The role of artificial intelligence in accurate interpretation of HER2 IHC 0 and 1+ in breast cancers

The role of artificial intelligence in accurate interpretation of HER2 IHC 0 and 1+ in breast cancers | 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 Help Center Sign In Submit a Preprint Cite Share Download PDF Article The role of artificial intelligence in accurate interpretation of HER2 IHC 0 and 1+ in breast cancers Si Wu, Meng Yue, Jun Zhang, Xiaoxian (Bill) Li, Zaibo Li, Huina Zhang, and 8 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-1967645/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 The new HER2-targeting antibody drug conjugate offers the opportunity to treat patients with HER2-low breast cancer. Distinguishing HER2 immunohistochemistry (IHC) scores of 0 and 1+, is critical but also challenging due to HER2 heterogeneity and variability of observers. In this study, we aimed to increase interpretation accuracy and consistency of HER2 IHC 0 and 1 + evaluations through assistance from artificial intelligence (AI) algorithm. In addition, we examined the value of AI algorithm in evaluating HER2 IHC scores in tumors with heterogeneity. The AI-assisted interpretation consisted of AI algorithms and an augmenting reality module with microscope. Fifteen pathologists (5 junior, 5 mid-level and 5 senior) participated this multi-institutional two-round ring study that included 246 infiltrating duct carcinoma not otherwise specified (NOS) cases. In round 1, pathologists analyzed 246 HER2 IHC slides by microscope without AI assistance. After 2 weeks of washout period, the pathologists read the same slides with AI algorithm assistance and rendered the final results by adjusting to the AI algorithm. The interpretation accuracy was significantly increased with AI assistance (Accuracy 0.93 vs 0.80), as well as the evaluation precision of HER2 0 and the recall of HER2 1+. The AI algorithm also improved the total consistency (ICC = 0.542 to 0.812), especially in HER2 1 + cases. In cases with heterogeneity, the accuracy was improved significantly (Accuracy 0.68 to 0.89) and to similar level as cases without heterogeneity (Accuracy 0.95). Both accuracy and the consistency of junior pathologists were better improved than the mid-level and senior pathologists. To the best of our knowledge, it is the first study to show that the accuracy and consistency of HER2 IHC 0 and 1 + evaluations and the accuracy of HER2 IHC evaluation in breast cancers with heterogeneity can be significantly improved using AI-assisted interpretation. Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Introduction Traditionally, only patients with human epidermal growth factor receptor 2 (HER2)- overexpressing breast cancer were eligible for HER2-targeting therapy 1 . Recent clinical trials have shown promising results of antibody drug conjugate (ADCs) in treating patients with HER2-low breast cancer. HER2-low status is currently defined as immunohistochemistry (IHC) 1 + or 2 + with in situ hybridization (ISH) non-amplification. Modi et al. showed that the novel HER2-targeted ADC trastuzumab deruxtecan (T-Dxd) significantly improved outcome in patients with HER2-low breast cancer. Patients who received T-Dxd had a median progression-free survival of 9.9 months and overall survival of 23.4 months, compared with 5.1 months and 16.8 months for those who received chemotherapies only 2 . Studies have shown that HER2 0 and HER2-low tumors have different clinicopathological features and genetic backgrounds 3 , 4 . Therefore, accurately distinguishing HER2-low from HER2 0 breast cancers becomes critical. However, accurate evaluation of low HER2 IHC staining is challenging with high intra- and inter-observer variation. A study between local and central laboratories showed that up to 85% of patients with HER2 IHC 0 breast cancer evaluated at local laboratories were scored as IHC 1 + or 2 + on central review 5 . In addition, distinguishing HER2 0 from 1 + tumors is more challenging than distinguishing between HER2 2 + and 3 + tumors 6 . Therefore, there is an urgent unmet need to develop a protocol to improve the accuracy of interpreting HER2 IHC results, especially at low levels of HER2 expression. The intra-tumoral heterogeneity of HER2 in breast cancer may contribute to the poor consistency of HER2 interpretation. The heterogeneity is higher in HER2-low than HER2 3 + tumors 7 . According to the 2013 American Society of Clinical Oncology/College of American Pathologists (ASCO/CAP) guidelines, breast cancer containing 5–50% HER2 gene amplified tumor cells should be reported as HER2 genetic heterogeneity 8 . However, there is lack of consensus regarding the definition of HER2 heterogeneity at the protein level. Intratumoral HER2 IHC heterogeneity can manifest as an uneven distribution of HER2 expression or different intensity of HER2 staining in tumor cells. Artificial intelligence (AI) plays an increasingly important role in assisting pathological diagnoses, especially in image analysis and quantitative evaluations. With AI assistance, it is possible to develop a computer algorithm to analyze images, quantify HER2 membrane staining, and provide accurate and reproducible scoring results 9 – 13 . Previous studies focused on using AI to improve evaluations of HER2-positive and HER2-negative tumors 14 , 15 . However, few studies evaluated the role of AI in differentiating between HER2 0 and HER2 1 + tumors. In this study, we aim to improve the accuracy and consistency of interpretation of HER2 0 and HER2 1 + using AI-assisted interpretation and evaluate the role of AI in assessing HER2 heterogeneity in the HER2-low breast cancer. Methods And Materials Clinical data Five hundred seventy-six consecutive cases of infiltrating duct carcinoma at the Fourth Hospital of Hebei Medical University were retrospectively reviewed and 246 cases of with HER2 0 (n = 120) or 1+ (n = 126) were included in this study. Sections of formalin-fixed and paraffin-embedded samples were stained with hematoxylin-eosin (H&E) and antibodies via immunohistochemical analysis (4b5, rabbit monoclonal antibody, Ventana Medical Systems, Oro Valley, AZ, USA). HER2 test results were interpreted according to the 2018 ASCO/CAP guidelines 16 . The sections were interpreted by two experienced pathologists (> 15 years’ experience) according to the 2018 ASCO/CAP guidelines and 2021 French recommendations 17 for the assessment of HER2 status. Cases in which the assessments of the two pathologists were inconsistent, were reassessed by a third experienced pathologist. Consensus scores were used as the gold standard. HER2 0 tumors were further divided into “HER2 0 (no staining)” and “HER2 ultra-low” tumors 18 . In this study, we defined heterogeneity as the presence of faint/barely perceptible and incomplete HER2 membrane staining in 5%-50% of tumor cells. According to the distribution and staining intensity of tumor cells, sections with heterogeneity were classified into three main types—“scattered type,” “clustered type,” and “mixed type.” The specific screening flow is shown in Fig. 1. Use of AI to interpret HER2 score The reading equipment utilized in this study was a microscope equipped with a computer unit, which included a HER2 scoring algorithm based on the 2018 ASCO/CAP guidelines. The entire system was also designed to provide AI results to pathologists in real time without interfering with their routine microscopic reading by embedding an augmented reality (AR) module under the microscope eyepiece to display the calculation results. Specifically, the HER2 scoring algorithm was built based on membrane delineation and cell classification. The cells were identified by a heatmap regression model of the fully convolutional network. To further filter out non-cancerous cells, a well-trained cancerous region segmentation model was also involved in the cell identification process. Since accurate cancerous region recognition is significantly important in the scoring stage, we have built a large-scale annotation dataset (more than 20,000 image patches with the size of 2,048 x 2,048 in 40 \(\times\) magnification). The cancerous region segmentation model was trained using a Swin-Transformer 19 . Benefiting from the large-scale annotation data and powerful deep learning model, our segmentation model could achieve robust and accurate segmentation performance. Currently, our model cannot automatically differentiate between in situ carcinomas and invasive carcinomas, and the pathologists need to select the invasive field of views or manually draw regions of interest to exclude in situ carcinoma. On the other hand, the membrane staining status was described by skeleton analysis of the segmented membrane. The RGB image was first decomposed to generate the DAB channel to enhance the DAB staining membrane. Then, the DAB image was segmented according to the empirical threshold. Note that the threshold could be used as a parameter for the pathologists to perform the calibration according to the sample slides. A skeleton algorithm was used to extract the contour to describe the cell membrane 20 . With the extracted contours, the cancerous cells could be classified into seven classes according to the completeness of membrane and staining intensity, and the suggested score could be calculated by the scoring criteria in the guidelines. The AI results for the field of view (FOV) were also displayed on the AR screen for online illustration. The detailed principle and operation process could be found in the study of Yue et al 21 . A simplified version can also be tested through the following website ( https://aihealthcare.tencent.com/research/ ). The threshold was further adjusted according to the French recommendations for the assessment of HER2 status, which is more sensitive to weak recognition. Example images of IHC sections and their AI results are shown in Fig. 2. Study design HER2 IHC was interpreted by 15 pathologists from three different hospitals. The 15 pathologists were grouped according to their practicing experience. The junior pathologists had 1–2 years’ experience, the mid-level pathologists had 3–5 years’ experience, and the senior pathologists had 6–10 years’ experience. They all had experience in interpreting HER2 IHC sections during routine clinical practice. First, pathologists reviewed the 2018 ASCO/CAP guidelines and were trained to use the AI-assisted device. The entire study included two rounds of ring study. In the first ring study (RS1), 15 pathologists interpreted 246 HER2 IHC sections via microscopic examination. After a 2-week washout period, the second ring study (RS2) was conducted, in which pathologists reinterpreted the same sections using AI assistance. Pathologists examined the whole section under low magnification lens and then selected five representative typical visual fields under a 40-magnification lens. The AI algorithm calculated the overall proportion of cells with different staining patterns for these five FOVs and displayed a comprehensive suggestion score as a reference for pathologists. Pathologists determined the final score by taking into account of the AI results. Statistical methods The interpretation results were analyzed using IBM SPSS Statistics (version 24.0; IBM Corp., Armonk, NY, USA) and GraphPad Prism 8.01 (GraphPad Software, La Jolla California, USA). The accuracy of the interpretation results was evaluated using confusion matrix and Cohen’s kappa. The intraclass correlation coefficient (ICC) was used to evaluate consistency between observers. The Kruskal-Wallis test and Wilcoxon rank-sum test were used to analyze differences in accuracy between pathologists using different interpretation methods. P values < 0.05 were considered statistically significant. Results Overall study interpretation results The results of RS1 and RS2 are shown in Fig. 3. Visually, compared to the gold standard, the accuracy and consistency of the AI-assisted assessment in RS2 were better than those of the conventional microscopic assessment in RS1. In particular, the number of cases, in which HER2 1 + lesions were misinterpreted as HER2 0 lesions, was significantly reduced in RS2. Accuracy assessment in each ring study for all pathologists Confusion Matrix was used to compare the accuracy of pathologists’ HER2 scores (Fig. 4A). The final interpretation results in RS2, called “pathologist-review”, were produced by adjusting the score according to the AI results and their perception. As shown in Fig. 4B, the accuracy was significantly improved in RS2 (Accuracy 0.93) with AI-assisted approach than in RS1 (Accuracy 0.80). The gap in accuracy between the gold standard and results of all pathologists was narrowed by 0.07 with AI-assisted approach. Next, the accuracy of the pathologists’ interpretation of HER2 0 and HER2 1 + tumors was separately evaluated. In RS1, the precision for HER2 0 tumors (Precision 0.76), the recall for HER2 1+ (Recall 0.70), the F1-score for HER2 0 (F1-score 0.82) and HER2 1+ (F1-score 0.78) were poor. In RS2, these values were improved with varying degrees. For example, the precision for HER2 0 was increased to 0.90 and the recall for HER2 1 + increased to 0.90. The F1-scores for HER2 0 and for HER2 1 + were both increased to 0.93. To investigate the validity of AI results, we compared the results of AI with pathologist-reviewed results with AI-assisted approach. The accuracy of pathologist-reviewed results and AI results was consistent. The AI had the superior results of the precision for HER2 0 (Precision 0.93) and the recall for HER2 1+ (Recall 0.93). There was no difference for F1-score of HER2 0 and HER2 1 + between AI and pathologist-reviewed results. We further investigated the causes of discordant results. As shown in Fig. 4C, most errors were caused by interpreting HER2 ultra-low tumors as HER2 1 + tumors. In addition, the AI-assisted device was able to reduce these errors. HER2 Intratumoral heterogeneity was examined and the performance of each HER2 interpretation approach was evaluated in cases with or without heterogeneity (Fig. 5A-D). All HER2 0 (no staining) cases showed homogeneity, while 28% (12/43) of HER2 0 (ultra-low) cases showed heterogeneity (Fig. 5E). Eighty-six percent of HER2 1 + cases showed heterogeneity (108/126). The accuracy of pathologist-review results in cases with homogenous staining was 0.91, but it was extremely poor with an accuracy of 0.68 in cases with heterogeneity (Fig. 5F). The AI improved the accuracy to 0.92 in heterogenous cases similar to the accuracy in homogenous cases (Accuracy 0.95). After reviewed by pathologists, the accuracy of heterogenous cases had decreased slightly (Accuracy 0.89). As shown in Fig. 5G, in HER2 ultra-low cases, the type of heterogeneity was mainly scattered type, followed by the clustered type and mixed type. In HER2 1 + cases, the three types of heterogeneity appeared in different proportions, especially mixed type. In the analysis of the accuracy assessment of different types of heterogeneous cases using the confusion matrix, the accuracy of identifying scattered type cases was relatively poor (Accuracy 0.49, Fig. 5H). The use of AI led to a significant improvement in the accuracy of detecting scattered-type heterogeneity (Accuracy 0.79). The accuracy of detection of the other two types of heterogeneity was also increased by AI. Consistency assessment in each study for all pathologists A heatmap was used to visualize changes in concordance between RS1 and RS2 (Fig. 6A-B). Pathologists obtained general consistency using a conventional microscope in RS1 (ICC = 0.542, 95% CI 0.496–0.592). There was significant improvement in concordance in RS2 in AI-assisted assessment (ICC = 0.812, 95% CI 0.783–0.840) compared to that in RS1. Concordance was similar between the AI (ICC = 0.804, 95% CI 0.744 − 0.733) and pathologist-reviewed results. We further analyzed the consistency of evaluating HER2 0 and HER2 1 + cases and used a cutoff of ≥ 90% of evaluators to obtain acceptable consistency (Fig. 6C-F). If at least 14 of 15 pathologists agreed on the interpretation of the results of the case, consensus would be achieved. In RS1, 218 cases were read as HER2 0 by at least one pathologist, and 84 cases achieved consensus (38.5%); while 185 cases were read as HER2 1 + by at least one pathologist, and 41 cases achieved consensus (22.2%). In RS2, 171 cases were read as HER2 0 by at least one pathologist, and 108 cases reached an agreement (63.2%); while 151 cases were interpreted as HER2 IHC 1 + cases by at least one pathologist and a consensus view was obtained for 89 cases (58.9%). On conventional microscopy, the interpretation consistency of HER2 IHC 1 + was worse than that of HER2 IHC 0, but it improved more significantly with the help of AI. Assessment among pathologists of different experience Fifteen pathologists were grouped according to the length of practice. Figure 7A shows a comparison of the accuracy among three groups of pathologists in RS1 and RS2. Their accuracies were similar for RS1 ( p ༞0.05). The accuracies of junior and intermediate pathologists were more volatile than that of senior pathologists. Comparing the accuracy of RS1 with that of RS2, all three groups of pathologists benefited from AI-assisted approach ( p ༜0.05). Using AI-assisted approach, the gap in accuracy between the gold standard and all pathologists with different experience was narrowed. In particular, for junior pathologists, the improvement in accuracy was the greatest from 0.57 (95% CI 0.47–0.67) in RS1 to 0.86 (95% CI 0.80–0.92) in RS2. The variation in concordance among the different groups is shown in the bar chart (Fig. 7B), which demonstrates that the consistency of all pathologists was generally higher in RS2 than in RS1. In RS1, senior pathologists had the highest concordance (ICC = 0.570, 95% CI 0.514–0.626). In RS2, intermediate pathologists had the highest concordance (ICC = 0.825, 95% CI 0.794–0.854). The concordance of junior pathologists was improved most significantly from 0.498 to 0.804. The average acceptance of all pathologists with 0.95, revealed that the pathologists had perfect agreement with AI results (Fig. 7C). Discussion With the advent that HER2-low breast cancers become targetable by the new generation of HER2-directing ADCs, more accurate and reliable methods are urgently needed to ensure accurate identification of patients with HER2-low breast cancers who may benefit from these newer ADCs. In this study, we selected 246 consecutive, real-world cases and conducted a two-round study to explore the role of AI in accurate interpretation of HER2 IHC scores of 0 and 1 + in breast cancers. Our results demonstrated that AI-assisted interpretation could significantly improve the accuracy and consistency of interpretation of HER2 0 and HER2 1 + tumors, including breast cancers with heterogeneity. It is by far the first study to provide direct evidence that AI-assistance can help pathologists better interpret HER2 IHC 0 and 1 + in breast cancers. The current definition of HER2-low breast cancers predominately relies on the HER2 IHC method to select patients who have HER2 low breast cancers and may benefit from the new HER2-targeted agent. One significant challenge using IHC as the primary testing method is the subjectivity of HER2 IHC scoring and significant intra- and inter-observer variability, especially in the low levels of HER2 expression 4 . For example, Fernandez et al. recently reported that among 18 pathologists, the concordance between HER2 0 and 1 + was 26.0%, compared with a 58.0% concordance between 2 + and 3 + 6 . AI has great advantages in the interpretation of IHC results and can be one of possible solutions to increase the accuracy and consistency in HER2 interpretation in the HER2 low-expressing tumors. Although several AI models have been proposed 22 – 24 , the AI algorithm which we proposed in this study was the first one to focus on the accurate interpretation of HER2 0 and HER2 1+. In the RS1 of this study, there were no significant differences among pathologists, and the accuracy in distinguishing HER2 0 from HER2 1 + tumors was poor, regardless of their levels of experience. In RS2, both the precision of HER2 0 and the recall of HER2 1 + increased significantly with AI-assistance, indicating that AI-assisted technology could help pathologists identify patients with HER2 0 tumors more accurately, and decrease the misinterpretation of HER2 1 + tumors. Our results support that the proposed AI algorithm in this study may be more suitable for identifying the patient population that could benefit from the new ADCs. In the current study, all pathologists with different levels of experience benefited from AI-assisted approach, and unsurprisingly, the improvement in accuracy was the greatest in the group of junior pathologists. Interestingly, the pathologists with mid-level experience achieved the best accuracy in the RS2 with AI-assistance. The possible explanations include less experience in differentiating invasive tumors from in-situ component by the junior pathologists, and more resistance of senior pathologists, who are more confident in their evaluation, to accept the AI results. It has been reported that HER2 heterogeneity was more common in HER2 1 + and HER2 2 + tumors than in HER2 3 + tumors 7 , and the HER2 heterogeneity partly contributes to the poor consistency of HER2 IHC scoring. In this study, we also investigated whether AI-assistance could improve the HER2 scoring in breast cancers with heterogeneity. The findings confirmed that AI-assistance also increased the accuracy of HER2 heterogeneous cases greatly and reached the same level as that of homogenous cases. Additionally, recognizing HER2 heterogeneity in the HER2-low breast cancers is important since the bystander killing effect of T-DXd in the HER2-low breast cancers may not be sufficient to eradicate tumors with HER2 heterogeneity, especially the tumors with clustered and scattered HER2 staining patterns 25 . Therefore, it is worth considering HER2 interpretation as a continuous variable and improving the accuracy of tumor staining percentage interpretation. For this purpose, the AI-based technology could be a promising strategy to provide more accurate and quantitative evaluation. There are some limitations in this study. First, the final comprehensive scores and the determination of HER2 heterogeneity were obtained by selecting ADCs with AI assistance, and the whole-slide evaluation could be influenced by the pathologist’s subjective choice. Second, the gold standard of HER2 scoring used in this study was based on consensus reading from two or three experienced pathologists, which is still somewhat subjective. Third, our AI model cannot automatically differentiate between in situ carcinomas and invasive carcinomas, and the pathologists need to select the invasive field of views or manually draw regions of interest to exclude in situ carcinoma, which may cause inconsistence among pathologists. In conclusion, this multi-institutional two-round ring study demonstrated that AI-assisted interpretation could significantly improve the accuracy and consistency of interpretation of HER2 0 and HER2 1 + tumors, including breast cancers with heterogeneity. The results highlight the potential of applying AI-based technology in better identifying patients with HER2-low breast cancer who are more likely to benefit from the newer ADC therapeutic agents Declarations ACKNOWLEDGEMENTS We would like to sincerely thank the pathologists who participated in this interpretation study. Except for 5 authors (Xinran Wang, Lijing Cai, Jiuyan Shang, Zhanli Jia, Jinze Li), the 10 pathologists are Lingling Zhang, Fang Li, Xu Wang, Xuemei Sun, Hanxu Jiang, Jiankun He, Shuyao Niu, Chun Wu, Mengxue Han, Xiaoyan Pei, and Yanli Wu who participated in interpreting HER2 IHC sections. ETHICS APPROVAL AND CONSENT TO PARTICIPATE The study protocol was reviewed and approved by the ethics committee of The Fourth Hospital of Hebei Medical University (approval no. 2021KY124). The study was performed in accordance with the ethics standards of the participating institutions and the Declaration of Helsinki. AUTHOR CONTRIBUTION S. W and M.Y had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. Study concept and design: S. W, M.Y, J.Z, X.W, J.S, Z.J, J.L, Y.L, X.L, Z.L. Acquisition, analysis, or interpretation of data: S. W, M.Y, J.Z, Y.L, X.W, L.C. Drafting of the manuscript: S. W, M.Y, J.Z, Y.L, X.L, Z.L. Statistical analysis: S. W, M.Y, J.Z. Study supervision: Y.L, X.L, Z.L, H.Z. Critical revision of the manuscript for important intellectual content: all authors. FUNDING This work was supported by funds from the Beijing Science And Technology Innovation Medical Development Foundation (grant number: KC2022-ZZ-0091-8). DATA AVAILABILITY All data generated or analyzed during this study are included in this published article and its supplementary information files. Subject ontology. breast cancer; HER2-low; artificial intelligence-assisted interpretation；heterogeneity References Comprehensive molecular portraits of human breast tumours. Nature 490 , 61–70 (2012). Modi S, Jacot W, Yamashita T, Sohn J, Vidal M, Tokunaga E, et al . Trastuzumab Deruxtecan in Previously Treated HER2-Low Advanced Breast Cancer. N Engl J Med 387 , 9–20 (2022). Denkert C, Seither F, Schneeweiss A, Link T, Blohmer JU, Just M, et al . Clinical and molecular characteristics of HER2-low-positive breast cancer: pooled analysis of individual patient data from four prospective, neoadjuvant clinical trials. Lancet Oncol 22 , 1151–1161 (2021). Schettini F, Chic N, Braso-Maristany F, Pare L, Pascual T, Conte B, et al. . Clinical, pathological, and PAM50 gene expression features of HER2-low breast cancer. NPJ Breast Cancer 7 , 1 (2021). Kathleen L, Mieke VB, Lies V, Sofie G, Isabelle R, Ann N, et al . Distinguishing score 0 from score 1 + in HER2 immunohistochemistry-negative breast cancer: clinical and pathobiological relevance. Am J Clin Pathol 140 , 561–566 (2013). Fernandez AI, Liu M, Bellizzi A, Brock J, Fadare O, Hanley K, et al . Examination of Low ERBB2 Protein Expression in Breast Cancer Tissue. Jama Oncol 8 , 1–4 (2022). Seol H, Lee HJ, Choi Y, Lee HE, Kim YJ, Kim JH, et al . Intratumoral heterogeneity of HER2 gene amplification in breast cancer: its clinicopathological significance. Mod Pathol 25 , 938–948 (2012). Wolff AC, Hammond ME, Hicks DG, Dowsett M, Mcshane LM, Allison KH, et al . Recommendations for human epidermal growth factor receptor 2 testing in breast cancer: American Society of Clinical Oncology/College of American Pathologists clinical practice guideline update. J Clin Oncol 31 , 3997–4013 (2013). Farahmand S, Fernandez AI, Ahmed FS, Rimm DL, Chuang JH, Reisenbichler E, et al . Deep learning trained on hematoxylin and eosin tumor region of Interest predicts HER2 status and trastuzumab treatment response in HER2 + breast cancer. Mod Pathol 35 , 44–51 (2022). Qaiser T, Mukherjee A, Reddy PC, Munugoti SD, Tallam V, Pitkaaho T, et al . HER2 challenge contest: a detailed assessment of automated HER2 scoring algorithms in whole slide images of breast cancer tissues. Histopathology 72 , 227–238 (2018). Helin HO, Tuominen VJ, Ylinen O, Helin HJ, Isola J. Free digital image analysis software helps to resolve equivocal scores in HER2 immunohistochemistry. Virchows Arch 468 , 191–198 (2016). Tuominen VJ, Tolonen TT, Isola J. ImmunoMembrane: a publicly available web application for digital image analysis of HER2 immunohistochemistry. Histopathology 60 , 758–767 (2012). Laurinaviciene A, Dasevicius D, Ostapenko V, Jarmalaite S, Lazutka J, Laurinavicius A. Membrane connectivity estimated by digital image analysis of HER2 immunohistochemistry is concordant with visual scoring and fluorescence in situ hybridization results: algorithm evaluation on breast cancer tissue microarrays. Diagn Pathol 6 , 87 (2011). Zakrzewski F, de Back W, Weigert M, Wenke T, Zeugner S, Mantey R, et al . Automated detection of the HER2 gene amplification status in Fluorescence in situ hybridization images for the diagnostics of cancer tissues. Sci Rep 9 , 8231 (2019). Koopman T, Buikema HJ, Hollema H, de Bock GH, van der Vegt B. What is the added value of digital image analysis of HER2 immunohistochemistry in breast cancer in clinical practice? A study with multiple platforms. Histopathology 74 , 917–924 (2019). Antonio C. Wolff MEHH, M. S. Bartlett MBIO, Patricia A. Spears GHVG. Human Epidermal Growth Factor Receptor 2 Testing in Breast Cancer: American Society of Clinical Oncology/ College of American Pathologists Clinical Practice Guideline Focused Update. Arch Pathol Lab Med 142 , 1364–1382 (2018). Franchet C, Djerroudi L, Maran-Gonzalez A, Abramovici O, Antoine M, Becette V, et al . Mise à jour 2021 des recommandations du GEFPICS pour l’évaluation du statut HER2 dans les cancers infiltrants du sein en France. Ann Pathol 41 , 507–520 (2021). Venetis K, Crimini E, Sajjadi E, Corti C, Guerini-Rocco E, Viale G, et al . HER2 Low, Ultra-low, and Novel Complementary Biomarkers: Expanding the Spectrum of HER2 Positivity in Breast Cancer. Front Mol Biosci 9 , 834651 (2022). Liu Z, Lin Y, Cao Y, Hu H, Wei Y, Zhang Z, et al . Swin Transformer: Hierarchical Vision Transformer using Shifted Windows.; 2021. Zhang TY, Suen CY. A Fast Parallel Algorithm for Thinning Digital Patterns. Commun Acm 27 , 236–239 (1984). Yue M, Zhang J, Wang X, Yan K, Cai L, Tian K, et al . Can AI-assisted microscope facilitate breast HER2 interpretation? A multi-institutional ring study. Virchows Arch 479 , 443–449 (2021). Tewary S, Arun I, Ahmed R, Chatterjee S, Mukhopadhyay S. AutoIHC-Analyzer: computer-assisted microscopy for automated membrane extraction/scoring in HER2 molecular markers. J Microsc 281 , 87–96 (2021). Rawat RR, Ortega I, Roy P, Sha F, Shibata D, Ruderman D, et al . Deep learned tissue "fingerprints" classify breast cancers by ER/PR/Her2 status from H&E images. Sci Rep 10 , 7275 (2020). Khameneh FD, Razavi S, Kamasak M. Automated segmentation of cell membranes to evaluate HER2 status in whole slide images using a modified deep learning network. Comput Biol Med 110 , 164–174 (2019). Tarantino P, Hamilton E, Tolaney SM, Cortes J, Morganti S, Ferraro E, et al . HER2-Low Breast Cancer: Pathological and Clinical Landscape. J Clin Oncol 38 , 1951–1962 (2020). Additional Declarations There is NO conflict of interest to disclose. 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 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-1967645","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":131492923,"identity":"70732bdb-d01f-489b-9ca2-215ea198f7e5","order_by":0,"name":"Si Wu","email":"","orcid":"","institution":"the Fourth Hospital of Hebei Medical University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Si","middleName":"","lastName":"Wu","suffix":""},{"id":131492924,"identity":"44a32281-be00-4e15-b257-b84a31850e6b","order_by":1,"name":"Meng Yue","email":"","orcid":"","institution":"the Fourth Hospital of Hebei Medical University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Meng","middleName":"","lastName":"Yue","suffix":""},{"id":131492925,"identity":"8c325c30-09e5-4333-9e29-23b21633b2be","order_by":2,"name":"Jun Zhang","email":"","orcid":"","institution":"Tencent","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Jun","middleName":"","lastName":"Zhang","suffix":""},{"id":131492926,"identity":"ce85a97c-ce2a-4958-9a09-c4f7a2edbe10","order_by":3,"name":"Xiaoxian (Bill) Li","email":"","orcid":"https://orcid.org/0000-0002-0995-1721","institution":"Emory University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Xiaoxian","middleName":"(Bill)","lastName":"Li","suffix":""},{"id":131492927,"identity":"7d1f8ce7-0989-4afc-92a6-7344f7da4f23","order_by":4,"name":"Zaibo Li","email":"","orcid":"https://orcid.org/0000-0003-1325-1696","institution":"The Ohio State University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Zaibo","middleName":"","lastName":"Li","suffix":""},{"id":131492928,"identity":"9b00fc3e-9184-4165-85b5-ffe0e9f4b97d","order_by":5,"name":"Huina Zhang","email":"","orcid":"https://orcid.org/0000-0002-4208-7680","institution":"University of Rochester","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Huina","middleName":"","lastName":"Zhang","suffix":""},{"id":131492929,"identity":"22c61dc3-83ab-443e-a19c-f26ea8844216","order_by":6,"name":"Xinran Wang","email":"","orcid":"","institution":"the Fourth Hospital of Hebei Medical University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Xinran","middleName":"","lastName":"Wang","suffix":""},{"id":131492930,"identity":"771b4177-9540-457b-9877-09cc682967f6","order_by":7,"name":"Xiao Han","email":"","orcid":"","institution":"Tencent Inc","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Xiao","middleName":"","lastName":"Han","suffix":""},{"id":131492931,"identity":"c3384de2-c409-46d5-9278-c1fe309b46d4","order_by":8,"name":"Lijing Cai","email":"","orcid":"","institution":"the Fourth Hospital of Hebei Medical University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Lijing","middleName":"","lastName":"Cai","suffix":""},{"id":131492932,"identity":"8b7bf19c-1371-4d8c-83a7-7efb1a948b98","order_by":9,"name":"Jiuyan Shang","email":"","orcid":"","institution":"the Fourth Hospital of Hebei Medical University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Jiuyan","middleName":"","lastName":"Shang","suffix":""},{"id":131492933,"identity":"ba9c32b4-b832-4ac8-8c94-8df3d092e044","order_by":10,"name":"Zhanli Jia","email":"","orcid":"","institution":"The fourth hospital of Hebei medical university","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Zhanli","middleName":"","lastName":"Jia","suffix":""},{"id":131492934,"identity":"1aee0ada-376c-438d-b918-5b7464309d42","order_by":11,"name":"Xiaoxiao 9 Wang","email":"","orcid":"","institution":"the Fourth Hospital of Hebei Medical University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Xiaoxiao","middleName":"9","lastName":"Wang","suffix":""},{"id":131492935,"identity":"fabaea66-b5de-452b-969b-a05df79cfd60","order_by":12,"name":"Jinze Li","email":"","orcid":"","institution":"the Fourth Hospital of Hebei Medical University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Jinze","middleName":"","lastName":"Li","suffix":""},{"id":131492936,"identity":"f041a4af-2a76-4654-b0f9-379f6c0dc712","order_by":13,"name":"Yueping Liu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAp0lEQVRIiWNgGAWjYPACGx5+/gbStKTJSM44QJqWwzYGDQlEquWf3Xvw042a8zwGDAcYP3zMIUKLxJ1zydI5x27zmDM3MEvO3EaMNTdyDKRz2G7zWDYcYGPmJUaL/I0c4985/87xGBxIIFKLwY0cM+nctgMkaDEEarHO7UvmkZxxsJk4v8gBHXY755udPT9/88EPH4nyPgIwNpCmfhSMglEwCkYBbgAANBs08sehZ/0AAAAASUVORK5CYII=","orcid":"https://orcid.org/0000-0002-4582-114X","institution":"The Fourth Hospital of Hebei Medical University","correspondingAuthor":true,"submittingAuthor":false,"prefix":"","firstName":"Yueping","middleName":"","lastName":"Liu","suffix":""}],"badges":[],"createdAt":"2022-08-16 13:25:55","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-1967645/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-1967645/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":25793954,"identity":"c42730d8-02b5-47c3-86a6-1a4651655fa2","added_by":"auto","created_at":"2022-08-29 14:00:44","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":325410,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCase screening flow-chart. \u003c/strong\u003eFlow-chart showed the selection process of patients, and the causes for exclusion. Note: According to 2018 ASCO/CAP guidelines, some patterns, such as basolateral or lateral staining, were not applicable to general evaluation protocol. To prevent the special staining from confusing the AI results, we chose cases of infiltrating duct carcinoma not otherwise specified (NOS).\u003c/p\u003e\u003cp\u003e\u003cem\u003eIDC \u003c/em\u003einvasive ductal carcinoma \u003cem\u003eNAC\u003c/em\u003e neoadjuvant chemotherapy\u003c/p\u003e","description":"","filename":"Fig.1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-1967645/v1/c8f484ff8a6e7abf601024ae.jpg"},{"id":25793953,"identity":"0fe0f036-7878-42f1-bbf6-5354d618674e","added_by":"auto","created_at":"2022-08-29 14:00:44","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":2419930,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eHER2 images under conventional microscope and AI-assisted interpretation.\u003c/strong\u003e \u003c/p\u003e","description":"","filename":"Fig.2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-1967645/v1/b177487704d4c7d6b0934340.jpg"},{"id":25794712,"identity":"12f63f0f-1ff1-4c39-8b20-1fde73a03101","added_by":"auto","created_at":"2022-08-29 14:10:44","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":556511,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eHER2 scoring of 246 cases in two rounds of ring studies.\u003c/strong\u003e The gold standard results were displayed at the top. RS1 and RS2 were the results of the first and second rounds of interpretation, respectively. The white color represented HER2 0 and the black color represented HER2 1+.\u003c/p\u003e","description":"","filename":"Fig3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-1967645/v1/36d1020dad3dbdec2a6b682d.jpg"},{"id":25793956,"identity":"a0b3b90e-89e4-4d7e-81b2-13bc2a4a6338","added_by":"auto","created_at":"2022-08-29 14:00:44","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1570001,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eHER2 scoring of 246 cases in two rounds of ring studies and accuracy of two ring studies for pathologists. A\u003c/strong\u003e Confusion Matrix in two ring studies.\u003cstrong\u003e B \u003c/strong\u003eThe evaluation of total accuracy and the statistical comparisons of intergroup significance\u003cstrong\u003e. \u003c/strong\u003ePaired t-test were performed to identify the significance between groups, where ‘ns’ indicated p≥0.05, ‘*’ indicated 0.01≤p\u0026lt;0.05, ‘**’ indicated 0.001≤p\u0026lt;0.01, and ‘***’ indicated p\u0026lt;0.001.\u003cstrong\u003e C \u003c/strong\u003eRefined HER2 0 of the gold standard into HER2 0 (no staining) and HER2 ultra-low. The light color represented HER2 0 which was interpreted by 15 pathologists, and the dark color represented HER2 1+ and error cases. \u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Fig4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-1967645/v1/facc73cb1ab3f90fb83e973a.jpg"},{"id":25794519,"identity":"12fb8f6e-fb79-46d2-a2d4-e4988911d062","added_by":"auto","created_at":"2022-08-29 14:05:44","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":2563827,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThe heterogeneity analysis of HER2 sections in two ring studies. A \u003c/strong\u003eThe HER2 section of homogenous cases with no HER2 membrane staining under 20×lens. \u003cstrong\u003eB \u003c/strong\u003eThe HER2 section of homogenous cases with \u0026gt;50% HER2 membrane staining under 20×lens.\u003cstrong\u003eC \u003c/strong\u003eThe HER2 section of clustered-type heterogenous cases under 20×lens. \u003cstrong\u003eD\u003c/strong\u003e The HER2 section of mixed-type heterogenous cases under 20×lens. \u003cstrong\u003eE \u003c/strong\u003eThe existence of heterogeneity in HER2 0 (no staining), HER2 ultra-low, and HER2 1+. \u003cstrong\u003eF\u003c/strong\u003e The accuracy of homogenous and heterogenous cases` interpretation in two ring studies. \u003cstrong\u003eG\u003c/strong\u003e The distribution of three heterogenous types cases in HER2 ultra-low and HER2 1+ cases. \u003cstrong\u003eH \u003c/strong\u003eThe accuracy of three types of heterogenous cases` interpretation in two ring studies.\u003c/p\u003e","description":"","filename":"Fig5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-1967645/v1/b50e1b366669e52ae2ac42de.jpg"},{"id":25795418,"identity":"ce040d5d-a158-49e0-9a84-081de2d1e203","added_by":"auto","created_at":"2022-08-29 14:15:44","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":1708403,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThe concordance of HER2 scoring in two ring studies. A \u003c/strong\u003eThe total concordance in RS1 \u003cstrong\u003eB\u003c/strong\u003e The total concordance in RS2\u003cstrong\u003e C \u003c/strong\u003eThe concordance of HER2 0 in RS1 \u003cstrong\u003eD\u003c/strong\u003e The concordance of HER2 1+ in RS1 \u003cstrong\u003eE \u003c/strong\u003eThe concordance of HER2 0 in RS2 \u003cstrong\u003eF\u003c/strong\u003e The concordance of HER2 1+ in RS2.\u003c/p\u003e","description":"","filename":"Fig6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-1967645/v1/4ca92a1c8ae5f872978d8a01.jpg"},{"id":25794710,"identity":"438e9274-514c-4800-91ae-896e1cbaa11a","added_by":"auto","created_at":"2022-08-29 14:10:44","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":1251302,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAccuracy, acceptance rate, and concordance of pathologists in different levels. A \u003c/strong\u003eComparison of accuracies of junior, mid-level, and senior pathologists in two ring studies. \u003cstrong\u003eB\u003c/strong\u003e Intra-observer concordance of pathologists at different levels between RS1 (yellow) and RS2 (green). \u003cstrong\u003eC \u003c/strong\u003eAcceptance rate of AI results by pathologists in different levels.\u003cstrong\u003e \u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Fig7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-1967645/v1/9fd819b0dfd9099917301cb9.jpg"},{"id":28363659,"identity":"abd09c1c-0fd3-48cc-b165-cbab0838c528","added_by":"auto","created_at":"2022-10-28 11:32:55","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1413986,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-1967645/v1/1422767d-c0c4-4e73-b943-8ccbc8884890.pdf"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e conflict of interest to disclose.","formattedTitle":"The role of artificial intelligence in accurate interpretation of HER2 IHC 0 and 1+ in breast cancers","fulltext":[{"header":"Introduction","content":"\u003cp\u003eTraditionally, only patients with human epidermal growth factor receptor 2 (HER2)- overexpressing breast cancer were eligible for HER2-targeting therapy\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. Recent clinical trials have shown promising results of antibody drug conjugate (ADCs) in treating patients with HER2-low breast cancer. HER2-low status is currently defined as immunohistochemistry (IHC) 1\u0026thinsp;+\u0026thinsp;or 2\u0026thinsp;+\u0026thinsp;with in situ hybridization (ISH) non-amplification. Modi et al. showed that the novel HER2-targeted ADC trastuzumab deruxtecan (T-Dxd) significantly improved outcome in patients with HER2-low breast cancer. Patients who received T-Dxd had a median progression-free survival of 9.9 months and overall survival of 23.4 months, compared with 5.1 months and 16.8 months for those who received chemotherapies only\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Studies have shown that HER2 0 and HER2-low tumors have different clinicopathological features and genetic backgrounds\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. Therefore, accurately distinguishing HER2-low from HER2 0 breast cancers becomes critical. However, accurate evaluation of low HER2 IHC staining is challenging with high intra- and inter-observer variation. A study between local and central laboratories showed that up to 85% of patients with HER2 IHC 0 breast cancer evaluated at local laboratories were scored as IHC 1\u0026thinsp;+\u0026thinsp;or 2\u0026thinsp;+\u0026thinsp;on central review\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. In addition, distinguishing HER2 0 from 1\u0026thinsp;+\u0026thinsp;tumors is more challenging than distinguishing between HER2 2\u0026thinsp;+\u0026thinsp;and 3\u0026thinsp;+\u0026thinsp;tumors\u003csup\u003e6\u003c/sup\u003e. Therefore, there is an urgent unmet need to develop a protocol to improve the accuracy of interpreting HER2 IHC results, especially at low levels of HER2 expression.\u003c/p\u003e \u003cp\u003eThe intra-tumoral heterogeneity of HER2 in breast cancer may contribute to the poor consistency of HER2 interpretation. The heterogeneity is higher in HER2-low than HER2 3\u0026thinsp;+\u0026thinsp;tumors\u003csup\u003e7\u003c/sup\u003e. According to the 2013 American Society of Clinical Oncology/College of American Pathologists (ASCO/CAP) guidelines, breast cancer containing 5\u0026ndash;50% HER2 gene amplified tumor cells should be reported as HER2 genetic heterogeneity\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. However, there is lack of consensus regarding the definition of HER2 heterogeneity at the protein level. Intratumoral HER2 IHC heterogeneity can manifest as an uneven distribution of HER2 expression or different intensity of HER2 staining in tumor cells.\u003c/p\u003e \u003cp\u003eArtificial intelligence (AI) plays an increasingly important role in assisting pathological diagnoses, especially in image analysis and quantitative evaluations. With AI assistance, it is possible to develop a computer algorithm to analyze images, quantify HER2 membrane staining, and provide accurate and reproducible scoring results\u003csup\u003e\u003cspan additionalcitationids=\"CR10 CR11 CR12\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. Previous studies focused on using AI to improve evaluations of HER2-positive and HER2-negative tumors\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e,\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. However, few studies evaluated the role of AI in differentiating between HER2 0 and HER2 1\u0026thinsp;+\u0026thinsp;tumors.\u003c/p\u003e \u003cp\u003eIn this study, we aim to improve the accuracy and consistency of interpretation of HER2 0 and HER2 1\u0026thinsp;+\u0026thinsp;using AI-assisted interpretation and evaluate the role of AI in assessing HER2 heterogeneity in the HER2-low breast cancer.\u003c/p\u003e"},{"header":"Methods And Materials","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eClinical data\u003c/h2\u003e \u003cp\u003eFive hundred seventy-six consecutive cases of infiltrating duct carcinoma at the Fourth Hospital of Hebei Medical University were retrospectively reviewed and 246 cases of with HER2 0 (n\u0026thinsp;=\u0026thinsp;120) or 1+ (n\u0026thinsp;=\u0026thinsp;126) were included in this study.\u003c/p\u003e \u003cp\u003eSections of formalin-fixed and paraffin-embedded samples were stained with hematoxylin-eosin (H\u0026amp;E) and antibodies via immunohistochemical analysis (4b5, rabbit monoclonal antibody, Ventana Medical Systems, Oro Valley, AZ, USA). HER2 test results were interpreted according to the 2018 ASCO/CAP guidelines\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e. The sections were interpreted by two experienced pathologists (\u0026gt;\u0026thinsp;15 years\u0026rsquo; experience) according to the 2018 ASCO/CAP guidelines and 2021 French recommendations\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e for the assessment of HER2 status. Cases in which the assessments of the two pathologists were inconsistent, were reassessed by a third experienced pathologist. Consensus scores were used as the gold standard. HER2 0 tumors were further divided into \u0026ldquo;HER2 0 (no staining)\u0026rdquo; and \u0026ldquo;HER2 ultra-low\u0026rdquo; tumors\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eIn this study, we defined heterogeneity as the presence of faint/barely perceptible and incomplete HER2 membrane staining in 5%-50% of tumor cells. According to the distribution and staining intensity of tumor cells, sections with heterogeneity were classified into three main types\u0026mdash;\u0026ldquo;scattered type,\u0026rdquo; \u0026ldquo;clustered type,\u0026rdquo; and \u0026ldquo;mixed type.\u0026rdquo; The specific screening flow is shown in Fig.\u0026nbsp;1.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003eUse of AI to interpret HER2 score\u003c/h2\u003e \u003cp\u003e The reading equipment utilized in this study was a microscope equipped with a computer unit, which included a HER2 scoring algorithm based on the 2018 ASCO/CAP guidelines. The entire system was also designed to provide AI results to pathologists in real time without interfering with their routine microscopic reading by embedding an augmented reality (AR) module under the microscope eyepiece to display the calculation results.\u003c/p\u003e \u003cp\u003eSpecifically, the HER2 scoring algorithm was built based on membrane delineation and cell classification. The cells were identified by a heatmap regression model of the fully convolutional network. To further filter out non-cancerous cells, a well-trained cancerous region segmentation model was also involved in the cell identification process. Since accurate cancerous region recognition is significantly important in the scoring stage, we have built a large-scale annotation dataset (more than 20,000 image patches with the size of 2,048 x 2,048 in 40\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\times\\)\u003c/span\u003e\u003c/span\u003e magnification). The cancerous region segmentation model was trained using a Swin-Transformer\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. Benefiting from the large-scale annotation data and powerful deep learning model, our segmentation model could achieve robust and accurate segmentation performance. Currently, our model cannot automatically differentiate between in situ carcinomas and invasive carcinomas, and the pathologists need to select the invasive field of views or manually draw regions of interest to exclude in situ carcinoma.\u003c/p\u003e \u003cp\u003eOn the other hand, the membrane staining status was described by skeleton analysis of the segmented membrane. The RGB image was first decomposed to generate the DAB channel to enhance the DAB staining membrane. Then, the DAB image was segmented according to the empirical threshold. Note that the threshold could be used as a parameter for the pathologists to perform the calibration according to the sample slides. A skeleton algorithm was used to extract the contour to describe the cell membrane\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. With the extracted contours, the cancerous cells could be classified into seven classes according to the completeness of membrane and staining intensity, and the suggested score could be calculated by the scoring criteria in the guidelines. The AI results for the field of view (FOV) were also displayed on the AR screen for online illustration. The detailed principle and operation process could be found in the study of Yue et al\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. A simplified version can also be tested through the following website (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://aihealthcare.tencent.com/research/\u003c/span\u003e\u003cspan address=\"https://aihealthcare.tencent.com/research/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). The threshold was further adjusted according to the French recommendations for the assessment of HER2 status, which is more sensitive to weak recognition. Example images of IHC sections and their AI results are shown in Fig.\u0026nbsp;2.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003eStudy design\u003c/h2\u003e \u003cp\u003eHER2 IHC was interpreted by 15 pathologists from three different hospitals. The 15 pathologists were grouped according to their practicing experience. The junior pathologists had 1\u0026ndash;2 years\u0026rsquo; experience, the mid-level pathologists had 3\u0026ndash;5 years\u0026rsquo; experience, and the senior pathologists had 6\u0026ndash;10 years\u0026rsquo; experience. They all had experience in interpreting HER2 IHC sections during routine clinical practice. First, pathologists reviewed the 2018 ASCO/CAP guidelines and were trained to use the AI-assisted device. The entire study included two rounds of ring study. In the first ring study (RS1), 15 pathologists interpreted 246 HER2 IHC sections via microscopic examination. After a 2-week washout period, the second ring study (RS2) was conducted, in which pathologists reinterpreted the same sections using AI assistance. Pathologists examined the whole section under low magnification lens and then selected five representative typical visual fields under a 40-magnification lens. The AI algorithm calculated the overall proportion of cells with different staining patterns for these five FOVs and displayed a comprehensive suggestion score as a reference for pathologists. Pathologists determined the final score by taking into account of the AI results.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003eStatistical methods\u003c/h2\u003e \u003cp\u003eThe interpretation results were analyzed using IBM SPSS Statistics (version 24.0; IBM Corp., Armonk, NY, USA) and GraphPad Prism 8.01 (GraphPad Software, La Jolla California, USA). The accuracy of the interpretation results was evaluated using confusion matrix and Cohen\u0026rsquo;s kappa. The intraclass correlation coefficient (ICC) was used to evaluate consistency between observers. The Kruskal-Wallis test and Wilcoxon rank-sum test were used to analyze differences in accuracy between pathologists using different interpretation methods. \u003cem\u003eP\u003c/em\u003e values\u0026thinsp;\u0026lt;\u0026thinsp;0.05 were considered statistically significant.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eOverall study interpretation results\u003c/h2\u003e \u003cp\u003eThe results of RS1 and RS2 are shown in Fig.\u0026nbsp;3. Visually, compared to the gold standard, the accuracy and consistency of the AI-assisted assessment in RS2 were better than those of the conventional microscopic assessment in RS1. In particular, the number of cases, in which HER2 1\u0026thinsp;+\u0026thinsp;lesions were misinterpreted as HER2 0 lesions, was significantly reduced in RS2.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003eAccuracy assessment in each ring study for all pathologists\u003c/h2\u003e \u003cp\u003eConfusion Matrix was used to compare the accuracy of pathologists\u0026rsquo; HER2 scores (Fig.\u0026nbsp;4A). The final interpretation results in RS2, called \u0026ldquo;pathologist-review\u0026rdquo;, were produced by adjusting the score according to the AI results and their perception. As shown in Fig.\u0026nbsp;4B, the accuracy was significantly improved in RS2 (Accuracy 0.93) with AI-assisted approach than in RS1 (Accuracy 0.80). The gap in accuracy between the gold standard and results of all pathologists was narrowed by 0.07 with AI-assisted approach.\u003c/p\u003e \u003cp\u003eNext, the accuracy of the pathologists\u0026rsquo; interpretation of HER2 0 and HER2 1\u0026thinsp;+\u0026thinsp;tumors was separately evaluated. In RS1, the precision for HER2 0 tumors (Precision 0.76), the recall for HER2 1+ (Recall 0.70), the F1-score for HER2 0 (F1-score 0.82) and HER2 1+ (F1-score 0.78) were poor. In RS2, these values were improved with varying degrees. For example, the precision for HER2 0 was increased to 0.90 and the recall for HER2 1\u0026thinsp;+\u0026thinsp;increased to 0.90. The F1-scores for HER2 0 and for HER2 1\u0026thinsp;+\u0026thinsp;were both increased to 0.93.\u003c/p\u003e \u003cp\u003eTo investigate the validity of AI results, we compared the results of AI with pathologist-reviewed results with AI-assisted approach. The accuracy of pathologist-reviewed results and AI results was consistent. The AI had the superior results of the precision for HER2 0 (Precision 0.93) and the recall for HER2 1+ (Recall 0.93). There was no difference for F1-score of HER2 0 and HER2 1\u0026thinsp;+\u0026thinsp;between AI and pathologist-reviewed results.\u003c/p\u003e \u003cp\u003eWe further investigated the causes of discordant results. As shown in Fig.\u0026nbsp;4C, most errors were caused by interpreting HER2 ultra-low tumors as HER2 1\u0026thinsp;+\u0026thinsp;tumors. In addition, the AI-assisted device was able to reduce these errors.\u003c/p\u003e \u003cp\u003eHER2 Intratumoral heterogeneity was examined and the performance of each HER2 interpretation approach was evaluated in cases with or without heterogeneity (Fig.\u0026nbsp;5A-D). All HER2 0 (no staining) cases showed homogeneity, while 28% (12/43) of HER2 0 (ultra-low) cases showed heterogeneity (Fig.\u0026nbsp;5E). Eighty-six percent of HER2 1\u0026thinsp;+\u0026thinsp;cases showed heterogeneity (108/126).\u003c/p\u003e \u003cp\u003eThe accuracy of pathologist-review results in cases with homogenous staining was 0.91, but it was extremely poor with an accuracy of 0.68 in cases with heterogeneity (Fig.\u0026nbsp;5F). The AI improved the accuracy to 0.92 in heterogenous cases similar to the accuracy in homogenous cases (Accuracy 0.95). After reviewed by pathologists, the accuracy of heterogenous cases had decreased slightly (Accuracy 0.89). As shown in Fig.\u0026nbsp;5G, in HER2 ultra-low cases, the type of heterogeneity was mainly scattered type, followed by the clustered type and mixed type. In HER2 1\u0026thinsp;+\u0026thinsp;cases, the three types of heterogeneity appeared in different proportions, especially mixed type. In the analysis of the accuracy assessment of different types of heterogeneous cases using the confusion matrix, the accuracy of identifying scattered type cases was relatively poor (Accuracy 0.49, Fig.\u0026nbsp;5H). The use of AI led to a significant improvement in the accuracy of detecting scattered-type heterogeneity (Accuracy 0.79). The accuracy of detection of the other two types of heterogeneity was also increased by AI.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003eConsistency assessment in each study for all pathologists\u003c/h2\u003e \u003cp\u003eA heatmap was used to visualize changes in concordance between RS1 and RS2 (Fig.\u0026nbsp;6A-B). Pathologists obtained general consistency using a conventional microscope in RS1 (ICC\u0026thinsp;=\u0026thinsp;0.542, 95% CI 0.496\u0026ndash;0.592). There was significant improvement in concordance in RS2 in AI-assisted assessment (ICC\u0026thinsp;=\u0026thinsp;0.812, 95% CI 0.783\u0026ndash;0.840) compared to that in RS1. Concordance was similar between the AI (ICC\u0026thinsp;=\u0026thinsp;0.804, 95% CI 0.744\u0026thinsp;\u0026minus;\u0026thinsp;0.733) and pathologist-reviewed results.\u003c/p\u003e \u003cp\u003eWe further analyzed the consistency of evaluating HER2 0 and HER2 1\u0026thinsp;+\u0026thinsp;cases and used a cutoff of \u0026ge;\u0026thinsp;90% of evaluators to obtain acceptable consistency (Fig.\u0026nbsp;6C-F). If at least 14 of 15 pathologists agreed on the interpretation of the results of the case, consensus would be achieved. In RS1, 218 cases were read as HER2 0 by at least one pathologist, and 84 cases achieved consensus (38.5%); while 185 cases were read as HER2 1\u0026thinsp;+\u0026thinsp;by at least one pathologist, and 41 cases achieved consensus (22.2%).\u003c/p\u003e \u003cp\u003eIn RS2, 171 cases were read as HER2 0 by at least one pathologist, and 108 cases reached an agreement (63.2%); while 151 cases were interpreted as HER2 IHC 1\u0026thinsp;+\u0026thinsp;cases by at least one pathologist and a consensus view was obtained for 89 cases (58.9%).\u003c/p\u003e \u003cp\u003eOn conventional microscopy, the interpretation consistency of HER2 IHC 1\u0026thinsp;+\u0026thinsp;was worse than that of HER2 IHC 0, but it improved more significantly with the help of AI.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eAssessment among pathologists of different experience\u003c/h2\u003e \u003cp\u003eFifteen pathologists were grouped according to the length of practice. Figure\u0026nbsp;7A shows a comparison of the accuracy among three groups of pathologists in RS1 and RS2. Their accuracies were similar for RS1 (\u003cem\u003ep\u003c/em\u003e༞0.05). The accuracies of junior and intermediate pathologists were more volatile than that of senior pathologists. Comparing the accuracy of RS1 with that of RS2, all three groups of pathologists benefited from AI-assisted approach (\u003cem\u003ep\u003c/em\u003e༜0.05). Using AI-assisted approach, the gap in accuracy between the gold standard and all pathologists with different experience was narrowed. In particular, for junior pathologists, the improvement in accuracy was the greatest from 0.57 (95% CI 0.47\u0026ndash;0.67) in RS1 to 0.86 (95% CI 0.80\u0026ndash;0.92) in RS2.\u003c/p\u003e \u003cp\u003eThe variation in concordance among the different groups is shown in the bar chart (Fig.\u0026nbsp;7B), which demonstrates that the consistency of all pathologists was generally higher in RS2 than in RS1. In RS1, senior pathologists had the highest concordance (ICC\u0026thinsp;=\u0026thinsp;0.570, 95% CI 0.514\u0026ndash;0.626). In RS2, intermediate pathologists had the highest concordance (ICC\u0026thinsp;=\u0026thinsp;0.825, 95% CI 0.794\u0026ndash;0.854). The concordance of junior pathologists was improved most significantly from 0.498 to 0.804.\u003c/p\u003e \u003cp\u003eThe average acceptance of all pathologists with 0.95, revealed that the pathologists had perfect agreement with AI results (Fig.\u0026nbsp;7C).\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eWith the advent that HER2-low breast cancers become targetable by the new generation of HER2-directing ADCs, more accurate and reliable methods are urgently needed to ensure accurate identification of patients with HER2-low breast cancers who may benefit from these newer ADCs. In this study, we selected 246 consecutive, real-world cases and conducted a two-round study to explore the role of AI in accurate interpretation of HER2 IHC scores of 0 and 1\u0026thinsp;+\u0026thinsp;in breast cancers. Our results demonstrated that AI-assisted interpretation could significantly improve the accuracy and consistency of interpretation of HER2 0 and HER2 1\u0026thinsp;+\u0026thinsp;tumors, including breast cancers with heterogeneity. It is by far the first study to provide direct evidence that AI-assistance can help pathologists better interpret HER2 IHC 0 and 1\u0026thinsp;+\u0026thinsp;in breast cancers.\u003c/p\u003e \u003cp\u003eThe current definition of HER2-low breast cancers predominately relies on the HER2 IHC method to select patients who have HER2 low breast cancers and may benefit from the new HER2-targeted agent. One significant challenge using IHC as the primary testing method is the subjectivity of HER2 IHC scoring and significant intra- and inter-observer variability, especially in the low levels of HER2 expression\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. For example, Fernandez et al. recently reported that among 18 pathologists, the concordance between HER2 0 and 1\u0026thinsp;+\u0026thinsp;was 26.0%, compared with a 58.0% concordance between 2\u0026thinsp;+\u0026thinsp;and 3\u0026thinsp;+\u0026thinsp;\u003csup\u003e6\u003c/sup\u003e. AI has great advantages in the interpretation of IHC results and can be one of possible solutions to increase the accuracy and consistency in HER2 interpretation in the HER2 low-expressing tumors. Although several AI models have been proposed\u003csup\u003e\u003cspan additionalcitationids=\"CR23\" citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e, the AI algorithm which we proposed in this study was the first one to focus on the accurate interpretation of HER2 0 and HER2 1+. In the RS1 of this study, there were no significant differences among pathologists, and the accuracy in distinguishing HER2 0 from HER2 1\u0026thinsp;+\u0026thinsp;tumors was poor, regardless of their levels of experience. In RS2, both the precision of HER2 0 and the recall of HER2 1\u0026thinsp;+\u0026thinsp;increased significantly with AI-assistance, indicating that AI-assisted technology could help pathologists identify patients with HER2 0 tumors more accurately, and decrease the misinterpretation of HER2 1\u0026thinsp;+\u0026thinsp;tumors. Our results support that the proposed AI algorithm in this study may be more suitable for identifying the patient population that could benefit from the new ADCs.\u003c/p\u003e \u003cp\u003eIn the current study, all pathologists with different levels of experience benefited from AI-assisted approach, and unsurprisingly, the improvement in accuracy was the greatest in the group of junior pathologists. Interestingly, the pathologists with mid-level experience achieved the best accuracy in the RS2 with AI-assistance. The possible explanations include less experience in differentiating invasive tumors from in-situ component by the junior pathologists, and more resistance of senior pathologists, who are more confident in their evaluation, to accept the AI results.\u003c/p\u003e \u003cp\u003eIt has been reported that HER2 heterogeneity was more common in HER2 1\u0026thinsp;+\u0026thinsp;and HER2 2\u0026thinsp;+\u0026thinsp;tumors than in HER2 3\u0026thinsp;+\u0026thinsp;tumors\u003csup\u003e7\u003c/sup\u003e, and the HER2 heterogeneity partly contributes to the poor consistency of HER2 IHC scoring. In this study, we also investigated whether AI-assistance could improve the HER2 scoring in breast cancers with heterogeneity. The findings confirmed that AI-assistance also increased the accuracy of HER2 heterogeneous cases greatly and reached the same level as that of homogenous cases. Additionally, recognizing HER2 heterogeneity in the HER2-low breast cancers is important since the bystander killing effect of T-DXd in the HER2-low breast cancers may not be sufficient to eradicate tumors with HER2 heterogeneity, especially the tumors with clustered and scattered HER2 staining patterns\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e. Therefore, it is worth considering HER2 interpretation as a continuous variable and improving the accuracy of tumor staining percentage interpretation. For this purpose, the AI-based technology could be a promising strategy to provide more accurate and quantitative evaluation.\u003c/p\u003e \u003cp\u003eThere are some limitations in this study. First, the final comprehensive scores and the determination of HER2 heterogeneity were obtained by selecting ADCs with AI assistance, and the whole-slide evaluation could be influenced by the pathologist\u0026rsquo;s subjective choice. Second, the gold standard of HER2 scoring used in this study was based on consensus reading from two or three experienced pathologists, which is still somewhat subjective. Third, our AI model cannot automatically differentiate between in situ carcinomas and invasive carcinomas, and the pathologists need to select the invasive field of views or manually draw regions of interest to exclude in situ carcinoma, which may cause inconsistence among pathologists.\u003c/p\u003e \u003cp\u003eIn conclusion, this multi-institutional two-round ring study demonstrated that AI-assisted interpretation could significantly improve the accuracy and consistency of interpretation of HER2 0 and HER2 1\u0026thinsp;+\u0026thinsp;tumors, including breast cancers with heterogeneity. The results highlight the potential of applying AI-based technology in better identifying patients with HER2-low breast cancer who are more likely to benefit from the newer ADC therapeutic agents\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eACKNOWLEDGEMENTS\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe would like to sincerely thank the pathologists who participated in this interpretation study. Except for 5 authors (Xinran Wang, Lijing Cai, Jiuyan Shang, Zhanli Jia, Jinze Li), the 10 pathologists are Lingling Zhang, Fang Li, Xu Wang, Xuemei Sun, Hanxu Jiang, Jiankun He, Shuyao Niu, Chun Wu, Mengxue Han, Xiaoyan Pei, and Yanli Wu who participated in interpreting HER2 IHC sections.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eETHICS APPROVAL AND CONSENT TO PARTICIPATE\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study protocol was reviewed and approved by the ethics committee of\u0026nbsp;The Fourth Hospital of Hebei Medical University\u0026nbsp;(approval no. 2021KY124). The study was performed in accordance with the ethics standards of the participating institutions and the Declaration of Helsinki.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAUTHOR CONTRIBUTION\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eS. W and M.Y had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. Study concept and design: S. W, M.Y, J.Z, X.W, J.S, Z.J, J.L, Y.L, X.L, Z.L. Acquisition, analysis, or interpretation of data: S. W, M.Y, J.Z, Y.L, X.W, L.C. Drafting of the manuscript: S. W, M.Y, J.Z, Y.L, X.L, Z.L. Statistical analysis: S. W, M.Y, J.Z. Study supervision: Y.L, X.L, Z.L, H.Z. Critical revision of the manuscript for important intellectual content: all authors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFUNDING\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by funds from the Beijing Science And Technology Innovation Medical Development Foundation (grant number: KC2022-ZZ-0091-8).\u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDATA AVAILABILITY\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAll data generated or analyzed during this study are included in this published article and its supplementary information files.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSubject ontology.\u003c/strong\u003e breast cancer; HER2-low; artificial intelligence-assisted interpretation；heterogeneity\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eComprehensive molecular portraits of human breast tumours. Nature \u003cb\u003e490\u003c/b\u003e, 61\u0026ndash;70 (2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eModi S, Jacot W, Yamashita T, Sohn J, Vidal M, Tokunaga E, \u003cem\u003eet al\u003c/em\u003e. Trastuzumab Deruxtecan in Previously Treated HER2-Low Advanced Breast Cancer. N Engl J Med \u003cb\u003e387\u003c/b\u003e, 9\u0026ndash;20 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDenkert C, Seither F, Schneeweiss A, Link T, Blohmer JU, Just M, \u003cem\u003eet al\u003c/em\u003e. Clinical and molecular characteristics of HER2-low-positive breast cancer: pooled analysis of individual patient data from four prospective, neoadjuvant clinical trials. Lancet Oncol \u003cb\u003e22\u003c/b\u003e, 1151\u0026ndash;1161 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchettini F, Chic N, Braso-Maristany F, Pare L, Pascual T, Conte B, \u003cem\u003eet al.\u003c/em\u003e. Clinical, pathological, and PAM50 gene expression features of HER2-low breast cancer. NPJ Breast Cancer \u003cb\u003e7\u003c/b\u003e, 1 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKathleen L, Mieke VB, Lies V, Sofie G, Isabelle R, Ann N, \u003cem\u003eet al\u003c/em\u003e. Distinguishing score 0 from score 1 + in HER2 immunohistochemistry-negative breast cancer: clinical and pathobiological relevance. Am J Clin Pathol \u003cb\u003e140\u003c/b\u003e, 561\u0026ndash;566 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFernandez AI, Liu M, Bellizzi A, Brock J, Fadare O, Hanley K, \u003cem\u003eet al\u003c/em\u003e. Examination of Low ERBB2 Protein Expression in Breast Cancer Tissue. Jama Oncol \u003cb\u003e8\u003c/b\u003e, 1\u0026ndash;4 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSeol H, Lee HJ, Choi Y, Lee HE, Kim YJ, Kim JH, \u003cem\u003eet al\u003c/em\u003e. Intratumoral heterogeneity of HER2 gene amplification in breast cancer: its clinicopathological significance. Mod Pathol \u003cb\u003e25\u003c/b\u003e, 938\u0026ndash;948 (2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWolff AC, Hammond ME, Hicks DG, Dowsett M, Mcshane LM, Allison KH, \u003cem\u003eet al\u003c/em\u003e. Recommendations for human epidermal growth factor receptor 2 testing in breast cancer: American Society of Clinical Oncology/College of American Pathologists clinical practice guideline update. J Clin Oncol \u003cb\u003e31\u003c/b\u003e, 3997\u0026ndash;4013 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFarahmand S, Fernandez AI, Ahmed FS, Rimm DL, Chuang JH, Reisenbichler E, \u003cem\u003eet al\u003c/em\u003e. Deep learning trained on hematoxylin and eosin tumor region of Interest predicts HER2 status and trastuzumab treatment response in HER2 + breast cancer. Mod Pathol \u003cb\u003e35\u003c/b\u003e, 44\u0026ndash;51 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eQaiser T, Mukherjee A, Reddy PC, Munugoti SD, Tallam V, Pitkaaho T, \u003cem\u003eet al\u003c/em\u003e. HER2 challenge contest: a detailed assessment of automated HER2 scoring algorithms in whole slide images of breast cancer tissues. Histopathology \u003cb\u003e72\u003c/b\u003e, 227\u0026ndash;238 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHelin HO, Tuominen VJ, Ylinen O, Helin HJ, Isola J. Free digital image analysis software helps to resolve equivocal scores in HER2 immunohistochemistry. Virchows Arch \u003cb\u003e468\u003c/b\u003e, 191\u0026ndash;198 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTuominen VJ, Tolonen TT, Isola J. ImmunoMembrane: a publicly available web application for digital image analysis of HER2 immunohistochemistry. Histopathology \u003cb\u003e60\u003c/b\u003e, 758\u0026ndash;767 (2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLaurinaviciene A, Dasevicius D, Ostapenko V, Jarmalaite S, Lazutka J, Laurinavicius A. Membrane connectivity estimated by digital image analysis of HER2 immunohistochemistry is concordant with visual scoring and fluorescence in situ hybridization results: algorithm evaluation on breast cancer tissue microarrays. Diagn Pathol \u003cb\u003e6\u003c/b\u003e, 87 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZakrzewski F, de Back W, Weigert M, Wenke T, Zeugner S, Mantey R, \u003cem\u003eet al\u003c/em\u003e. Automated detection of the HER2 gene amplification status in Fluorescence in situ hybridization images for the diagnostics of cancer tissues. Sci Rep \u003cb\u003e9\u003c/b\u003e, 8231 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKoopman T, Buikema HJ, Hollema H, de Bock GH, van der Vegt B. What is the added value of digital image analysis of HER2 immunohistochemistry in breast cancer in clinical practice? A study with multiple platforms. Histopathology \u003cb\u003e74\u003c/b\u003e, 917\u0026ndash;924 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAntonio C. Wolff MEHH, M. S. Bartlett MBIO, Patricia A. Spears GHVG. Human Epidermal Growth Factor Receptor 2 Testing in Breast Cancer: American Society of Clinical Oncology/ College of American Pathologists Clinical Practice Guideline Focused Update. Arch Pathol Lab Med \u003cb\u003e142\u003c/b\u003e, 1364\u0026ndash;1382 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFranchet C, Djerroudi L, Maran-Gonzalez A, Abramovici O, Antoine M, Becette V, \u003cem\u003eet al\u003c/em\u003e. Mise \u0026agrave; jour 2021 des recommandations du GEFPICS pour l\u0026rsquo;\u0026eacute;valuation du statut HER2 dans les cancers infiltrants du sein en France. Ann Pathol \u003cb\u003e41\u003c/b\u003e, 507\u0026ndash;520 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVenetis K, Crimini E, Sajjadi E, Corti C, Guerini-Rocco E, Viale G, \u003cem\u003eet al\u003c/em\u003e. HER2 Low, Ultra-low, and Novel Complementary Biomarkers: Expanding the Spectrum of HER2 Positivity in Breast Cancer. Front Mol Biosci \u003cb\u003e9\u003c/b\u003e, 834651 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Z, Lin Y, Cao Y, Hu H, Wei Y, Zhang Z, \u003cem\u003eet al\u003c/em\u003e. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows.; 2021.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang TY, Suen CY. A Fast Parallel Algorithm for Thinning Digital Patterns. Commun Acm \u003cb\u003e27\u003c/b\u003e, 236\u0026ndash;239 (1984).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYue M, Zhang J, Wang X, Yan K, Cai L, Tian K, \u003cem\u003eet al\u003c/em\u003e. Can AI-assisted microscope facilitate breast HER2 interpretation? A multi-institutional ring study. Virchows Arch \u003cb\u003e479\u003c/b\u003e, 443\u0026ndash;449 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTewary S, Arun I, Ahmed R, Chatterjee S, Mukhopadhyay S. AutoIHC-Analyzer: computer-assisted microscopy for automated membrane extraction/scoring in HER2 molecular markers. J Microsc \u003cb\u003e281\u003c/b\u003e, 87\u0026ndash;96 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRawat RR, Ortega I, Roy P, Sha F, Shibata D, Ruderman D, \u003cem\u003eet al\u003c/em\u003e. Deep learned tissue \"fingerprints\" classify breast cancers by ER/PR/Her2 status from H\u0026amp;E images. Sci Rep \u003cb\u003e10\u003c/b\u003e, 7275 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKhameneh FD, Razavi S, Kamasak M. Automated segmentation of cell membranes to evaluate HER2 status in whole slide images using a modified deep learning network. Comput Biol Med \u003cb\u003e110\u003c/b\u003e, 164\u0026ndash;174 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTarantino P, Hamilton E, Tolaney SM, Cortes J, Morganti S, Ferraro E, \u003cem\u003eet al\u003c/em\u003e. HER2-Low Breast Cancer: Pathological and Clinical Landscape. J Clin Oncol \u003cb\u003e38\u003c/b\u003e, 1951\u0026ndash;1962 (2020).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"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-1967645/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-1967645/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe new HER2-targeting antibody drug conjugate offers the opportunity to treat patients with HER2-low breast cancer. Distinguishing HER2 immunohistochemistry (IHC) scores of 0 and 1+, is critical but also challenging due to HER2 heterogeneity and variability of observers. In this study, we aimed to increase interpretation accuracy and consistency of HER2 IHC 0 and 1\u0026thinsp;+\u0026thinsp;evaluations through assistance from artificial intelligence (AI) algorithm. In addition, we examined the value of AI algorithm in evaluating HER2 IHC scores in tumors with heterogeneity. The AI-assisted interpretation consisted of AI algorithms and an augmenting reality module with microscope. Fifteen pathologists (5 junior, 5 mid-level and 5 senior) participated this multi-institutional two-round ring study that included 246 infiltrating duct carcinoma not otherwise specified (NOS) cases. In round 1, pathologists analyzed 246 HER2 IHC slides by microscope without AI assistance. After 2 weeks of washout period, the pathologists read the same slides with AI algorithm assistance and rendered the final results by adjusting to the AI algorithm. The interpretation accuracy was significantly increased with AI assistance (Accuracy 0.93 vs 0.80), as well as the evaluation precision of HER2 0 and the recall of HER2 1+. The AI algorithm also improved the total consistency (ICC\u0026thinsp;=\u0026thinsp;0.542 to 0.812), especially in HER2 1\u0026thinsp;+\u0026thinsp;cases. In cases with heterogeneity, the accuracy was improved significantly (Accuracy 0.68 to 0.89) and to similar level as cases without heterogeneity (Accuracy 0.95). Both accuracy and the consistency of junior pathologists were better improved than the mid-level and senior pathologists. To the best of our knowledge, it is the first study to show that the accuracy and consistency of HER2 IHC 0 and 1\u0026thinsp;+\u0026thinsp;evaluations and the accuracy of HER2 IHC evaluation in breast cancers with heterogeneity can be significantly improved using AI-assisted interpretation.\u003c/p\u003e","manuscriptTitle":"The role of artificial intelligence in accurate interpretation of HER2 IHC 0 and 1+ in breast cancers","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2022-08-29 14:00:41","doi":"10.21203/rs.3.rs-1967645/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":"a8107148-6b42-497a-8c91-1848707f6a08","owner":[],"postedDate":"August 29th, 2022","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2022-10-28T11:32:43+00:00","versionOfRecord":[],"versionCreatedAt":"2022-08-29 14:00:41","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-1967645","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-1967645","identity":"rs-1967645","version":["v1"]},"buildId":"zQwnuV7TCBrMSSSToR1PI","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}