WBCAtt+: Fine-Grained Pixel-Level Morphological Annotations for White Blood Cell Images

Detailed annotations improve AI explainability in medical imaging: WBCAtt+ provides pixel-level morphological details (11 attributes and 5 cell components) rather than just cell category labels, enabling AI models to explain why they make diagnostic decisions—a critical requirement for responsible deployment of AI in clinical settings.

Dataset transparency supports responsible AI development: By publicly releasing 113,000 image-level labels and 10,000 segmentation maps, the researchers enable independent validation and auditing of AI models used in blood disorder diagnosis, helping ensure these systems are trustworthy before clinical use.

Compositional AI models can better reflect expert reasoning: The paper demonstrates that models incorporating the structural composition of cells (how morphological features relate to each other) perform better than simpler approaches, suggesting that AI systems designed to mirror how pathologists actually think may be more reliable and interpretable.

Responsible AI requires comprehensive training data: The gap the authors identify—that existing datasets lacked the morphological details pathologists actually use—illustrates why responsible AI in healthcare requires datasets that capture the full complexity of expert decision-making, not just simplified labels.

Explainable AI applications become possible with rich annotations: The dataset enables advanced responsible AI techniques like counterfactual example generation, which can help clinicians understand edge cases and potential failure modes of diagnostic AI systems.