Image classification for early detection of lung cancer

Nowadays, only 10 - 15% of people diagnosed with lung cancer survive more than 5 years after the diagnostic. The main cause is the delay in detecting it. One way of early detecting nodules is to use a system over chest x-rays that can classify them. A project is being carried out in order to develop this system, but this work is a previous and necessary step: it aims to separate anteroposterior and lateral images in order to make the classifier to perform better. To do so, we have studied four deep learning methods: logistic regression, multi-layer perceptron (MLP), restricted Boltzmann machines (RBM) and convolutional networks (CNN). We applied all four methods to a random sample of our dataset and registered their accuracy, specificity, sensitivity and AUC (area under curve). With these, we observed that MLP is the one with the best performance along with CNN, but the latter requires more runtime. However, if one would like to use the simplest method, logistic regression also performs well enough