Multi-Class Multi-Instance Learning for Lung Cancer Image Classification Based on Bag Feature Selection

In lung cancer image classification, the label concepts are usually given out for the whole image but not for a single cell, which leads to a low predict accuracy if we use supervised learning methods on cell-level. In this paper we model lung cancer image classification as a multi-class multi-instance learning problem. A lung cancer image is treated as a bag. Each bag contains a set of instances that are lung cancer cells. In our approach, we first extract the features for cells in all images as bags, and then transform each bag into a new bag feature space by computing the Hausdorff distance in all of the bags. At last we use AdaBoost algorithm to select the bag features and build two-level classifiers to solve the multi-class classification problem. Experiments on the lung cancer image dataset show that our approach is an effective solution for the lung cancer classification problem.