Training with 16 bits images

[robotechnics]

I am currently working with XRay 16 bits images on object detection. I noticed that YOLO dataloaders uses cv2 to load images, which as it is currently coded automatically converts the image to 8 bits, hence, loosing plenty of subtle information.
Is there any limitation or modification required that I may introduced to train 16 bits images?. How could I do it?. I do not mind if the model cannot be used on standard flows, but I think it will learn more features that are missed when converted to 8 bits.