinto, M. C. \u003cem\u003eet al.\u003c/em\u003e Impact of artificial intelligence decision support using deep learning on breast cancer screening interpretation with single-view wide-angle digital breast tomosynthesis. \u003cem\u003eRadiology\u003c/em\u003e \u003cstrong\u003e300\u003c/strong\u003e, 529\u0026ndash;536 (2021).\u003c/li\u003e\n \u003cli\u003eRodr\u0026iacute;guez-Ruiz, A. \u003cem\u003eet al.\u003c/em\u003e Detection of breast cancer with mammography: Effect of an artificial intelligence support system. \u003cem\u003eRadiology\u003c/em\u003e \u003cstrong\u003e290\u003c/strong\u003e, 305\u0026ndash;314 (2019).\u003c/li\u003e\n \u003cli\u003eRodriguez-Ruiz, A. \u003cem\u003eet al.\u003c/em\u003e Stand-Alone Artificial Intelligence for Breast Cancer Detection in Mammography: Comparison With 101 Radiologists. \u003cem\u003eJ Natl Cancer Inst\u003c/em\u003e \u003cstrong\u003e111\u003c/strong\u003e, 916\u0026ndash;922 (2019).\u003c/li\u003e\n \u003cli\u003eWanders, J. O. P. \u003cem\u003eet al.\u003c/em\u003e The combined effect of mammographic texture and density on breast cancer risk: A cohort study. \u003cem\u003eBreast Cancer Research\u003c/em\u003e \u003cstrong\u003e20\u003c/strong\u003e, 1\u0026ndash;10 (2018).\u003c/li\u003e\n \u003cli\u003eEisemann, N. \u003cem\u003eet al.\u003c/em\u003e Nationwide real-world implementation of AI for cancer detection in population-based mammography screening. \u003cem\u003eNat Med\u003c/em\u003e \u003cstrong\u003e31\u003c/strong\u003e, 917\u0026ndash;924 (2025).\u003c/li\u003e\n \u003cli\u003eDembrower, K. \u003cem\u003eet al.\u003c/em\u003e Effect of artificial intelligence-based triaging of breast cancer screening mammograms on cancer detection and radiologist workload: a retrospective simulation study. \u003cem\u003eLancet Digit Health\u003c/em\u003e \u003cstrong\u003e2\u003c/strong\u003e, e468\u0026ndash;e474 (2020).\u003c/li\u003e\n \u003cli\u003eLarsen, M., Aglen, C. F., Hoff, S. R., Lund-Hanssen, H. \u0026amp; Hofvind, S. Possible strategies for use of artificial intelligence in screen-reading of mammograms, based on retrospective data from 122,969 screening examinations. \u003cem\u003eEur Radiol\u003c/em\u003e (2022) doi:10.1007/s00330-022-08909-x.\u003c/li\u003e\n \u003cli\u003eNg, A. Y. \u003cem\u003eet al.\u003c/em\u003e Prospective implementation of AI-assisted screen reading to improve early detection of breast cancer. \u003cem\u003eNat Med\u003c/em\u003e \u003cstrong\u003e29\u003c/strong\u003e, 3044\u0026ndash;3049 (2023).\u003c/li\u003e\n \u003cli\u003ePacil\u0026egrave;, S. \u003cem\u003eet al.\u003c/em\u003e Improving Breast Cancer Detection Accuracy of Mammography with the Concurrent Use of an Artificial Intelligence Tool. \u003cem\u003eRadiol Artif Intell\u003c/em\u003e \u003cstrong\u003e2\u003c/strong\u003e, e190208 (2020).\u003c/li\u003e\n \u003cli\u003eRodriguez-Ruiz, A. \u003cem\u003eet al.\u003c/em\u003e Can we reduce the workload of mammographic screening by automatic identification of normal exams with artificial intelligence? A feasibility study. \u003cem\u003eEur Radiol\u003c/em\u003e \u003cstrong\u003e29\u003c/strong\u003e, 4825\u0026ndash;4832 (2019).\u003c/li\u003e\n \u003cli\u003eLeibig, C. \u003cem\u003eet al.\u003c/em\u003e Combining the strengths of radiologists and AI for breast cancer screening: a retrospective analysis. \u003cem\u003eLancet Digit Health\u003c/em\u003e \u003cstrong\u003e4\u003c/strong\u003e, e507\u0026ndash;e519 (2022).\u003c/li\u003e\n \u003cli\u003eVejborg, I. \u003cem\u003eMammografiscreening i Danmark - Kliniske Retningslinjer\u003c/em\u003e. rkkp.dk/siteassets/de-kliniske-kvalitetsdatabaser/databaser/mammografiscreening/kliniske-retningslinjer-for-mammografiscreening-i-danmark.-revideret-2.maj-2022_._..__.pdf (2022).\u003c/li\u003e\n \u003cli\u003eAuthor, P. \u0026amp; Fisher, R. A. \u003cem\u003eOn the Interpretation of \u0026chi; 2 from Contingency Tables, and the Calculation Of\u003c/em\u003e. \u003cem\u003eSource: Journal of the Royal Statistical Society\u003c/em\u003e vol. 85 (1922).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-7487727/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7487727/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"Artificial intelligence (AI) has been proposed to enhance mammography screening. Published literature suggested that AI increases detection and reduces workload. No published study has investigated the prospective impact of AI on sensitivity, specificity, and interval cancers. This was a population-based study on screening quality before and after implementation of AI in screening for triage and decision support. Among 270,974 screened women (156,151 with AI-support) AI-assistance improved sensitivity (73.3% with AI vs. 69.7% before AI, P=0.049), specificity (98.3% vs. 97.9%, P\u003c0.001), and detection rates (7.8 vs. 6.5 per 1,000, P\u003c0.001), while reducing recall and false-positive rates. Importantly, the 2-year interval cancer ratio decreased (26.7% vs. 30.3%, P=0.049). The reading workload was reduced by 36%. Cancers detected with AI-support presented with more favorable subtypes, including higher rates of ER-positive and luminal A tumors. These results provide evidence that AI can safely and effectively improve population-based screening by earlier detection of breast cancer.","manuscriptTitle":"Comprehensive evaluation of AI in a large regional mammography screening program for breast cancer: detection, interval cancers, and radiologist workload","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-05 08:19:44","doi":"10.21203/rs.3.rs-7487727/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"78aea6f2-40dc-427c-8c69-30518e4801ba","owner":[],"postedDate":"September 5th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":54140936,"name":"Health sciences/Diseases/Cancer/Breast cancer"},{"id":54140937,"name":"Health sciences/Health care/Medical imaging"},{"id":54140938,"name":"Health sciences/Medical research/Outcomes research"}],"tags":[],"updatedAt":"2025-09-05T08:19:47+00:00","versionOfRecord":[],"versionCreatedAt":"2025-09-05 08:19:44","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7487727","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7487727","identity":"rs-7487727","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}