OnLine Images from the History of Medicine:
Technical Limitations & Future Plans

Technical Limitations

• The intensity of images will vary depending upon the gamma value of the display device. The images as stored are appropriate for the system on which they were extracted and displayed, which employs a 8-bit color monitor (gamma = 2.6).
• The resolution of the images is rather low, though adequate for browsing purposes. The limited resolution is most troublesome for reproductions with fine print, such as posters and manuscript pages; print is sometimes unreadable. Original materials were photographed on black-and-white or color 35 mm film. These were transferred to a laser videodisc, reducing them to the 525-line (maximum) resolution provided by NTSC television images. (Refer to the history and credits documentation to read more about this work, from an earlier project). The laser videodisc images were extracted on a Sun SPARCstation 10/41 workstation using a Sun VideoPix frame grabber board, which reduced resolution further. Finally, the captured images were stored on disk using 8-bit JPEG compression, a "lossy" compression method which entails further image degradation. If interest in this system is sufficient to justify the effort, the images should be rescanned directly from 35 mm film.

  For comparison purposes, here are several images which are provided in in several of the forms just discussed. One is a black-and-white photograph of nurses taken aboard a hospital ship, the U.S.S. Relief, during the Spanish-American war. Images scanned from 35 mm film were obtained using a Nikon film scanner attached to a Macintosh II computer running Adobe PhotoShop. Scanned film images were saved as TIFF files. The nurses image required 7,802,140 bytes of storage in TIFF format. When converted to GIF, the image required 694,104 bytes. When converted to 8-bit JPEG, storage was reduced to 120,416 bytes. Such reductions are common. The size of the 8-bit JPEG file extracted from the laser videodisc is even smaller, due to the vertical resolution limit of NTSC (this image is smaller than the one scanned from film), and required only 48,540 bytes. Because there is no difference between the resolution of the TIFF and GIF images, only the GIF image is shown here. There is little difference between the 8-bit JPEG and GIF images, but marked differences between both of these and the image extracted from laser videodisc. Note also that the color characteristics of the laser videodisc-extracted image are different (explained below). Selecting the thumbnail representations below will illustrate these differences:

  Black-and-White photographic image, scanned from color 35 mm film, saved as grayscale image in GIF format (results: larger, clearer; 694,104 bytes)

  Black-and-White photographic image, scanned from color 35 mm film, saved as grayscale image in 8-bit JPEG format (results: larger, only slightly less clear; 120,416 bytes)

  Black-and-White photographic image, extracted from laser videodisc, saved as color 8-bit JPEG file (results: smaller, color artifacts, low resolution; 48,540 bytes)

  Here is a similar set of thumbnails which point to full images, this time for a color medical caricature by the French artist Benard. The overcropping of the laser videodisc-extracted file is explained below. The resolution of the scans from film is high enough that dust on the film is clearly visible (care will have to be taken to eliminate such artifacts if the collection is scanned from film):

  Color photographic image, scanned from color 35 mm film, saved as color image in GIF format (results: larger, clearer; 438,465 bytes)

  Color photographic image, scanned from color 35 mm film, saved as color image in 8-bit JPEG format (results: larger, only slightly less clear; 97,144 bytes)

  Color photographic image, extracted from laser videodisc, saved as color 8-bit JPEG file (results: smaller, overcropped, colors faded, low resolution; 41,386 bytes)

• Arbitrary decisions were made during image capture and processing which may bear upon their usefulness. Frames obtained from the laser videodisc contained gray borders between the (arbitrarily shaped) images and the edges of the full television image. To remove this gray border (for aesthetic reasons, and to conserve storage space for the images), a simple automatic cropping algorithm was devised in perl, using various components from the freely available pbmplus image processing toolkit (authored by Jef Poskanzer and others) and the JPEG toolkit. This algorithm evolved during the course of image capture, and some images, particularly those captured early, will have been overcropped. Images were also subjected to cropping and image processing (such as color adjustments) during their capture for videodisc; for example, the titles associated with the French caricature shown above were cropped for the videodisc to maximize resolution of the central artwork.

  The frame grabber could not easily detect whether images were color or black-and-white, nor was the cataloging information adequate for making this determination. Quick visual screening revealed that the images at the beginning of sides A and B were all black-and-white. These were all retrieved, processed, and stored as grayscale images. Black-and-white and color images were intermixed at the ends of both sides A and B. On side A, images in the intermixed region were all processed as if they were color. On side B, the images were quickly screened to ascertain which category they belonged to.

  Images which could not be processed by the automatic cropping script were stored by the script in portable colormap (pnm) format. These files were processed manually using the shareware program xv (author: John Bradley, University of Pennsylvania). First, the gray border was cropped away. If the image was clearly black-and-white in nature, it was converted to grayscale (if and only if the file was captured as a color image), and then the image contrast was normalized (enhanced). Normalization greatly enhanced the clarity of most images, although the processed image may not accurately reflect the appearance of the archived original. Thus, black-and-white images which received this manual processing may be more visually striking (and easier to interpret) than black-and-white images which were successfully handled by the automatic cropping script. The nurses image presented earlier provides an example. This image appeared in the part of side A on which color and black-and-white images were mixed, and so it was captured and processed as if it were a color image. The automated cropping algorithm succeeded in processing the file, so it was not changed to a grayscale file, nor was its contrast enhanced. Selecting the thumbnail representations below will illustrate these differences:

  Photographic image as saved by automatic cropper (saved as 8-bit JPEG color image, no contrast enhancement; results: color artifacts, low resolution; 48,540 bytes)

  Photographic image as saved in 8-bit JPEG format after manual cropping, conversion to grayscale, and contrast enhancement (result: enhanced contrast, easier to interpret, but possibly less like original photograph; 48,684 bytes)

• The textual catalog information is incomplete. For example, only 10 percent of records have precise date information. Completing a proper catalog will require the time-consuming efforts of an experienced human cataloger familiar with treating visual materials.

• The database mechanism initially chosen for this prototype, the freely available Berkeley-developed object-oriented relational database, POSTGRES, was limited in its usefulness as a text search engine. In order to implement the kind of features we needed (truncation, booleans, phrase searching, etc.), numerous workarounds were implemented, all of which contributed to relatively slow searches. We eventually abandoned POSTGRES in favor of purpose-written local software based on binary-tree searches and the use of dbm, which proved to be much faster.

• The Hypertext Transport Protocol (HTTP) underlying World-Wide Web is a stateless protocol. The state engine hidden behind OnLine Images is what allows multiple related documents (such as consecutive sets of browseable images from the retrieval set) to be sent in the correct fashion. This state engine was purposely kept as simple as possible. Information for the current search is all that is maintained.

• Parsing of search expressions is very simple, and disallows the full power of Boolean logic (see here for a technical note).

• The wild-card (truncation) character may only be applied to the end of words.

• Special characters such as foreign accent marks are correctly mapped into their HTML encoding prior to display. However, special codes employed in catalog records, such as those used to indicate geographical location, are not translated into human-readable form.

• Retrieved images are not sorted in any useful way.

  ---

  Future Plans

• Rescan the images from film, at much higher resolution.

• Offer the larger images in multiple sizes, providing the browser a means to select which size to view.

• Complete the catalog data.

• Provide alternate search engines; for example, allow a WAIS-like search using a non-Boolean free-text search expression, with the production of matching images ranked according to some weighting scheme (the catalog information is too sparse to make this very useful at present).

• Create interactions with other applications that might be used to enhance the power of searching, such as UMLS Metathesaurus browsers or other UMLS tools.

• Map all special codes employed in catalog records, such as those used to indicate geographical location, into appropriate human-readable descriptive text. Display both forms in the catalog record display.

• Expand the state engine. For example: support multiple searches within a session, allowing operations between multiple retrieval sets.

• Enhance the sophistication of Boolean expressions, supporting grouping by parentheses, appropriate rules of precedence, and the NOT operator.

• Generalize proximity searching to allow words to occur within a specified non-zero distance from one another. Presently, phrase searching support is a specific instance of proximity searching, in which the distance between the specified words is 0.

• Implement a more general partial pattern-matching capability, perhaps even a full implementation of UNIX-like regular expressions (unfortunately, elaborate indexing mechanisms would be needed to accomplish this).

• Allow more sophisticated retrievals by frame number. Allow ranges and lists of non-consecutive frames to be specified. Automatically skip frames which do not contain images, rather than displaying a warning image, as is done at present.

• Expand the random browsing mechanism to allow the user to specify the number of images to return.

• Allow the user to specify sorting criteria for ordering of the returned images.

• Create a more flexible search page in which features are configurable by the user.