Investigation of cardinality classification for bacterial colony counting using explainable artificial intelligence

Visual similarity, not model design, is the main barrier: Using explainable AI techniques, researchers found that MicrobiaNet's difficulty distinguishing colonies of three or more individuals stems primarily from high visual similarity across different colony count categories in the data itself, rather than inherent flaws in the model architecture as previously believed.

Explainable AI revealed hidden data constraints: The study demonstrates that XAI methods can uncover how dataset properties—not just model capabilities—limit performance in real-world applications, suggesting that improving AI systems sometimes requires fixing data problems rather than redesigning algorithms.

Imbalanced datasets pose broader challenges: The findings have implications beyond bacterial counting, suggesting that neural network classifiers trained on imbalanced datasets (where some categories have fewer examples than others) may face similar visual similarity challenges that XAI could help identify.

Future solutions should match the actual problem: Rather than continuing to refine cardinality classification models, the research recommends that future work should either develop models that explicitly account for visual similarity between classes or explore alternative approaches like density estimation, ensuring development efforts target the real bottleneck.