Transfer learned model from binary to multiclass

Hi, I’m working on a disease detection project. I found 10+ datasets with binary labels (disease/no disease) and 4 dataset with labels indicating the gravity of the disease (0-3, 0 being no disease).

I was wondering if it was possible to train on the 10 datasets binary datasets and then transfer what I learned to a multiclass training.

Sure, this is equivalent to finetuning. The low-level features (closer to your input) will be pretrained by the large datasets and you only learn the high-level features (the ones helping decision / classification) when finetuning on the smaller datasets.

Great! Do I simply have to load the weightss of the binary pretrained model on the multiclass one, and then train some more on the multiclass?

(Sorry for the delay, I was away)
That is one way of doing it, but the classical approach is generally to freeze the low-level features (disable gradient computing & updating of weights) and only finetune the high-level ones with a lower learning-rate.