Automated Tracing of Neuronal Processes via 3D Image Segmentation and Contrast Enhancement
Philip Cho
pnc10@duke.edu
| Paper PDF |
|
Tracing of neuronal processes is regularly performed in virtually any research laboratory dedicated to cellular neuroscience, yet it remains a laborious procedure reliant on error-prone human judgement. Here, I introduce an image segmentation algorithm, built with a U-Net architecture and a contrast-modulating physical layer, and I demonstrate its capability in automatically tracing fluorescently labeled neurons. The effects of contrast are also analyzed deeply as a physical parameter by experimenting with various point-wise intensity transformations, and a comparison between models indicates that a power-law transformation in the physical layer – although inducing a longer necessary training period – can potentially yield optimal tracing performance with further modifications and investigation. |
|
|
| Paper: |