Ki-67 is widely used as a marker to assess cellular proliferation in biopsy and surgical specimens from patients with breast cancer. However, determining Ki-67 expression by visually assessing stained tissues is time-consuming and has limited inter-observer repeatability.
Researchers from the Department of Pathology, The Fourth Hospital of Hebei Medical University, used artificial intelligence (AI) to establish an automated pipeline for assessing Ki-67 expression in breast cancer tissues.1 The use of AI to measure Ki-67 expression provided high inter-observer repeatability and might open new avenues in the implementation of Ki-67-based digital pathology applications in diagnostic pathology. The study was published in the journal Diagnostic Pathology on January 30, 2022.
Study Rationale: Improving Ki-67 Staining Reproducibility in Breast Cancer Tissues
Ki-67 staining using immunohistochemistry (IHC) is often employed to assess cancer cell proliferation in breast cancer tissues. Mounting evidence also suggests that Ki-67 is a prognostic factor associated with the risk of recurrence and survival in patients with breast cancer. Nevertheless, the lack of automated scoring methods and standardized Ki67 interpretation guidelines results in high intra-observer variability and hinders the adoption of Ki-67 testing in diagnostic pathology of breast cancer.1
According to the updated recommendations from the International Ki-67 in Breast Cancer Working Group, the clinical use of immunohistochemical staining for Ki-67 in breast cancer is limited to prognosis assessment in patients with stage I or II disease.2
Building an AI-Assisted Model for Ki-67 Staining Interpretation
To improve the interpretation and repeatability of Ki-67 IHC staining in breast cancer, Li and coworkers developed an AI model to assess Ki-67 expression in breast cancer tissues.1 They collected surgical specimens and pathological data from 300 patients with invasive breast cancer who underwent radical surgical resection and divided the samples into training and validation sets (150 specimens per set).
After staining the specimens with hematoxylin and eosin (H&E) and Ki-67 IHC, they scanned the slides and built a standard reference set of tissues with Ki-67 labeling index ranging from 5% to 90%. Three blinded breast pathologists reviewed the digital slides to determine the Ki-67 labeling index, which was defined as the percentage of Ki-67-positive tumor cells.
The training set was used to train the AI model to determine the Ki-67 labeling index based on multiple regions (interpretation frames) of digital slides. This was achieved by counting the number of Ki-67-positive tumor cells and the total number of tumor cells in each interpretation frame using a deep learning network. For comparison, slides in the training set were also reviewed under a microscope by three blinded pathologists. The pathologists manually counted Ki-67-positive tumor cells in three tissue areas to determine the Ki-67 labeling index.
Three pathologists analyzed the validation set manually, as well as using the AI model, to confirm the ability of AI to interpret Ki-67 IHC staining and determine the intra-group correlation coefficient as a measure of repeatability.
AI Can Accelerate Interpretation of Ki-67 Labeling Index
Manual counting of Ki-67-positive cells in tumor tissues is time-consuming and may contribute to diagnostic delays. Li and colleagues showed that AI could accelerate the interpretation of Ki-67 labeling index in breast cancer tissues, with an average evaluation time of 100–120 seconds per slide. In comparison, manual counting of Ki-67-positive cells required an average of 240–480 seconds per slide.1
In addition to being less time-consuming, AI-assisted measurement of the Ki-67 labeling index provided results consistent with the gold standard method for Ki-67 labeling assessment in breast cancer. The gold standard method was considered a combination of AI and manual methods to determine the percentage of Ki-67-positive cells.1
AI Provides Low Inter-Observer Variability in Ki-67 Labeling Index
In the training set, AI provided an intra-group correlation coefficient (ICC) of 0.972 (95% confidence interval [CI], 0.964–0.978), whereas manual counting provided an ICC of 0.803 (95% CI, 0.763–0.841). The high repeatability of determining the Ki-67 labeling index using AI was confirmed in the validation set, in which the model provided an ICCs of 0.988 (95% CI, 0.985–0.911).
Among the breast cancer tissues with homogeneous Ki-67 expression, both manual counting and the AI model provided low inter-observer variability. The ICC in the training set was 0.941 (95% CI, 0.914–0.963) for manual counting and 0.984 (95% CI, 0.976–0.990) for the AI model. However, when breast cancer tissues had heterogeneous Ki-67 expression, AI provided lower inter-observer variability than manual counting. Among tissues with heterogeneous Ki-67 expression, the ICC was 0.672 (95% CI, 0.605–0.738) for manual counting and 0.959 (95% CI, 0.947,0.970) for the AI model.
Future Perspectives
Although AI can help pathologists determine Ki-67 expression in breast cancer tissues, several improvements are warranted to allow the implementation of AI architectures in diagnostic laboratories. One of the most important aspects that needs to be improved in AI-assisted methods for determining Ki-67 expression is their ability to distinguish different cell types in the complex tumor microenvironment. Particularly in lymphocyte-rich tumors, the accuracy of AI methods in identifying tumor cells is lower than that of manual tissue observation. This may result in overestimation or underestimation of Ki-67 labeling index. Combining AI and manual identification of tumor cells may help overcome this problem.1
The effect of using different scanners and analysis systems on Ki-67 staining interpretation remains unclear. Future efforts are needed to ensure inter-platform standardization of AI-assisted methods for determining Ki-67 expression in breast cancer.2
Furthermore, validation of standardized AI architectures in large multicenter studies is required to confirm the performance and reproducibility of AI-assisted methods in determining Ki-76 labeling index in breast cancer tissues.
To learn more about the use of AI for determining Ki-67 labeling index in breast cancer, read the article by Li et al., “Artificial intelligence-assisted interpretation of Ki-67 expression and repeatability in breast cancer,” Diagnostic Pathology 17(1):20 (2022).
References
- Li L, Han D, Yu Y, Li J, Liu Y. Artificial intelligence-assisted interpretation of Ki-67 expression and repeatability in breast cancer. Diagn Pathol. 2022;17(1):20. doi:10.1186/s13000-022-01196-6
- Nielsen TO, Leung SCY, Rimm DL, et al. Assessment of Ki67 in Breast Cancer: Updated Recommendations From the International Ki67 in Breast Cancer Working Group. J Natl Cancer Inst. 2021;113(7):808-819. doi:10.1093/jnci/djaa201
Christos Evangelou, PhD
Christos received his Masters in Cancer Biology from Heidelberg University and PhD from the University of Manchester. After working as a scientist in cancer research for ten years, Christos decided to switch gears and start a career as a medical writer and editor. He is passionate about communicating science and translating complex science into clear messages for the scientific community and the wider public.
