Hi,
MAISI’s Diffusion Net doesn’t train adequately with a dataset of around 250 medical volumes. I’ve tried to train it on a similar dataset (MSD Task07_Pancreas), but I get the same result: the loss value stops falling after the first few epochs, and the generated images collapse into noise with no adequate intermediate results.
Also, I carried out a sanity-overfitting experiment, training on only 10 samples, but after around 500 epochs, no improvement was observed, and the generations collapsed into noise again.
Does this mean the training data is insufficient? Does MAISI require training its models on high volumes of data, as stated in the original paper (around 11K samples)?
Thanks very much in advance.