The proposed projects are:
Visualization techniques for medical data
Content-based medical image retrieval systems
Machine learning approaches for texture-based segmentation
Automatic detection of intensity spots in micro-array images
Accelerating visualization techniques
User interface development
Evaluation of segmentation algorithms
Texture classification algorithms testing
Project 1 (Raicu): Visualization techniques for medical data. The size and complexity of medical data sets from various medical imaging modalities (such as Magnetic Resonance Imaging (MRI), Computed Tomography (CT), Positron Emission Tomography (PET) and Ultrasonography) makes it increasingly difficult to understand, compare, analyze and communicate the data. Visualization is an attempt to simplify these tasks by representing as a single image the relationships among data, and enabling the recognition of the patterns in the data and the navigation through complex and disparate sets of data. The students working on this project will experience, develop, and assess different visualization tools (http://www.cc.nih.gov/cip/visualization/vis_packages.html), including VTK (a popular freely available object-oriented visualization toolkit, http://public.kitware.com/VTK), for exploring Computed Tomography data of the chest and abdomen.
Project 2 (Furst): Volumetric segmentation. This project will continue work begun last summer, in which a user clicks on a point in a 3D computed tomography of the human abdomen, and an algorithm automatically segments the organ clicked. The algorithm currently uses Haralick texture descriptors and a distance histogram within a breadth-first search to segment organs. Students working on this project will investigate which descriptors are best for conducting the segmentation, modifications to the breadth-first search, distance metrics for populating the distance histogram, and stopping criteria for the search.
Project 3 (Raicu): Content-based medical image retrieval systems. In medicine to date, virtually all Picture Archiving and Communications Systems (PACS) retrieve images simply by textual indices based on patient name, technique, or some-observer-coded text of diagnostic findings. This textual approach, however, may suffer from considerable observer variability, high cost of manual classification and manipulation of images by medical experts, and failure to fully account for quantitative relationships of medically relevant structures within an image that are visible to a trained observer but not codable in conventional database terms. Therefore, suitable database structures addressing the visual/spatial properties of medical images and more effective techniques to deal with different types of knowledge are necessary; Content-based Image Retrieval (CBIR) systems can be the tool that will allow researchers and medical practioners to automatically access images directly by their content and complement the textual retrieval approach currently in use. In particular, students working on this project will develop a CBIR system for the Lung Image Database Consortium (LIDC, http://radiology.rsnajnls.org/cgi/content/full/232/3/739) and evaluate the system with respect to 1) its retrieval performance (precision and recall); 2) different low-level feature vector representations calculated at local or/and global image level; and 3) various similarity measures.
Project 4 (Furst): Machine learning approaches for texture-based segmentation. This project will give students the opportunity to work alongside a doctoral student in achieving segmentation of computed tomography using split and merge algorithms. The segmentation will be based on pixel-level Haralick texture descriptors, and students will investigate issues relating to window size of the texture calculations, which descriptors are most effective, splitting and merging criteria and machine learning algorithms to classify individual pixels and by extension, regions.
Project 5 (Furst): Automatic detection of intensity spots in micro-array images. Micro-array data analysis attempts to learn information about unknown DNA by submitting it to a substrate of RNA probes arranged in a grid of points on a glass slide. Where the RNA and DNA bind, a radioactive marker will fluoresce, signifying the binding. An image of the fluorescing probes will generate a pattern that can be used to learn about the composition of the unknown DNA. A significant problem in the analysis of micro-array data is the determination of relative intensities of active probes in images of the micro-array. This project will examine the task of reading images that contain significant noise and imaging artifacts, and generating a reliable and robust way of identifying the relative intensity of the spots with respect to the image background. Applications of microarray technologies include forensics, medical diagnosis, gene mapping, and disease tracking.
Project 6 (Channin): Ontology development. An ontology describes a set of concepts and the relationships among them. It also defines attributes (“slots”) for those terms, and relationships of various types among those terms. We want to build pieces of an Ontology of Radiology that can be used to provide knowledge to imaging informatics applications. Students participating in this project will learn how to use the Protégé ontology editor from Stanford University and its JAVA API. We will build ontology components from medical imaging standards such as DICOM and develop prototype classes where standards do not exist.
Project 7 (Channin): Accelerating visualization technology. Students participating in this project will assist in porting visualization software from general purpose CPU execution to GPU (Graphical Processing Unit) execution on high end graphics cards, such as those from NVIDIA and ATI.
Project 8 (Channin): User interface development. Navigating large image datasets requires new human interface devices. Students participating in this project will work to interface a variety of devices into workstation development efforts. They will also experiment with graphical user interfaces to these devices and their evaluation.
Project 9 (Dettori): Evaluation of segmentation algorithms. Image segmentation is often the first critical step in the process of such tasks in the medical domain as tissue classification, content-based image retrieval, and computer-aided diagnosis. Many segmentation methods are still evaluated using a subjective human opinion of quality with a lack of quantitative analysis. Ideally, segmentation would be performed on an image with as little aid from a human user as possible, so solid quantitative analysis of results and optimization of user-defined parameters are a must. Last summer a methodology for evaluating unsupervised segmentation algorithm was developed and a metric based on a statistical analysis of the overlap between machine segmented and referenced ground truth images was introduced. Students working on this project will test the effectiveness of such metric in evaluating and optimize intra-algorithm parameters, and compare inter-algorithm performance for unsupervised segmentation algorithms.
Project 10 (Dettori): Texture classification algorithms testing. Several algorithms for texture-based classification of normal tissues in computed tomography images of the chest and abdomen have been developed by this group in the last year using texture features derived from a variety of image features including co-occurrence, run-length, wavelet, ridgelet and curvelet transforms. Mostly, such algorithms have been test on a single dataset obtained from same-machine scans on two patients. A variety of new datasets from different machines and modalities are now available. Students working on this project will evaluate the performance of existing classification algorithms when applied to the new datasets.