Working Group on Robotics, Telemedicine and Image Processing

Medical Applications

The interaction of HPC and Health Care in the specific areas of robotics, telemedicine and image processing is best demonstrated by considering a set of sample narratives. Each of these narratives helps eludicate the opportunities for HPC to leverage technologies to solve important problems in health care, and at the same time helps elucidate unusual requirements that such problems place on HPC.

Minimally invasive surgery
There is a strong trend within the medical community towards minimally invasive surgical procedures, with the expected benefits of reduced complications, reduced trauma for the patient, and reduced length of hospital stays, all with the expected benefit of reduced costs and increased quality of life for the patient. To effectively utilize minimally invasive procedures, one needs automatic or semi-automatic methods for localizing anatomical structures for the surgeon, where such localization is coupled with pre-surgical planning, and one needs automatic or semi-automatic methods to support navigation within the body and delivery of treatment and procedures in minimally invasive ways. Several examples serve to motivate the idea of tools for minimal invasion. One such example is enhanced reality visualizations, in which segmented and labelled anatomical models, acquired through three dimensional medical sensors (MRI, CT) are automatically registered with the actual patient, and displayed in a superimposed visualization so that the surgeon can see internal structures directly overlaid on top of the patient, from the correct viewpoint. Ideally, such structures would be tracked and their registration refined over time, to maintain a consistent visualization as the surgeon changes view, as the patient moves, or as the tissue deforms. This problem is particular relevant in endoscopic applications, since here the surgeon, by definition, has a limited field of view, and navigation and localization become critically important.

A second example is the use of robotic devices to assist a surgeon. Such devices include remote manipulation and tactile feedback devices for palpation of internal tissue, systems to deliver surgical tools and procedures to inaccesible locations, such as in sinus surgery, and tools to improve the accuracy and reliability of surgical procedures, such as the RoboDoc system used in hip replacement surgery. Key issues in this scenario are real-time processing needs, band-width needs, and computational and data reliability.

Remote Area Telemedicine Interaction
The driving problem in this scenario is one of extending the reach of quality medical care to areas outside the normal scope of major hospitals. While existing initial approaches have focused primarily on medical teleconferencing, ultimately, such telemedical applications should involve collaborative interactions. Such collaborative interactions might include telepresence expert advise, in which a remote physician is able to interact with the patient, including feeling the patient through telerobotic sensors, viewing the patient from arbitrary directions, and viewing the patient with registered sensory data overlaid on top of the patient. Clearly such an interactive ability would be of considerable utility, not only for collaborative consultation, but also as a training tool, in which a medical student can observe an expert in action, not just passively, but actively through tactile and visual interfaces. Key issues in this scenario are real-time processing needs, band-width needs, and computational and data reliability.
Routine testing
There are a number of high-volume, computationally intensive image screening applications (e.g. mammograms, PAP smears) in which semi-automated, reliable image processing methods could have a huge impact. While real-time processing is not criticial here, the huge volume of data to be processed does put serious constraints on the computational power needed. As well, while some routine testing examples would simply involve the analysis of individual acquisitions, more robust methods would also include database acquisition and manipulation. For example, in mammography the use of change detection algorithms, in which a current scan is normalized and registered to previously acquired scans of the patient, then the two scans are compared to highlight potential differences, would be of considerable interest. Such an application would be further enhanced by the ability to automatically register a new scan to a canonical atlas, including the deformation of the scan to account for patient variability. By registering to an atlas, any detected anatomical changes can be further interpreted based on knowledge of the tissue type associated with the matched portion of the atlas.
Device design
The creation of new medical devices can benefit from more extensive use of simulation. Simulations can reduce the time required to complete a design as well as the time needed for testing. With good models, the designer can evaluate the effect of various device parameters in its future physiological environment. For example, the ability to perform accurate simulations of blood flow through the heart with an artificial valve would help in the design of such devices. High performance computing can allow the implementation of a more accurate model of the heart and greatly reduce the time it takes to perform such a complex simulation.

Underlying image processing, vision and robotics problems

Each of these narrative scenarios involves one or more of a common set of technical problems, including:
Image Segmentation
Almost any acquired medical imagery needs to be segmented into tissue type, and needs to be done in a fast, reliable manner, while demonstrating insensitivity to sensor variations and artifacts. Such segmentation may be enhanced by automatic registration of newly acquired sensory data to existing anatomical atlases.
Registration
Many of the problems in this realm involve the registration of data sets. Often this registration is multi-modal (MRI vs CT vs SPECT vs PET vs Xray vs Laser) and ideally such registration should be performed in real time. This latter requirement is especially true for problems involving minimally invasive procedures, in which the surgeon needs information that can help her adjust almost instantly to situations within a patient.
Biomedical Simulation/ Physical Modeling/ Physics-Based Modeling
The structures of the body that need to be simulated are highly complex and involve complicated interactions. This stretches the performance limits of existing simulation engines.
Spatial Planning and Optimization
A prime example of the need for high performance methods in planning is the navigation of instruments, both internal and external to the patient. This is especially important in robotically assisted procedures and remote diagnosis and consultation.
Visualization
Especially in scenarios such as minimally invasive surgery, it is essential that visual information be provided to the surgeon in a fast and useful way.
Image-based connections to data bases
Given the vast quantity of visual information that can be stored, especially from routine testing such as mammograms, it is critical that such information be accessible to the physician in relevant ways. This means not only the traditional data base acquisition methods based on textual information, but also tools that support visual accessing of data. For example, given an unusal medical image, a useful tool for the physician is to ask for examples of similar imagery from a database, as a possible guide to diagnosis.
This partial list of technical areas does serve to highlight a series of common themes.
  1. While many of these technical areas are ones that have existed for some time, the interaction of these technical problems in the medical arena raise unusual problems. The size of the data sets is extreme compared to many other problem domains. The real time aspects of the domain are not common to many other domains. Reliability of processing and of derived information is extremely important (e.g. if a manufacturer makes a few unacceptable parts as part of a processing operation, this is not a big deal, while an occasional slip of a robotically wielded scalpel is clearly not acceptable).
  2. The complexity of the structures to be handled, both in terms of the complexity of representating and manipulation shape based models, as well as the physical variability of such structures, both across patients and across time, are much larger than in most other domains.
These special aspects of the medical imaging and robotic domain place strong demands on realistic solutions: in terms of the algorithmic tools needed, as well as in terms of the computational engines required.

HPCC Infrastructure Issues

There are several HPCC infrastructure issues that interact with the development of succesful solutions to these problems.

Interactive HPC
One of the unusual aspects of this particular domain is the need for interactive computation. This includes needs for high bandwith (bidirectional), high-speed I/O, reliable networking, guaranteed availability of resources, and real-time operating systems.
Software integration
There is a need for a clean integration of image analysis software with modeling software with visualization software, etc.
Robustness
This is particularly important in cases where safety is an issue, such as in robotic interventions, or in automated diagnosis systems.