Evaluation of the deep learning-based detection of dopaminergic neurons in primary culture: A practical alternative to manual counting

Abstract

Background

Manual counting remains the gold standard for assessing neurotoxicity in cultured neurons. However, it is labor-intensive and susceptible to subjective variability, limiting its scalability and reproducibility in high-throughput studies.

New method

To address these limitations, we evaluated two artificial intelligence–based object detection methods for identifying tyrosine hydroxylase-positive dopaminergic neurons in immunostained primary cultures. Specifically, we compared a traditional cascade classifier with a deep learning-based model employing the YOLOv3 algorithm.

Results

The cascade classifier performed reasonably well in detecting healthy dopaminergic neurons but showed a high rate of false positives under neurotoxic conditions involving neuronal degeneration. In contrast, the deep learning-based model maintained high precision under both healthy and neurotoxic conditions. The deep learning model detected the neuroprotective effect of a test compound, consistent with expert manual counting. In terms of processing time, the deep learning model was more than seven times faster than manual counting.

Comparison with existing methods

While expert manual counting is commonly accepted in biological image analysis, it lacks objectivity and is not suitable for large-scale analyses. The cascade classifier provides limited utility under neurotoxic conditions. The deep learning-based model outperformed the cascade-based approach in terms of precision, especially under neurotoxic conditions.

Conclusions

Deep learning-based analysis offers a practical and reproducible alternative to manual cell counting in dopaminergic neurons. It is particularly useful in studies involving neurotoxicity or neuroprotection and has the potential to support scalable and reliable quantification in preclinical research.