U-Net

U-Net is pretty much the go-to model architecture for biomedical image segmentation. Yet stock U-Net has not yet been shown to be particularly effective when applied to images of neurons.

On the other hand, recurrent variants might well be a fruitful avenue of research. For example, there is a 2018 pre-print¹ which seems promising. The recurrent aspect allows the model to maintain an internal memory/state about remote areas of a given neuron. Neurons are very, very long compared to other cells. The recurrence can span such distances. In the pre-print recurrent U-Net is applied to retinas.

References

• U-Net: Convolutional Networks for Biomedical Image Segmentation
• Recurrent
  • Recurrent U-Net for Resource-Constrained Segmentation
    • (Summer-2019 pre-print out of China)
    • This sort of model archetype might have enough smarts to tackle computer viz object recognition of neurons. Very complex structures, very repetitive though. One might even say recurrent.
  • Recurrent Residual Convolutional Neural Network based on U-Net (R2U-Net) for Medical Image Segmentation

Test

This is text with a ¹ rando foot note