Inference failed saving prediction

[ATATC]

See #84.

[ATATC]

@perctrix can you check

[perctrix]

For 3D, that works. For 2D cases, it failed due to SimpleITK limited support for dtypes:
RuntimeError: Exception thrown in SimpleITK ImageFileWriter_Execute PNG supports unsigned char and unsigned short
Solution is adding dtype convert in save_image to convert dtype to torch.uint8. Additionally, only torch.uint8 is appropriate here because other unsigned short (such as torch.uint16, torch.uint32) are limited support by torch and will save unexpected image to file.

[perctrix]

BTW, I suggest adding os.mkdir() before saving the image. Currently, if the folder does not exist, without making the folder manually (which people tend to forget and hard to understand from the error message), SimpleITK will also raise the error RuntimeError: Exception thrown in SimpleITK ImageFileWriter_Execute: Problem while opening the file.

[ATATC]

I added error handling.

[perctrix]

Also, dtype not supported is the most important problem here.

[ATATC]

What if the image is float?

[perctrix]

Well, just convert to an integer... because for .PNG, SimpleITK only supports 8-bit unsigned integer pixels.

[ATATC]

There is a helper function for it, actually. I have included auto_convert() in save_image(). Does it work now?

[perctrix]

rn the auto_convert returns a tensor with dtype torch.int32. In save_image, we have to convert to uint8 manually

[ATATC]

What abt now?

[perctrix]

It technically works, but the output is still not aligned with our expected visualization result. Currently, we would like the prediction output to look like this:

However, what we actually obtain is:

In the second sample, some skeletal structure is visible, but it remains extremely faint. This is because the model outputs raw logits with varying dynamic ranges (e.g., [-5, 11]), and Normalize(domain=(0, 255), strict=False) adapts the target upper bound to the actual maximum value. When logits max at 11, the output becomes [0, 11] instead of [0, 255], resulting in only 4.3% brightness.
The solution is to use strict=True:

Normalize(domain=(0, 255), strict=True)(image).int()

to force full-range normalization regardless of input dynamic range.

[perctrix]

Additionally, I tested that if we normalize after sigmoid, we also need strict=True.

[ATATC]

That normalization should be done separately. Some images are intended to have a maximum value that is not 255. We should not narrow that down to a strict range.