Point positioning and counting network: A deep learning-based method for automatic axon counting.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computational Biology

Background:

• Accurate axon density quantification is crucial for understanding neurological disorders affecting the optic nerve.
• Manual counting is labor-intensive and prone to errors.
• Existing automated methods have limitations in accuracy and reliability.

Purpose of the Study:

• To introduce the Point Positioning and Counting Network (PPCNet), a novel deep learning framework for automated axon quantification.
• To overcome the limitations of manual counting and current automated tools for axon density measurement.
• To validate PPCNet's performance against state-of-the-art methods on optic nerve datasets.

Main Methods:

• Developed PPCNet, a point-annotation-based deep learning model with a VGG16 backbone and dual-branch architecture for localization and confidence scoring.
• Integrated an optimized Hungarian algorithm for accurate point correspondence.
• Trained the model end-to-end using a hybrid loss function (MSE and cross-entropy).

Main Results:

• PPCNet demonstrated superior performance in axon counting on goat optic nerve datasets compared to Axonet 2.0 and AxonDeepSeg.
• Achieved higher agreement with manual counts (R²=0.939) than comparative methods.
• Exhibited lower Mean Absolute Error (MAE=45.93) and narrower Bland-Altman limits of agreement, indicating high reliability and reduced bias.

Conclusions:

• PPCNet provides a reliable and accurate automated solution for axon quantification in optic nerve research.
• The model surpasses traditional manual counting and segmentation-based automated methods.
• PPCNet facilitates efficient and precise analysis of axonal degeneration in neurological studies.