An automatic counting algorithm for the quantification and uncertainty analysis of the number of microglial cells trainable in small and heterogeneous datasets

A research team led by Luca Martino has published a novel AI algorithm in *Expert Systems with Applications* that automates the tedious and error-prone task of counting microglial cells in biological tissue samples. The work addresses a critical bottleneck in neuroscience and pathology, where manual counting of immunopositive cells is time-consuming and requires extensive personnel training. Traditional automated methods often fail with high-resolution images containing significant noise, providing only rough area and intensity measurements rather than accurate cell counts. This new approach, detailed in arXiv:2602.22974, omits complex cell detection pipelines entirely, focusing the AI's effort solely on the counting task itself for rat lumbar spinal cord cross-sections.

The core innovation is a 'kernel counter'—a non-parametric and non-linear method designed for efficiency and flexibility. Its most significant technical advantage is its ability to be trained on very small datasets, as its basic version relies on tuning only a single hyper-parameter. Despite this simplicity, the model is flexible enough to handle rich, heterogeneous data and can even 'overfit' if required for maximum accuracy on specific datasets. Crucially, it provides uncertainty estimates for its predictions, a feature that allows researchers to gauge the reliability of automated counts and even reconcile differing opinions from multiple human experts on the same image. The accompanying release of MATLAB code ensures other labs can implement and build upon this tool, potentially accelerating research in neuroinflammation and related fields where precise cellular quantification is essential.