Transfer Learning for Kernel Matching Pursuit on Computer-Aided Diagnoses

It is commonplace for the lack of labeled data in novel domains on medical image computer-aided diagnosis but there have been some labeled data or prior knowledge in old correlative domains. In this paper, instance-transfer approach is introduced into medical image processing. And then we present a novel transfer learning model based on kernel matching pursuit called TLKMP, which extends KMP (kernel matching pursuit learning machine, Vincent & Bengio, 2002). TLKMP uses the Greedy Approximation Residue to transfer instances into target domains which have little labeled set different distributions from the source domains. So, valuable instances in resource domains are reused to construct high quality classification model for the unlabeled set of the target domains. The experiment is performed datasets on Gastric Cancer of Lymph Node database which comes from some a hospital. The results show that the proposed algorithm has better classification performance compared with traditional KMP methods, and it improves diagnosis accuracy rate of medical images effectively the same as the algorithm need lest labeled data for training a good classification model.