I am trying to train a model for reconstruction. My CBCT dataset is of DICOM format with datatype uint16. When I normalize the volume to uint8 based on the maximum intensity of the volume, my model doesn’t give expected performance. But when I normalize each axial slices separately, my model performs better. But the images have dark lines in the middle. I understand the reason is because the middle regions are where teeth is located and since the teeth intensity are much larger compared to other regions these slices becomes dark.
Does anyone have suggestions on how to solve this issue?
Quantizing to 8 bit is not a good idea in general of course. You are going to lose detail that may be important.
But also normalizing on a slice-by-slice level will change the relative intensities. You might choose to normalize by the standard deviation over the volume instead of the max.
1 Like