Parametric Transfer Learning on Prostate and Breast Cancer
Intricate structure is analyzed by deep learning algorithms by extracting hierarchical feature representations. Employing deep learning algorithms for such computer-aided classification and detection has been exhaustively analyzed and has proven to out-perform human detection from many perspectives. Heterogeneous transfer learning algorithms learn features by cross-mapping between feature spaces by estimating millions of parameters. The convolutional neural network model analysis that we perform uses a cross-domain transfer learning approach to learn the feature-spaces from one source domain and map correlated feature instance correspondences in the target domain with noted transitions from generality to specificity of the neurons from one domain to the other, across the layers of the deep neural network. In this work, we exploit methods of transferring parameters across layers and analyzing the correlation between the domains to extract features to develop computer-aided classification and detection systems. A method of transferring parameters learnt from one feature space of a correlated medical domain to support the learning of the parameters from another feature space of the same medical domain is performed in this work.