Bibliographic databases in medicine, e.g., MEDLINE, have been routinely used for the last thirty years. However, medicine is a domain in which images are critical elements of the diagnostic process. Till now, the technologies making it possible to extend databases to include images or to create image databases have been limited.
Several emerging technologies are pointing the way toward making image databases practical. Among them are: high resolution scanners to convert the information on film to digital files; high capacity storage media such as digital optical disks, both WORM and CDROM; increasingly reliable archival mechanisms such as jukeboxes holding such platters representing an archive of several scores of gigabytes, and yet providing random access rather than the sequential access provided by tape; increasingly ubiquitous wide area networks such as the INTERNET with the potential eventually of reaching speeds of gigabits per second; high performance computing alternatives such as hardware and software for parallel processing.
Why is it an important topic?
Image databases, or visual information management systems (VIMS), in medicine are important primarily because so much of diagnostic information resides in images. Radiographs (xray films) are an example of a traditional source of image-based information. In recent years other imaging modalities such as CT, MRI, PET and ultrasound images have grown in importance and today are used routinely. Other examples include 35mm slides of dermatologic conditions, histologic (tissue) slides, images in cytology and pathology.
Characteristics of an imaging system
Characteristics or features of VIMS in medicine depend on the specific application, i.e., the properties of the images themselves and the objectives of the user. One relevant property is file size which varies widely. Certain images when digitized can result in very large files. For example, cervical and lumbar spine xrays taken as part of a nationwide health and nutrition examination survey when digitized at a resolution of 2000x2000 pixels, and 12 bits per pixel, amount to 5 Megabytes and 10 Megabytes respectively. On the other hand, other modalities with resolutions of the order of 512x512 pixels result in files that are considerably smaller.
The user objectives vary widely depending on the setting. Clinical requirements often postulate very rapid response, i.e., on the order of a second or two. This might call for a highly compressed image, which might call for a lossy compression technique. On the other hand, clinical usage also comes with a strong concern for legal considerations, which could be interpreted to being a bar to using lossy techniques.
How would it benefit the user?
The user, a diagnostician, a biostatician or other investigator, could benefit from a VIMS from several points of view. One, he/she could access exemplary images from a "library" store for comparison with current images. In clinical work comparison might also be in order for images of the same "scene" taken at different times, to say detect the size change of a tumor. Two, images may be transferred among different sites linked by a high speed network for consultation more readily in an electronic form than as films. Three, a sequence of 2-dimensional imagery may be used to generate 3-D views for a heightened sense of perspective (it has been shown, for example by J. Rosenman, UNC, that 3-D images often reveal features easily missed in 2-D slices). Four, images may be linked to other information sources available electronically, such as bibliographic or factual databases.
The role that an intelligent system incorporating imaging might play in say, clinical oncology, might be to address the following:
a. Where was the disease initially [detection, location]? b. What happened to the disease [temporal analysis]? c. Assessment of the tumor [detect and dilineate tumor bulk; is it benign or malignant?] d. Assessment of the therapeutic response [did the therapy reduce the tumor?]
Desirable features of a VIMS
Among the desirable features of a VIMS in medicine are: rapid access to the image store for interactive usage; image access linked to searches of bibliographic, full text or factual databases; single point-of-contact multimedia workstations to handle outputs of different databases; ability to send images seamlessly over local TCP/IP networks (e.g., Ethernet), high speed HiPPI links, and increasingly available wide area networks such as INTERNET; interactive use of image databases where the images themselves are manipulable fields of the DBMS; the extraction of quantitative information (size, shape and proximity of objects) from images containing multiple objects; tools to enable computer- aided diagnosis.
Problems to be solved to realize a suitable VIMS
Image compression. Techniques that provide high compression ratio (CR) and are also lossless would be most desirable. However, existing lossless techniques such as DPCM, Lempel-Ziv and others do not exceed a CR of 3 or so, in contrast to lossy techniques that could deliver a CR of 30 or more. What is needed are studies by expert panels presented at consensus meetings to investigate and confirm that images (of different classes) contain all relevant information at different levels of loss. Another approach would be to mount serious research into more efficient lossless techniques.
Access speed. Research into efficient cacheing algorithms, high speed communication protocols and hierarchical storage methods would be important to provide higher access speeds.
Image DBMS. Research into multimedia database structures, pictorial query languages, spatial indexing techniques, and other tools to allow access not only to images but into images.
Computer-aided Diagnosis (CADx). Tools to perform image analysis (shape extraction, texture measurement), feature extraction and link-up with content knowledge to enable the automated detection of: heart size and size changes, clustered microcalcifications in mammograms, lung nodules, etc.