Projects
The proposed projects are:
Project 1 (Raicu): “Semantic and content-based Image Retrieval.” The long-term goal of this research study is to build a content-based image retrieval (CBIR) system for pulmonary CT nodules. Currently the texture-based retrieval methods of Gabor, Markov, and a version of Global Co-occurrence have been implemented by the REU students from summer 2006. The new task is to improve global co-occurrence and implement local co-occurrence, shape descriptors and Fourier descriptors as well as develop a retrieval system based on the annotations made by radiologists. The new image features will be combined using principal component analysis, mapped to the radiologists’ annotations using decision trees, and selected in the retrieval process based on the learned mappings.
Project 2 (Raicu): “Automatic Segmentation”. This project involves automatic liver segmentation in CT scan images of the torso. The student researchers will assess the accuracy with which a deformable parametric curve propagate to the edges of images of the Gabor Filter and Co-Occurrence matrix features when compared to propagation on the simple gray-level intensities of the original image.
Project 3 (Furst): “Volumetric texture-based classification” Volume segmentation of computed tomography (CT) images has become increasingly important as the between slice resolution of modern CT scanners has gotten better. Today, truly isotropic three dimensional CT images are available, while most of the segmentation algorithms still rely on slice-by-slice segmentation. This project will use isotropic 3D CT images of the human chest and abdomen and texture descriptions of healthy tissue to classify soft tissues with the final goal of automatically segmenting organs of interest. The emphases of this project will include: 1) minimal user interaction; 2) good boundary detection, and 3) repeatability and robustness across patients and organs.
Project 4 (Dettori): “Contrast enhancement of medical images”. Enhancing contrast in medical images offers radiologist a tool to progressively highlight regions of interests to improve diagnostic. It is particularly challenging to be able to highlight at the same time different regions, of potentially quite different grey level intensity. The goal of this project is to build a system that allows the user to specify several concurrent regions to be enhanced. We will explore and compare a variety of techniques built on histogram equalization, and other binning strategies. We will also investigate the portability of such techniques to a variety of image modality.
Project 5 (Channin): “Breast Density Estimation”. The goal of this research project is to create a volumetric breast density assessment tool using DICOM images in an XIP environment. The extensible Imaging Platform (XIP) was developed to quickly create medical imaging applications. This open source environment allows the user to choose from an extensible set of modular elements and easily create and change applications. Our approach is based on the Standard Mammographic Form algorithm that takes the assumption that every pixel of a mammographic image is composed of both fat and fibro glandular tissue. By knowing specific values from the DICOM header such as breast thickness and filter material, we are able to calculate the height of fibro glandular tissue at a given pixel. This value is then compared to the volume of the entire breast and given as a percentage. Creating this in the XIP environment will allow for access for both implementation and improvement.
Project 6 (Channin): “Tone mapping using HDR techniques”. Real-world scenes and other high dynamic range scans contain far more pixel information than current displaying devices can utilize. Tone mapping is a technique used to effectively compress high range pixel data into 8-bit image data so that devices can display these images. A tone mapping algorithm strikes a balance between keeping as much detail from the original image as possible while maintaining the original visual impression through contrast enhancement. The students will investigate several techniques for tone mapping and implement them in an XIP environment. This is a C/C++ project.