Like many tools in IPT, graythresh() depends on the class of the image array to know what its scale is. When you change classes, the data must be correctly scaled. Unless you know the difference, don't use uint16() or double() to change image class. Use im2uint16() or im2double().
% a properly-scaled uint8 image
A = uint8([1 2 3 ; 5 6 7; 10 11 12; 14 56 45])
% the threshold in unit-scale
graythresh(A)
Do the same thing in uint16. The result is the same.
% a properly-scaled uint16 image
B = im2uint16(A)
% the threshold in unit-scale
graythresh(B)
Casting without scaling means this image data is essentially black. The brightest pixel in this image is darker than 1LSB in uint8. Internally, the image is quantized to 256 levels for histogram generation. Since the entire image lies in the very bottom of the lowest bin, there's no histogram to split.
% a improperly-scaled uint16 image
% this is uint8-scale data in a uint16 container
B = uint16(A)
% so it gets interpreted completely wrong
graythresh(B)
