The sparse structure for traditional Convolutional Neural Network (CNN) can not preserve the high efficiency of the dense network-intensive computing and the low utilization of convolution features in the experimental process,which leads to inaccurate classification results and slow convergence rate. To solve above problems, a method of extracting image features that multi-scale method combined with a deconvolution network algorithm (MSDCNN) based on CNN was proposed, and applied in the glandular images classification. Firstly, the deconvolution operation was used to achieve the fusion of different scale features, then the multi-scales feature was extracted by using different scale convolution kernels of the Inception structure. Finally, the image was classified by Softmax method. Experiments were conducted on the pathological cell images of glandular, and the experimental results showed that, the accuracy of MSDCNN algorithm is about 14% higher than that of traditional CNN algorithm under the same feature scale as the last layer, and the accuracy of MSDCNN algorithm is about 1.2% higher than that of fusion network model based on multi-scale features.