Multi-task deep convolutional neural network for cancer diagnosis

Using computational techniques especially deep learning methods to facilitate and enhance cancer detection and diagnosis is a promising and important area. Nowadays, gene expression data has been widely used to train an effective deep neural network for precise cancer diagnosis. However, if a particular tumor has insufficient gene expressions, the trained deep neural networks may lead to a bad cancer diagnosis performance. In this paper, we propose a novel multi-task deep learning (MTDL) method to solve the data insufficiency problem. Since MTDL leverages the knowledge among the expression data of multiple cancers to learn a more stable representation for rare cancers, it can boost cancer diagnosis performance even if their expression data are inadequate. The experimental results show that MTDL significantly improves the performance of diagnosing every type of cancer when it learns from the aggregation of the expression data of twelve types of cancers.