Speaker
Description
Brains are composed of neurons, containing axons. Myelin aids in propagating electrical and chemical signals by insulating axons. Postmortem studies suggest that axon properties change with various neurological conditions. Imaging methods can aid in the diagnosis process and understanding how axons change over time. Electron microscopy (EM) is the current gold standard to measure microstructures in the brain due to its high resolution. Manual segmentation of EM images is extremely time-consuming, prohibiting large-scale studies. The ultimate goal of this project is to have an automated segmentation method for EM data that can accurately delineate axons, myelin, and other cells for comparison with a newly developed magnetic resonance (MR) in vivo method. The first step compared an automated segmentation model with the corresponding manual segmentations for axons. Eight EM images from one male mouse were manually segmented, defining intra-axonal space and myelin. The light microscopy model of Segment Anything for Microscopy (µsam) was used for automatic segmentation of the axons in the images. The manual and automated segmentation of each image was compared using the average Dice similarity coefficient (DSC), precision and accuracy. The average and standard deviation of the DSC, precision and accuracy over all images were 0.78±0.03, 0.63±0.04 and 0.70±0.03, respectively. Compared to the intra-axonal and myelin manual segmentations, the average DSC, precision and accuracy were 0.91±0.02, 0.93±0.05 and 0.85±0.03, respectively. Manual segmentation took approximately 42 hours per image, whereas the automated prediction was completed on average 3.9 seconds per image. The reduced segmentation time comes at the cost of µsam only being able to detect larger cells, whereas the longer manual segmentation time allowed for every cell to be segmented. The results demonstrate automated segmentation predictions better represent the intra-axonal space and myelin rather than just intra-axonal space. The next steps are to test other automated models to evaluate if they perform better compared to µsam and to MR results on the same tissue. Additional models should be investigated to determine if they are able to automatically segment other cells and separate the myelin from the axon.
| Keyword-1 | automated segmentation |
|---|---|
| Keyword-2 | medical imaging |