Semi-supervised Multi-domain Learning for Medical Image Classification

Ruchika Chavhan, Biplab Banerjee, Nibaran Das

Abstract: The limitations of domain dependence in neural networks and data scarcity are addressed in this paper by analyzing the problem of semi-supervised medical image classification across multiple visual domains using a single integrated framework. Under this premise, we learn a universal parametric family of neural networks, which share a majority of their weights across domains by learning a few adaptive domain-specific parameters. We train these universal networks on a suitable pretext task that captures a meaningful representation for image classification and further finetune the networks using a small fraction of training data. We perform our experiments on five medical datasets spanning breast, cervical, and colorectal cancer. Extensive experiments on architectures of domain-adaptive parameters demonstrate that our data-deficient universal model performs equivalently to a fully supervised setup, rendering a semi-supervised multi-domain setting with lower numbers of training samples for medical data extremely feasible in the real world.