We present a novel procedure to apply deep learning techniques to medical image classification. With increasing popularity of computed tomography (CT) lung screening, fully manual diagnosis of lung cancer puts a burden on the radiologists who need to spend hours reading through CT scanned images to identify Region of Interests (ROIs) to schedule follow-ups. Accurate computer-aided diagnosis of lung cancer can effectively reduce their workload and help training new radiologists. However, lung cancer detection is challenging because of the varying size, location, shape, and density of nodules. Many studies have approached this problem using image-processing techniques with the intention of developing an optimal set of features. Convolutional neural network has demonstrated to learn discriminative visual features automatically and has beat many state-of-art algorithms in image-processing tasks, such as pattern recognition, object detection, segmentation, etc. In this report, we evaluate the feasibility of implementing deep learning algorithms for lung cancer diagnosis with the Lung Image Database Consortium (LIDC) database. We compare the performance of two 2.5D convolutional neural network models, two 3D convolutional neural network models, and one 3D convolutional neural networks with a 3D spatial transformer network module on the task of detecting lung nodules. The performance of our best model is comparable with the state-of-art results in the lung nodule detection task.