Deep learning-based localization algorithms on fluorescence human brain 3D reconstruction: a comparative study using...

Area of Science:

• Neuroscience
• Computational Biology
• Medical Imaging

Background:

• Advancements in tissue clearing and fluorescence microscopy enable high-resolution 3D human brain reconstruction.
• Analyzing large-scale brain data necessitates automated, accurate cell counting and localization methods.
• Deep learning (DL) offers powerful cell segmentation tools, but human brain data presents unique quantification challenges.

Purpose of the Study:

• To empirically evaluate three DL techniques (StarDist, CellPose, BCFind-v2) for human neuron quantification.
• To compare DL methods against a 3D stereological design for large-scale brain analysis.
• To guide researchers in selecting optimal DL tools based on specific research needs.

Main Methods:

• Evaluation of StarDist, CellPose, and BCFind-v2 on 3D human brain data.
• Utilized a 3D stereological design as a reference standard for accuracy.
• Compared methods based on density/localization accuracy, computational efficiency, and annotation effort.

Main Results:

• DL approaches demonstrated high effectiveness in neuron quantification.
• Methods provided high-throughput 3D cell localization.
• Results showed accuracy comparable to established stereological methods.

Conclusions:

• Deep learning techniques are highly effective for analyzing large-scale 3D human brain datasets.
• These automated methods significantly improve efficiency and accuracy in neuron counting and localization.
• The study provides valuable insights for selecting appropriate DL tools for diverse neuroscience research objectives.