Tumor segmentation using the learned distance metric

A novel interactive segmentation method based on distance metric learning is proposed for segmentation of tumors in CT and MRI images. Firstly, the moments of the gray-level histogram are extracted as the image features for segmentation. Then, Neighborhood Components Analysis is employed to learn a task-specific distance metric in the feature space using the interactive inputs. The probability of each pixel which belongs to the tumor and the background region is estimated by the K-Nearest Neighbor classifier with the learned distance metric. The cost function for segmentation is constructed by these probabilities. Finally, the graph cut algorithm is used to optimize the cost function. The proposed method is evaluated on the CT images of liver tumors and the MR images of brain tumors. Experimental results show that the proposed method is more robust and accurate compared to the other methods using the intensity histogram and the Euclidean distance.