Quantifying morphologies of developing neuronal cells using deep learning with imperfect annotations

Area of Science:

• Neuroscience
• Biomedical Imaging
• Computational Biology

Background:

• Neuronal morphology quantification is vital for understanding human brain function.
• Existing deep learning (DL) methods often require extensive, precise manual annotations for training, which is time-consuming.
• Developing efficient and accurate automated methods for neuronal analysis is essential.

Purpose of the Study:

• To propose a novel DL-based framework for segmenting and quantifying neuronal structures in fluorescence microscopy images.
• To develop a method that reduces the burden of data preparation by accepting imperfect neuron annotations.
• To accelerate the analysis of neuronal morphology in cultured neuronal cells.

Main Methods:

• A modified PSPNet with an EfficientNet backbone, pre-trained on CityScapes, was utilized.
• A weighted combination of Dice loss and Lovász loss functions was incorporated to handle imperfect training data.
• The framework was trained and evaluated on a dataset of approximately 900 manually quantified cultured mouse neurons.

Main Results:

• The proposed method demonstrated a close correlation with manual quantification for neuron length and branch number.
• The framework achieved improved analysis speed compared to existing methods.
• High accuracy in neuron segmentation was confirmed through evaluations of neuron length and branch count.

Conclusions:

• The developed DL framework offers an efficient and accurate approach for neuronal morphology quantification.
• The method's ability to utilize imperfect annotations significantly speeds up training data preparation.
• This advancement facilitates more comprehensive investigations into brain functionality through neuronal analysis.