Domain Transfer in Histopathology using Multi-ProtoNets with Interactive Prototype Adaptation

Few-shot learning addresses the problem of classification when little data or few labels are available. This is especially relevant in histopathology, where labeling must be carried out by highly trained medical experts. Prototypical Networks promise transferability to new domains by using a pre-Trained encoder and classifying by way of a prototypical representation of each class learned with few samples. We examine the applicability of this approach by attempting domain transfer from colon tissue (for training the encoder) to urothelial tissue. Furthermore, we address the problems arising from representing a class via a small amount of representatives (prototypes) by testing two different prototype calculation strategies. We compare the original "Prototype per Class"(PPC) approach to our "Prototype per Annotation"(PPA) method, which calculates one prototype for each example annotation made by the pathologist. We test the domain transfer capability of our approach on a dataset of 55 whole slide images (WSIs) containing six subtypes of urothelial carcinoma in two granularities: "Superclasses", which combines the tumorous subtypes into a single "tumor"class on top of a aggregated "healthy"and additional "necrosis"class, and "subtypes", which considers all eleven classes separately. We evaluate the classic PPC approach as well as our PPA approach on this data set. Our results show that the adaptation of the Prototypical Network from colon tissue to urothelial tissue was successful, yielding an F1 score of 0.91 for the "superclasses". Furthermore, the PPA approach performs very comparably to the PPC strategy. This makes it a viable alternative that places more value on the intent of the pathologist during annotation.