by Christos Evangelou, MSc, PhD – Medical Writer and Editor
Researchers at Pattern Computer, Inc. and the University of California San Diego developed an explainable AI system that not only detects mitotic figures but also presents a set of similar prior examples, some supporting the call and others arguing against it. In validation studies, the explanation layer closely approximated the detector’s local decision-making behavior (R² = 0.96). According to the authors, the AI system is meant not just to automate mitosis detection, but to make its reasoning visible at the point where clinicians may want to verify, challenge, or learn from an AI output.
The study was published in Scientific Reports.
Study Rationale
Manual assessment of mitotic figures on H&E-stained slides is limited by the sparsity of true mitoses, which can look different across phases and are easy to confuse with apoptotic bodies or lymphocytes. Although deep learning can overcome some of the limitations of manual assessment of mitotic figures, most systems remain hard to interpret.
“Our motivation was to address a broader challenge in AI, which is the lack of trust in deep learning systems due to limited interpretability,” corresponding author Nidhi Singh, PhD, said in an interview with Pathology News. “While these models often perform well, their ‘black-box’ nature limits adoption in high-stakes domains like pathology.” She added that digital pathology, and mitosis detection in particular, offered “an ideal testbed, as it is both clinically important and visually challenging, even for experts.”
Singh argued that if pathologists often reason by comparing a suspicious structure with reference cases, an AI system should be able to do something similar and show its work.
“Our approach explains decisions in a way that aligns with how pathologists already reason,” Singh said.
How the System Works
The researchers combined a standard object detector (Faster R-CNN with a ResNet-50 backbone and feature pyramid network) with a post hoc explanation method they called Adaptive Example Selection, or AES. AES operates as a post hoc explanation layer applied to the detector’s predictions.
The detector scans image patches, identifies candidate mitotic figures, and assigns a confidence score to each predicted bounding box. AES then operates on those predictions by retrieving prototype images from prior annotated cases. Some are “supporting” prototypes that push the decision toward mitosis; others are “contradicting” prototypes that push away from it.
“We built an explainable AI system that not only detects mitotic figures but also explains each prediction by showing a small set of similar examples from prior cases,” Singh explained. “These examples include both supporting and contrasting evidence, helping users understand what influenced the model’s decision. In essence, the system mimics how pathologists reason by comparing new cases to known examples.”
Many explainability methods in pathology rely on saliency maps that highlight image regions believed to drive the output. The authors argue that those maps can be ambiguous, especially for small structures like mitotic figures. Their method instead retrieves real examples with signed influence, showing not only why the model favored a positive call but also why it rejected the alternative. The authors present this approach as a way to bring contrastive reasoning (supporting and contradicting evidence) into object detection, which is a harder problem than whole-image classification because the system must localize multiple small objects and estimate confidence for each one.
Model Performance
For the detection component, the team trained on the MIDOG++ dataset, a large multi-domain collection of expert-annotated mitotic figures spanning 503 histological specimens, seven tumor types, human and canine tissue, and images from multiple laboratories and scanners. The research team used 512 × 512-pixel patches and data augmentation to improve generalization across staining and imaging conditions.
Performance varied by tumor type. The F1-scores ranged from 0.57 in human neuroendocrine tumor to 0.71 in canine soft tissue sarcoma, 0.75 in human breast cancer, 0.77 in human melanoma, and 0.84 in canine cutaneous mast cell tumor.
When the best AES configuration was used, the R² was 0.96 between the sparse prototype model and the detector’s local confidence surface, with a median of ten prototypes needed to explain 90% of a prediction’s score and a global prototype dictionary of roughly 190 images.
“The key takeaway is that explanations can be both highly faithful to the model and easy for clinicians to interpret,” Singh said. “In our study, we achieved 96% fidelity while using only a small number of examples, typically around ten per case.”
What the Explanations Reveal in Practice
The team examined AES in scenarios of correct predictions, false positives, and false negatives. In high-confidence correct mitotic calls, AES usually returned strongly weighted mitotic prototypes and no contradicting examples. Borderline correct calls contained a few weak negative prototypes.
When non-mitotic structures were misclassified as mitoses (i.e., false positives), AES often retrieved several positively weighted mitotic prototypes and only weak negative examples.
“One interesting finding was how clearly the explanations revealed failure modes, such as false positives driven by subtle visual similarities, which are often difficult to identify without interpretability tools,” Singh noted.
False negatives (i.e., missed mitoses) were often accompanied by several negative prototypes, or only weak positive support. In a few cases, the explanation revealed mislabeled or ambiguous prototypes in the training data.
Implications and Future Work
According to Singh, this study helps bridge the gap between AI performance and clinical trust.
“By making model predictions transparent and evidence-based, pathologists can better understand, validate, and, when needed, challenge AI outputs,” Singh said. “This can improve diagnostic consistency, reduce uncertainty, and serve as a valuable training tool for junior pathologists by linking complex diagnostic decisions to concrete visual examples.”
The authors acknowledge that AES depends on the quality and breadth of its prototype library. Rare morphologies, staining shifts, scanner differences, and underrepresented tumor appearances can all make the retrieved examples less informative. The system has also not yet been validated in prospective clinical workflows or at the whole-slide scale.
“Several important directions remain,” Singh said. “Expanding to larger and more diverse datasets will improve both the underlying detector and the quality of explanations.”
She added that the group is exploring alternative model architectures, including ensemble methods, and plans to scale the method to whole-slide images, test it in prospective clinical trials, and study how pathologists interact with it in real-time workflows.
The study did not receive financial support from funding agencies in the public, commercial, or not-for-profit sectors.
References
- Banik M, Kreutz-Delgado K, Mohanty I, Brown JB, Singh N. Adaptive example selection for prototype based explainable mitosis detection in digital pathology. Sci Rep. 2026;16(1):9481. doi:10.1038/s41598-026-40283-2
