Since we want to train on both 2D (X-ray) and 3D data (CT and MRI), we need DINOv2 to work on 3D as well. A simple way of extending it is to apply identical spatial/color/noise transformations across slices, then investigate tube vs random masking across the slices.
Since we want to train on both 2D (X-ray) and 3D data (CT and MRI), we need DINOv2 to work on 3D as well. A simple way of extending it is to apply identical spatial/color/noise transformations across slices, then investigate tube vs random masking across the slices.