Continual-Zoo: Leveraging Zoo Models for Continual Classification of Medical Images

Abstract

In medical imaging, leveraging continual learning (CL) is key for models to adapt to new classes and data distributions without forgetting prior knowledge. Existing CL methods often overlook the use of off-the-shelf pretrained models that are equipped with informative and generalizable representations, opting instead to learn from scratch. In this paper, we propose Continual-Zoo, a novel CL paradigm that smartly leverages a zoo of pretrained models for continual medical image classification. For a given task, Continual-Zoo distills pertinent knowledge from the fixed zoo through cross-knowledge and semantic-knowledge attention mechanisms to obtain class prototypes. Since deploying a zoo could lead to scalability issues with a large number of models, we propose a novel prototypical variational autoencoder, pVAE, as a zoo knowledge encoder. During inference, Continual-Zoo utilizes pVAE as a feature extractor that maps images to the same space of class prototypes and returns the class whose prototype has the shortest distance in the latent space. To mitigate forgetting in CL, pVAE leverages the class prototypes to synthesize images from previously learned tasks before adapting to new ones. Experimental results on various clinical benchmarks demonstrate the superiority of Continual-Zoo over SOTA methods in class-incremental, domain-incremental, and domain and class-incremental learning scenarios, distinguishing it from most CL methods.Code is available at here.