Multimodal input substitution for RGB

Hi, I was wondering if it would be feasible to substitute the traditional RGB image style for three different medical modalities during SAM training and inference. For example, instead of having red, green, and blue channels, have three slices of data, each representing a different medical imaging modality (each is 224x224).

Would SAM reasonably be able to learn how to use information from each of these slices without too much fine tuning?

No file, test, or entry point is named. First inspect the repository’s training and inference entry points to locate the RGB-channel assumptions; done would require a concrete implementation scope and validation plan for three medical modalities.

Written by the indexing model from the issue text.