I am trying to understand the 8bit and 16bit input image for training deep neural network. DICOM Xray input pixel data is uncompressed 16bit (unsigned) while most of the deep neural network are available with 8bit input image format.
I wonder if anyone has faced similar problems/issues.
Though i am able to train the network after converting uncompressed 16bit to 8bit PNG format . I wonder if in real life deployment of such model, would there be any issues ?
I would suggest trying to read the dicom file directly before converting them to a PNG image. By convertion you are loosing a pixel resolution, However, it also depends what kind of precision you are maintaining while training. In my experience I didnt see any particular increase in accuracy.