Current automated 3D cell detection methods are not a suitable replacement for manual stereologic cell counting.

Area of Science:

• Neuroscience
• Biomedical Imaging
• Computational Biology

Background:

• Stereologic cell counting is crucial in neuroscience but is labor-intensive due to manual visual inspection.
• Current automated methods for cell detection and segmentation are being evaluated as alternatives.
• Manual counting is time-consuming and leads to the continued use of biased 2D methods.

Purpose of the Study:

• To evaluate the performance of automated cell detection and segmentation algorithms for stereologic cell counting.
• To compare automated methods against manual counting by an expert observer using 3D microscopic brain tissue images.

Main Methods:

• Three automated 3D cell detection algorithms were tested: FARSIGHT toolkit, 3D multiple level set methods, and ImageJ 3D object counter.
• Performance was evaluated by comparing automated results with manual counts from a 3D microscopic image dataset.
• Images used were thick brain tissue sections with common nuclear and cytoplasmic stains.

Main Results:

• FARSIGHT showed the best performance among the tested automated methods, with true-positive rates from 38-99% and false-positive rates from 3.6-82%.
• Current automated methods exhibit lower detection rates and higher false-positive rates than acceptable for unbiased stereologic counting.
• Significant discrepancies were observed between automated and manual cell identification and quantification.

Conclusions:

• Automated cell detection and segmentation algorithms are not yet sufficiently accurate for unbiased stereologic cell counting in neuroscience.
• Manual stereologic cell counting with expert decision-making remains the gold standard for accurate cell quantification in histologic sections.
• Further development is needed to improve the reliability and accuracy of automated methods for neuroscience research.