The VH Female dataset license was obtained from NLM and extensive image preprocessing was performed to computationally crop, "deblue", sharpen, downsample, compressed leading to the creation of two additional orthogonal views of the data, the coronal and the sagittal. The data was then compressed using the JPEG compression standard to ~70% and resampled at low, medium and full resolution. Images were arranged on a hard disk system linked to an HP-735 workstation configured as a WWW server, linked to the Internet by FDDI. A Java-based navigation tool was implemented in Netscape, which allows a "cube" of head (or pelvis) data to be browsed interactively in real time by clients. Cross-sectional slices are then selected by remote users and delivered via the Internet to the client station at the selected resolution. Performance using UNIX, PC, and Mac workstations linked to the Internet via , cable modem, FDDI, 10baseT, standard Ethernet, and by 28.8 kb/s modem has been measured and evaluated (data to be published and discussed elsewhere).
Stage 2 involved labeling areas of anatomical interest within each two-dimensional transverse slice, and storing that information in our relational database. Only select landmarks and anatomical regions taught during the Gross Anatomy courses were identified by Professor John Lillie, UM Medical School. We developed a modular, Java-based software package to facilitate the labeling activity. Java was chosen due to its platform independence and accessibility throughout the UM Medical School and elsewhere. Anatomists could work in their offices, teaching labs or any UM computer facility, regardless of the variability among clients. The Java graphical user interface (GUI) incorporated advanced data import, scaling, tracing, and labeling features that could store and download information into our database. The coordinates from a three-dimensional region of anatomical interest were associated with their anatomical label and visa versa. That is, a student could "click" on an area in the image and call up the label, or could enter the label in a text box and list the images that contain the region of interest.
During the third and final stage, we digitally processed the raw images for distribution on the W3 and developed a Java interface for navigation into the dataset. Image processing occurred in a series of multiple steps. First, the dataset was off-loaded from 8mm tape archive to local hard drives and uncompressed to produce the raw, transverse slices. Next, slices were converted into the portable pixel map (PPM) file format to allow the use of essential UNIX image processing and conversion tools, both readily available and proprietary. Each image was then cropped to remove the grayscale bar and post-it note along the bottom edge. Algorithms were developed and used to remove the blue gelatin background from each image and convert these pixel intensities to black. Finally, the original transverse (axial) slices were used to construct two orthogonal views (sagittal and coronal). Algorithms stripped a column or row of pixels from each transverse image to build the resulting sagittal or coronal slice. All data were converted to JPEG format with a compression factor that reduced the images to 70% quality (on average, file size was reduced by greater than half). JPEG compression is inherently lossy, and thus, image quality is slightly reduced. However, JPEG is currently the only available standard for distributing 24-bit imagery on the W3, currently a severe limitation of the W3. A 70% quality factor was selected to accommodate the storage capacity on our W3 server. Medium resolution images were downsampled to a 0.5 spatial resolution from the original data. Low resolution images have a 0.5 lower spatial resolution than the medium resolution images.
Figure 1 illustrates a central focus for the current UM system and future implementations. It shows the structure of the bioinformatics architecture that will allow for the anatomical labels to be effectively integrated with the VH images for recall and delivery to the user on demand using natural language linked to appropriate medical dictionaries.
The Java navigational tool was designed to work with both the head and pelvis data. Down-sampled images are displayed in the browser's three windows. Slice numbers can be roughly adjusted by dragging the colored bars, or fine-tuned using the arrow buttons. Selecting the "high", "medium", or "low" resolution button will display that image on-screen. Clicking in "hot spots", or regions of anatomical interest on the downloaded image will call appropriate labels. Conversely, entering text labels will list appropriate slice numbers.
Equipment from Silicon Graphics, Inc. (SGI) and Hewlett-Packard (HP) were used for image processing, Java development, and web serving. Both UNIX and Macintosh computers were used during the labeling effort. The database resides on a Sun workstation (see Figure 3).
A Java navigational tool (Figure 2) was designed to work with both the head and pelvis data. Downsampled images are displayed in the browser's three windows. Slice numbers can be roughly adjusted by dragging the colored bars, or fine-tuned using the arrow buttons. Selecting the "high", "medium", or "low" resolution button will display that image on-screen. Labels for discrete slices can be displayed by clicking the "labeled" button or removed by clicking the "unlabeled" button.