Abstract
Automated Ki-67 detection in breast-cancer histopathology has remained near macro-F $1\approx 86$ on BCData across successive architectures, where macro-F1 is the mean of the F1 scores for the positive- and negative-cell classes. We ask whether the detection errors reflect model capacity or a property of the data. Three architecturally distinct families — ResNet-50 +FPN, HRNet-W32, and MaxViT-Tiny are trained on BCData under matched protocols. Their test macro-F1 scores differ by at most 0.41 points, and their mined false-positive sets show substantial triple-overlap. In contrast, a lightweight ResNet-18 trained to distinguish Ki-67-positive from negative nuclei on crops of annotated cells reaches macro- $F1 = 97.12$ . This result suggests that positive-versus-negative label confusion is not a dominant source of error. To examine whether the remaining gap involves unannotated cell-like objects, we mine persistent out-of-fold false positives as an operational third label category. A three-class classifier trained to separate annotated positives, annotated negatives, and the mined category achieved a macro- $F1$ of 78.72. However, performance was strongly asymmetric: the mined category reached only $F1 = 56.29$ , indicating that separating mined false-positive samples from annotated cells is challenging. A blinded two-reader review of a 300-cell hard-enriched sample further supports this interpretation. The reviewers produced closely similar projected marginal vote distributions for the mined category; across all 300 cells, raw agreement was 65.0% with Cohen’s $\kappa = 0.48$ , while inter-reader disagreement was concentrated primarily at the boundary between positive/negative tumor-cell annotation and an unannotated outcome. Independent validation on SHIDC-BC-Ki-67 reproduced the same qualitative patterns of shared class-3 detections across model families and poor separation of class-3 from annotated tumor cells. Together, these findings suggest that part of the remaining BCData performance gap is associated with difficult-to-separate annotation-boundary cases involving unannotated cell-like objects. This motivates further investigation of supervision-focused approaches, including cleaner annotations and explicit modelling of annotation uncertainty, as potential ways to reduce this residual error component.