Cancer classification with multi-task deep learning

Abstract

Microarray technique can generate a large amount of gene expression profiles for thousands of genes simultaneously. The gene expression data has been widely used in disease diagnosis and deep learning approach has achieved great successes in this task. However, the deep learning approach may fail when the expression data for a particular tumor is insufficient for training an effective model. In this paper, we propose a novel multi-task deep learning (MTDL) to overcome the aforementioned deficiency by leveraging the knowledge among multiple expression data of related cancers. MTDL learns local features from each task with some private neurons, and learns shared features for all tasks simultaneously with some shared neurons, and learns to inference for each task separately in the end layer. Since MTDL leverages the expression data of multiple cancers, it can learn more stable representation for each cancer even its expression profiles are inadequate. The experimental results show that MTDL significantly improves the performance of diagnosing each type of cancer when it jointly learns from the expression data of twelve cancer datasets.