Approaches

From what I've read online, we have a few choices to isolate the brain (feel free to add to this list)

Possible Approaches

• Thresholding (Very Naive. Good Benchmark)
• Canny Edge Detection + Morphological "Fill in" (Subject to discontinuous boundaries. The canny edge detector tries to connect its edges through "Hysteresis thresholding", but this sometimes fails.)
• Sobel Filter + Markers + Watershed Transform (Seems like the best option so far -- without empirical validation)
• Histogram-Based segmentation -- what amounts to "Bump hunting" and cutting on modes

There are many choices of filters that will pick out edges. Both Sobel and Canny use a convolution kernel which essentially calculates derivatives of gray scale values in the 8 cardinal directions.

We might want to think about:

Preprocessing

• Denoising
• Smoothing with a Gaussian or "Disk Median"
• Morphological Opening and Closing to get rid of scattered members in contiguous regions