Invited Talks

Image Matching and Recognition from Invariant Local Features

David Lowe

University of British Columbia

Within the past few years, methods for identifying invariant local features in images have provided a powerful new approach to matching images.  Each feature is invariant to imaging scale, orientation, and location, yet carries enough information to select potential matches in a large database of previously seen features.  Reliable recognition is achieved by identifying clusters of consistent feature matches followed by detailed model fitting.  Unlike many other approaches to object recognition, the process uses no prior object segmentation and is unaffected by background clutter.  Recent work will be presented on applications to location recognition, augmented reality, and the detection of image panoramas from unordered sets of images.  A live demonstration will be given of a system that can recognize objects at near real-time speeds.

Image guided interventions: "The glass patient"

Wiro Niessen

Imaging Sciences Institute

University of Utrecht

The discovery of X-rays by Wilhelm Conrad Rontgen on November 8, 1895, ended a period in which anatomical information could only be acquired using invasive techniques. Already in the first months after their discovery, the potential of these "strange rays" for medicine were understood, not only for diagnosis, but also for therapy. In fact, the concept of image guided interventions is already more than one hundred years old.

While intra-operative imaging has been around for a long time, the concept of image guidance on preoperatively acquired data only took off in the last decade of the previous century; by registering a patient to a CT scan, and optically tracking surgical instruments, neurosurgical interventions could be carried out while navigating on a preoperatively acquired CT. This technology was subsequently applied to other anatomies. The principle of bringing diagnostic quality imaging to the tip of the surgeon's instrument was born.

Image guided surgery (or when minimally invasive: intervention) has some distinct advantages. First, current imaging modalities provide detailed four-dimensional anatomical and functional information that can be used for planning and guiding of interventions. Second, by providing intra-operative guidance, interventions can be carried out less invasive. Third, the use of intra-operative radiation may be limited or prevented.

After a historical overview, the concepts and technologies underlying image guided surgery will be discussed.

Subsequently, the state of the art and challenges in image guided interventions will be discussed, and ample examples in the fields of image guided neurosurgery, orthopaedic surgery, maxiollofacial surgery and cardiovascular interventions will be given. Finally the expected impact of the novel developments in imaging at the cellular and molecular level on image guided interventions will be discussed.

Understanding biological systems with the help of pattern discovery methods

Isidore Rigoutsos

Bioinformatics and Pattern Discovery Group

IBM Watson Research Center

Understanding biological systems with the help of pattern discovery methods In recent years, considerable amounts of research activity has been focused on the interpretation of large, diverse sets of biological measurements in order to elucidate the complex mechanisms that underly important and (seemingly simple) macroscopic phenotypes.

The problem at hand is hierarchical in nature, with the hierarchy spanning many levels. Each of these levels can be thought of as comprising multiple active agents that are diverse in their nature (e.g. genes, proteins, pathways, organelles, etc) and also in their behavior.   It is within this setting that one seeks to build an integrated view of the system under study, as soon as the relevant units and the complex inter- and intra-level relationships in which these units participate have been characterized.

Implicit in the above outline are the following assumptions: a complete, and, presumably correct list of parts exists for the system that is being studied; and, most, if not all, of the important relationships involving these parts are available.

Through the research work of my group and of others, there is increasing evidence that the situation is likely to be more complicated than initially estimated, and that one should be watchful when it comes to making or relying on the above two assumptions.

In this presentation, I will give an overview of several pattern discovery algorithms that we have developed and then show how they can be applied to solving a very wide range of biological problems.  The resulting answers are examples of features and relationships that are relevant for the kinds of questions that arise in the above hierarchical context.  Every effort will be made to make the presentation self-contained.