Artificial Intelligence for Assisted HER2 Immunohistochemistry Evaluation of Breast Cancer: A Systematic Review and Meta-Analysis

Si Wu,Xiang Li,Jiaxian Miao,Dongyi Xian,Meng Yue,Hongbo Liu,Shishun Fan,Weiwei Wei,Yueping Liu
DOI: https://doi.org/10.1016/j.prp.2024.155472
2024-01-01
Abstract:Accurate assessment of HER2 expression in tumor tissue is crucial for determining HER2-targeted treatment options. Nevertheless, pathologists' assessments of HER2 status are less objective than automated, computer-based evaluations. Artificial Intelligence (AI) promises enhanced accuracy and reproducibility in HER2 interpretation. This study aimed to systematically evaluate current AI algorithms for HER2 immunohistochemical diagnosis, offering insights to guide the development of more adaptable algorithms in response to evolving HER2 assessment practices. A comprehensive data search of the PubMed, Embase, Cochrane, and Web of Science databases was conducted using a combination of subject terms and free text. A total of 4994 computational pathology articles published from inception to September 2023 identifying HER2 expression in breast cancer were retrieved. After applying predefined inclusion and exclusion criteria, seven studies were selected. These seven studies comprised 6867 HER2 identification tasks, with two studies employing the HER2-CONNECT algorithm, two using the CNN algorithm, one with the multi-class logistic regression algorithm, and two using the HER2 4B5 algorithm. AI's sensitivity and specificity for distinguishing HER2 0/1+ were 0.98 [0.92 - 0.99] and 0.92 [0.80 - 0.97] respectively. For distinguishing HER2 2+, the sensitivity and specificity were 0.78 [0.50 - 0.92] and 0.98 [0.93 - 0.99], respectively. For HER2 3+ distinction, AI exhibited a sensitivity of 0.99 [0.98 - 1.00] and specificity of 0.99 [0.97 - 1.00]. Furthermore, due to the lack of HER2-targeted therapies for HER2-negative patients in the past, pathologists may have neglected to distinguish between HER2 0 and 1+, leaving room for improvement in the performance of artificial intelligence (AI) in this differentiation. AI excels in automating the assessment of HER2 immunohistochemistry, showing promising results despite slight variations in performance across different HER2 status. While incorporating AI algorithms into the pathology workflow for HER2 assessment poses challenges in standardization, application patterns, and ethical considerations, ongoing advancements suggest its potential as a widely effective tool for pathologists in clinical practice in the near future.
What problem does this paper attempt to address?