Network-Agnostic Knowledge Transfer from Latent Dataset for Medical Image Segmentation
Transfer learning often employs all or part of the weights of a pre-trained net-work to the problem at hand; this limits the flexibility of new neural architectures. We propose to transfer the knowledge of a neural network (teacher) from a latent dataset to another neural network (student) by training the student on a dataset agent whose annotations are generated by the teacher. The dataset agent requires no manual annotation and is independent of the teacher-training dataset. The student does not need to inherit the weights of the teacher, and such, the proposed algorithm can be flexibly conducted between heterogeneous neural architectures. Extensive experiments on six multi-organ medical image segmentation datasets have shown that the proposed algorithm was effective for knowledge transfer and easy to be used with fine-tuning. This algorithm has the potential to be employed in novel applications where the teacher-training dataset is not accessible, particularly in medical applications.
PDF Abstract