The Gleason grading system is the primary method pathologists use to assess prostate cancer aggressiveness. By examining tissue architecture under a microscope, they assign patterns ranging from 3 (resembling organized glandular structures) to 5 (showing the least gland-like organization), which together determine a Gleason score from 6 to 10. Higher scores indicate more aggressive cancer requiring more intensive treatment.
AI models have demonstrated strong ability to predict Gleason scores from histopathological images, sometimes matching or exceeding human performance. However, most existing AI approaches work as black boxes that produce a grade without explaining which specific tissue features drove the decision. This lack of transparency is a meaningful barrier to clinical adoption, particularly in fields like pathology where patients have a legal right to explanation under European regulations.
Common post-hoc explainability methods like Grad-CAM and LIME attempt to add transparency after the fact by highlighting image regions that influenced predictions. However, these methods provide only superficial visual explanations that are disconnected from pathological terminology, require specialized expertise to interpret, and carry a significant risk of confirmation bias, where viewers unconsciously see what they expect rather than what the model genuinely detected.
This paper introduces GleasonXAI, an inherently explainable AI model that describes tissue findings using the same histological terminology that pathologists use, making its reasoning transparent by design rather than through post-hoc approximation. The model was trained on annotations from 54 international pathologists who labeled specific tissue features according to standardized ISUP and GUPS guidelines.
The foundation of the approach is a medical ontology, a structured vocabulary of histological concepts developed in collaboration with an experienced uro-pathologist and reviewed by a panel of nine experts. The ontology organizes findings into three levels: Gleason patterns (3, 4, and 5), broader explanations such as poorly formed glands or cribriform glands, and detailed sub-explanations describing specific morphological variants within each broader category.
An international team of 54 pathologists from ten countries, with a median of 15 years of clinical experience, annotated 1,015 tissue microarray (TMA) core images from three institutional datasets. For each image region, pathologists marked which specific ontology concepts were present in which locations, rather than simply labeling the final Gleason pattern. This created annotations tied to clinical reasoning, not just outcomes.
Each image was independently annotated by three pathologists for the detailed explanatory features. When pathologists disagreed with provided Gleason grades, they could specify alternative patterns. This design captured the genuine diversity of expert opinion rather than enforcing artificial consensus, which is crucial for understanding where the tissue classification task is genuinely ambiguous.
Gleason patterns were merged across annotators using the STAPLE algorithm, a statistical method that estimates the true underlying label by accounting for each annotator's individual reliability. For the detailed explanation-level annotations, disagreements between annotators were preserved as probability distributions called soft labels, rather than forcing a single majority vote that would discard minority opinions.
Agreement between pathologists varied considerably depending on which histological feature was being assessed. Gleason pattern assignment showed relatively high agreement, with Fleiss' kappa values ranging from 0.23 to 1.00 across annotator groups, and most disagreements occurring at the clinically expected boundary between patterns 3 and 4 and between patterns 4 and 5.
At the explanation level, agreement dropped substantially. Some features like poorly formed glands (kappa 0.50) and individual glands (kappa 0.61) showed moderate to substantial agreement. However, rare features like single cells and compressed glands showed kappa values as low as 0.15 and 0.18, reflecting that even experienced uro-pathologists frequently disagree on the presence and extent of these specific tissue patterns.
At the pixel level, agreement was lower still. While 97.5% of foreground pixels could be assigned an unambiguous majority Gleason pattern, this dropped to 86.4% at the explanation level and only 67.8% at the sub-explanation level. Classes with the fewest annotated pixels were precisely those where annotators agreed least, suggesting that rarity and difficulty are linked properties of these tissue features.
This detailed characterization of variability is itself a scientific contribution: it maps exactly where human pathologists agree and where they do not, providing an upper bound on the performance any AI system can realistically achieve on each feature class and identifying which aspects of Gleason grading most urgently need standardization.
GleasonXAI uses a U-Net architecture with an EfficientNet-B4 encoder pre-trained on ImageNet. Instead of training directly to predict Gleason patterns, the model is trained to predict the ontology-based explanations at each pixel. The Gleason pattern classification emerges from aggregating these explanation predictions according to the ontology hierarchy, creating a natural pathway for explanation.
This design is called a concept bottleneck approach: the model must first predict which pathological concepts are present before it can assign a Gleason grade. A clinician reviewing the output can inspect which explanations the model found at each tissue location, and trace the grade assignment directly back to specific tissue feature detections, rather than receiving an unexplained final label.
Training used soft labels rather than majority-voted hard labels. Each pixel's target is represented as a probability distribution across classes, reflecting the genuine disagreement between annotators. A custom SoftDiceLoss function was developed to train on these distributional targets, extending the standard Dice loss from binary or categorical targets to continuous probability distributions.
The benefit of soft labels becomes especially clear in class imbalance settings. When comparing the Macro Dice score of cross-entropy models trained on soft labels (mean 0.625) versus those trained on majority-voted hard labels (mean 0.168), the soft label approach shows dramatically better performance. Majority voting discards 13.6% of foreground pixels where annotators disagree, while soft labels retain all of them, providing a more complete and honest learning signal.
The key performance finding is that training on explanation-level features did not reduce segmentation accuracy for Gleason patterns compared to training directly on Gleason patterns. GleasonXAI achieved a Dice score of 0.713 (standard deviation 0.003) on Gleason pattern segmentation, compared to 0.691 (SD 0.010) for the conventional direct approach trained on Gleason patterns without explanations. The explainable model performed comparably to and slightly exceeded the black-box baseline.
This result directly challenges the common assumption that explainability requires a performance tradeoff. The fact that adding the explanation layer did not degrade performance suggests that the pathological concepts in the ontology are genuinely informative features, not just post-hoc labels applied to already-made decisions.
Looking at individual explanation classes, the model predicted most explanations reliably, with per-class accuracy ranging from 44.4% for poorly formed glands to 73.8% for individual glands. Misclassifications predominantly occurred between adjacent Gleason patterns, exactly mirroring the pattern of human disagreement documented in the annotation study, confirming that the model's errors are clinically coherent rather than random.
The rarest features including comedonecrosis, glomeruloid glands, and single cells could not be reliably predicted, a limitation directly attributable to their scarcity in the training data and extremely low inter-rater agreement. However, the model's predictions were not entirely blind to these classes: they were consistently assigned as the second-most-probable class in affected regions, and misclassifications stayed within the same Gleason pattern, avoiding dangerous grade errors.
The ability of GleasonXAI to produce outputs in pathologist-defined language addresses a fundamental clinical adoption barrier. When an AI model can say that it identified cribriform glands and poorly formed glands as the basis for assigning Gleason pattern 4, a pathologist can agree or disagree with each specific finding. This is fundamentally different from reviewing a Grad-CAM heatmap that highlights a region without explaining what tissue property it detected.
Prior work in analogous domains has shown that pathologist-like explainable AI increases user trust and confidence compared to black-box alternatives, and can reduce reading time particularly for non-experts. In a healthcare context facing declining specialist numbers and rising cancer incidence, tools that help less experienced pathologists access expert-level explanations are particularly valuable.
The model currently focuses on tissue microarray core images, which have smaller physical size and are easier to annotate precisely than whole-slide images. Preliminary tests applying GleasonXAI to whole-slide images showed promising results but revealed a need for refinement, particularly for Gleason pattern 5. Extending the approach to whole-slide images is identified as the primary next development step.
A conceptual gap identified by the study is that current evaluation metrics for soft-label segmentation still rely on hard-label comparisons. Even a perfectly calibrated model that captures the full probability distribution of pathologist opinion cannot be fully evaluated with metrics based on majority votes. Developing evaluation frameworks that natively handle predictive uncertainty is an important open problem for the broader field of medical AI.
The study publicly releases its annotated dataset of 1,015 TMA core images with localized, concept-based pathological explanations, making it the largest openly available dataset of its kind. This resource enables other researchers to build on the GleasonXAI framework, train competing models under comparable conditions, and study the structure of pathologist reasoning as captured in expert annotations.
The annotation methodology itself is a contribution independent of the model. By recruiting 54 pathologists internationally, documenting their experience levels and caseloads, measuring inter-rater reliability at multiple levels of detail, and preserving disagreements rather than discarding them, the study provides a template for how to build training data for medical AI in tasks characterized by genuine expert subjectivity.
The inherent explainability approach may also improve regulatory pathway for clinical deployment. Post-hoc explainability methods like Grad-CAM are increasingly scrutinized by regulators because they do not reflect the model's actual reasoning process. An architecture where the explanation is integral to the prediction mechanism provides stronger transparency guarantees and may be easier to validate in a regulatory context.
The authors envision GleasonXAI as a tool that assists pathologists in grading decisions by surfacing specific tissue features for their review, rather than replacing their judgment. In a workflow context, this could reduce the cognitive load of reviewing complex cases, flag borderline regions that warrant additional scrutiny, and provide documentation of the specific tissue findings that support a grading decision, improving audit trails for quality assurance.