Automatic pathological image diagnosis is an important topic in medical image analysis, and the prerequisite for an accurate pathological image diagnosis is to capture the distinctive morphological features of normal and abnormal tissues. With a deep neural network as a tool, a boosting convolutional neural network is proposed, in which a pair of complementary networks is trained to optimize the accuracy of a pathological image diagnosis. To reduce the risk of over-fitting caused by the scarce training examples due to the high cost of obtaining pathological images, in the proposed algorithm, a basic classifier is first trained to estimate the probabilities of local tissues being abnormal, and then another heterogeneous network is trained to correct the predictions made by the basic one. The extensive experiments are carried out on the Cancer Metastasis Detection on Lymph Node dataset and the Animal Diagnostics Lab dataset provided by Pennsylvania State University which contains the pathological images of three organs (i.e., kidney, lung and spleen). The experimental results show that the proposed model can be used to achieve a high accuracy on the pathological images of different organs.