A deep fusion framework for unlabeled data-driven tumor recognition

Abstract Traditional pattern recognition problems are usually accomplished through two successive stages of representation and classification, the generalization ability and stability are difficult to guarantee for small samples and category imbalance. For tackling these problems, an unlabeled data-driven representation learning classification (RLC) fused model is constructed by integrating representation learning and classification into one model, rather than simple putting the two stages together. The RLC fused model mainly focuses on interactive iteratively optimizing representation learning and classification in a model, guiding and reinforcing each other. Under the framework of RLC, a deep nonnegative matrix factorization (NMF) is adopted for representation learning by complementing the advantages of NMF and deep learning, and avoiding complex network structure and parameter modulation. The framework is called deep NMF-RLC fusion model, which can achieve good performance for binary classification even the simplest linear regression classifier is used. The model explores useful information embedded in unlabeled data, and is suitable for small training samples and unbalanced classification. The performance of the proposed framework is verified on genetic-based tumor recognition, which contains all three stages of early diagnosis, tumor type recognition and postoperative metastasis. Experiments show that, compared with the published state-of-the-art methods and results, there are significant improvements in classification accuracy, specificity and sensitivity.