Overview: AI in pathology reveals promising accuracy with troubling disparities
Artificial intelligence systems trained to diagnose cancer from pathology slides have shown impressive overall accuracy, offering the promise of faster, more consistent readings. Yet a growing body of research indicates that these systems do not perform equally well for all patients. In some cases, diagnostic accuracy can vary by age, sex, race, or ethnicity, raising critical concerns about equity in cancer care. This article unpacks the latest findings and outlines the three main reasons researchers believe these disparities occur and what the field is doing to address them.
Reason 1: Data heterogeneity and representation gaps
One of the most cited factors behind AI biases in cancer detection is data heterogeneity. Pathology slides come from diverse laboratories, scanners, and staining protocols. If the training data underrepresents certain demographic groups or cancer subtypes, the AI may struggle to generalize to these cases in real-world practice. For example, slides from older patients or from underrepresented ethnic groups may present subtle morphological features that the model has not learned to recognize. As a result, accuracy can drop when the model encounters unfamiliar patterns. Experts stress the need for diverse, representative datasets and standardized imaging protocols to ensure models learn robust, generalizable features rather than artifacts tied to a particular cohort.
Reason 2: Confounding factors linked to biology and care contexts
Pathology is influenced by a complex mix of tumor biology and the surrounding tissue environment. Demographic factors often correlate with variations in biological subtypes or stage at diagnosis, which can influence AI performance. Additionally, social determinants of health—such as access to care, prior treatments, and biospecimen handling—can indirectly affect slide quality and annotation accuracy. If a model is trained on slides with systematic differences across groups (for instance, variations in fixation quality or slide thickness that correlate with certain institutions), it may inadvertently learn to associate a demographic label with a diagnostic outcome rather than the biology itself. Addressing this requires meticulous study design, inclusion of clinically varied data, and techniques that encourage the model to focus on tumor characteristics rather than incidental batch effects.
Reason 3: Annotation quality and ground-truth inconsistencies
AI systems for cancer detection rely on high-quality annotations that guide learning. When pathologist labels, segmentations, or cancer boundaries vary—whether due to institutional practices, inter-observer variability, or differing diagnostic criteria—models can inherit inconsistent signals. These annotation gaps often align with patient demographics simply because of who contributed which samples. As a result, the AI’s decisions may reflect inconsistencies in the ground truth rather than true biological differences. To reduce this source of bias, researchers are adopting multi-expert consensus labeling, standardized annotation protocols, and robust evaluation metrics that account for variability among human experts.
What researchers are doing to improve fairness and accuracy
Despite these challenges, the field is actively pursuing remedies. Key strategies include:
- Inclusive data curation: Building larger, more diverse datasets that represent multiple demographics, tumor subtypes, and clinical settings.
- Domain adaptation and fairness-aware modeling: Techniques that reduce reliance on non-biological artifacts and minimize disparate performance across groups.
- Standardized imaging and annotation: Harmonizing staining, scanning, and labeling processes to reduce batch effects and ground-truth variability.
- Transparent reporting and external validation: Publishing performance by subgroups and validating models on independent cohorts to detect and quantify bias.
Implications for patients and clinicians
AI-assisted pathology offers substantial benefits in throughput and consistency, but disparities in diagnostic accuracy can translate into unequal quality of care. Clinicians should remain aware of these limitations and consider AI assessments as one input among multiple diagnostic modalities. For patients, ongoing conversations about how AI tools were developed and validated can help build trust and ensure that care decisions reflect the best available evidence across diverse populations.
Looking ahead
As cancer diagnostics increasingly leverage AI, achieving equitable performance is as important as breaking new technical ground. The next generation of AI pathology tools aims to be robust across demographics, tumor types, and clinical settings, reducing gaps in accuracy while preserving the speed and interpretability that make AI valuable in the lab. The path forward involves diverse data, better ground-truth standards, and a commitment to fairness at every stage of model development and implementation.
