Abstract PR-06: Analysis of pathologists’ intraobserver, interobserver and AI agreement in breast cancer HER2 scoring: AI-assessed intra-sample tumor heterogeneity relates to lower agreement among pathologists and with AI
Bibliographic record
Abstract
Abstract Objective: This study aimed to evaluate the intra- and interobserver variability among pathologists in the assessment of HER2 immunohistochemistry (IHC) in breast cancer samples and to compare their performance with that of an artificial intelligence (AI) algorithm. We sought to measure the extent of the disagreements between observers and AI and explore potential underlying causes for variability. Methods: Thirty-four pathologists, including generalists and breast specialists, from the pathology department of a multicenter Brazilian hospital network were asked to assess 176 digital HER2 IHC samples in three testing sessions. The study design included an initial 10-image survey, a standardization workshop, and two subsequent 100-image surveys separated by one month. Repeated samples were included to allow for intraobserver variability analysis. A subset of 100 cases was also analyzed by an AI model that provided an overall HER2 score (0, 1+, 2+, or 3+) and a spatial breakdown of HER2 categories across the tumor area. Results: Four pathologists did not complete the final survey. Their data were excluded from intraobserver analysis but included in interobserver and AI comparisons. Median intraobserver concordance among pathologists was 67.68%, and median concordance with AI was 60.80%. Median interobserver agreement, using one response per pathologist, was 67.65% per sample. Median agreement with AI, using all responses, was 59.38% per sample. Out of 100 samples assessed by AI, 22 were scored as 0, 45 as 1+, 25 as 2+, and 8 as 3+. Significant positive correlations were found between interobserver and AI agreement (r = 0.86, p < 2.2×10-16), intra- and interobserver agreement (r = 0.74, p = 1.19×10-15), and intraobserver and AI agreement (r = 0.68, p = 1.47×10-11), indicating that more consistently scored samples had higher concordance across all metrics. Samples in which AI identified a higher proportion of the tumor area as matching the overall HER2 score (e.g., 90% of the tumor was 1+ and the final score was 1+) showed significantly higher interobserver (r = 0.65, p = 4.11x10-13) and AI agreement (r = 0.58, p = 2.57×10-10). In contrast, neither total tumor area nor absolute area matching the overall score showed correlation with variability metrics. This suggests that lower tumor heterogeneity within a sample, regardless of the tumor area assessed, is associated with more reproducible scoring. Conclusions: HER2 IHC interpretation remains subject to significant variability, even among experienced pathologists. Our findings highlight that tumor heterogeneity within a sample can challenge assessment reproducibility and suggest that AI can assist in identifying and evaluating challenging cases. While AI has the potential to improve reliability, overreliance on these tools may distance pathologists from critical reflection. Our results support the notion that proper integration of AI models into diagnostic workflows can enhance consistency in IHC scoring while helping pathologists recognize and address sources of variability in their assessments. Citation Format: Pedro S. S. M. Ferrari, Mariana P. Macedo, Isabela W. Cunha, Fernando A. Soares. Analysis of pathologists’ intraobserver, interobserver and AI agreement in breast cancer HER2 scoring: AI-assessed intra-sample tumor heterogeneity relates to lower agreement among pathologists and with AI [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr PR-06.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.029 | 0.077 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.001 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.002 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".