Update WSI foreground segmentation algorithm

[nkaenzig]

Closes #455

I used a simplified version of the segmentation algorithm from CLAM. CLAM on top of what is implemented here filters the approximate contours of the detected foreground objects based on an area threshold as a post processing step. We don't need this step here, because our patching logic anyways applies another foreground threshold filter to the sampled patches, so very small foreground areas in the mask most likely won't yield patches anyways.

Comparison

• "grayscale-threshold" is what we've been using so far (for datasets such as panda which have a clean white background this worked reasonably well, but it breaks for datasets such as camelyon16 where this is not the case)
• mb stands for median blur applied to saturation channel in HSV image
• otsu thresholding applied to saturation channel doesn't seem to perform better than using a fixed threshold
• I would recommend disabling the fill_holes in our default setup, as filled holes can result in patches sampled from within holes which might not contain relevant tissue data.

Avg Runtime per WSI [s]

grayscale_threshold 0.089027
hsv 0.071794
hsv + mb 0.138696
hsv + mb + holes 0.140928
hsv + mb + otsu 0.138782
hsv + mb + otsu + holes 0.143519

• median blur is computationally the most expensive step, but it generates cleaner and smoother masks. 138ms per WSI is still fast enough, so I recommend to enable it.

[ioangatop]

Looks great 🚀