ProtoMIL: Multiple Instance Learning with Prototypical Parts for Fine-Grained Interpretability