The emergency of whole slide imaging (WSI) in digital pathology is becoming a routine clinical diagnosis for many cancers. However, manual cancer regions review in WSIs for diagnosis is labor-intensive and error-prone task due to large scale, high-resolution and complexity of tumor heterogeneity. In this paper, we propose a fully automatic cancer region recognition framework for computer-assisted diagnostics in pathology WSIs based on deep convolutional neural network. Our framework leveraged patch-based images with image-level benign/malignant annotation for neural network training to perform a classification task instead of pixel-level segmentation, which could improve computation efficiency and alleviate annotation workload. The evaluation has been conducted on 100 liver digital whole-slide images and experimental results demonstrated our method can achieve the segmentation accuracy of 0.880 and 0.872 at 15x and 20x respectively, which is feasible and fast cancer detection for diagnosis on WSIs.