Modelo DAFO

Research into image databases began in the late 1970s when work in image interpretation created a need for systems able to store and retrieve large numbers of images. More recently researchers have been looking at image databases for use in multimedia applications and in what are called PACS (picture archiving and communi­ cation systems) for storing large numbers of clinical images. Grosky and M ehrota

[Groskyl992] identify three generations of image databases. The first is implemented using a relational model with little or no image interpretation being performed by the system. The current generation is implemented using the object-oriented data model and includes some routines for detecting feature on images. It also addresses some issues associated with conventional databases, issues such as integrity constraints to prevent the entry of erroneous or inappropriate information. The third generation of systems is thought likely to include more sophisticated image interpretation facilities. Grosky and Mehrota highlight the importance of user-defined feature detectors.

The development of this technology has led to the possibility of using image databases to provide decision support: if unsure of the significance of an image feature, a radiologist could inspect database images with similar features but known pathology. This requires the storage of large numbers of images and the capacity to retrieve them quickly on the basis of their visual content.

M uch of the research in the field of image databases generally involves attempts to develop “visual query languages” for this purpose, such as that o f Chang [Chang 1988] which allows “queries” to be built up from a reasonably small set of icons, a difficult technique to apply in clinical applications since it is likely that the user requiring assistance will have difficulty classifying the image feature. One group [W iederholdl989] has developed a system for a database o f 200 MRI images, in which the user indicates a section, or sections, of interest in the viewed image and this is used as the query in the retrieval of similar images. The designer of such a system faces two problems: first the process of matching the query with the images must be flexible enough to allow for normal variation between images of similar anatomy, secondly the process must be fast enough to be useful. In the Wiederholder et al. system an index consisting of binary images showing the outline of anatomical features is created from the database images, by noise reduction and simple thresholding. A set of graphical tools is provided to allow the user to specify image details, and relationships between details, to serve as a query. These details can be processed in a variety of ways to remove irrelevant data. The query is then matched with the binary images. The matching process is based on fuzzy mathematical morphology, using the query as a structuring element. In tests the system was able to retrieve similar images (those containing the same view) but the process took 40 seconds per database image. This is clearly too slow to be used interactively.

• VERACITY: the prototype contains 200 cases.

• PR A C TIC A LITY : a great deal of thought has clearly been given to the design of the user interface and facilities for specifying a query. However, the matching process is very slow, and even given rapid advances in hardware remains a barrier to practical use.

• RELEVANCE: the study suggests that the retrieval technique is successful in identifying appropriate images, thus the system could provide the user with useful information.

Cohn et al. [Cohn 1990] describe AXON, a knowledge-based system intended to assist in retrieving images from a database. In AXON each database im age (chest radiograph) is stored with a set of keywords labelling the lesions and the disease. Associated with the database are frame hierarchies representing taxonomies of the keywords. For example, the disease hierarchy includes a frame for tuberculosis which inherits properties from m ycobacterial infection, which in turn inherits properties from infectious diseases. The lowest level in the diseases hierarchy represents the cases known to the system. At this level there are links between the frames o f different hierarchies, which represent the occurrences of lesions and connect the images with the appropriate cases. The simplest way to use the system is to enter a keyword and a condition. The system will match the keyword with its frame and then search down the hierarchy retrieving images that match the condition. For example, all images of neoplastic disease showing pulmonary lymphadenopathy could be retrieved by indicating that the system should search for cases in frames underneath the n eoplastic disease frame and check for connections to the pulmonary lymphadenopathy frame before retrieving the images for the case.

This procedure has been made more sophisticated by using four “axes of clinical relevance” : radiographic findings, underlying aetiology, clinical findings and imaging modality. The stored images are grouped according to these four axes.

In addition, domain knowledge has been captured in search heuristics which have an application condition and a distance measure (an indication of similarity along an axis). One heuristic states that if fewer than five cases are found when searching along the aetiology axis, the immediate generalisation of the current aetiology should be used. When a user enters a query, the system first carries out a keyword search and gathers together all the heuristics whose application condition has been met. The heuristics are then applied in order of distance measure and the retrieved images are presented in order of merit.

AXON meets some of the criteria outlined in the introduction:

• VERACITY: the prototype contains only 60 cases and no mention is made of how an adequate coverage of image features and disease processes could be ensured.

• PR A C TIC A LITY : the system is provided as part of an image display system in which both AXON and other kinds of information source are available on request.

• RELEVANCE: if the retrieval technique is successful the system would provide the user with information he or she could use in decision making.

The method for assisting in content based retrieval seems appropriate for a decision support system, in that it is based on an understanding of clinicians’ ideas of relevance - although the paper does not explain where this understanding comes from. It is not clear, however, that the simple four-dimensional classification would be adequate to partition a search space much larger than 60 images, or that understanding what clinicians consider to be relevant could inform a sufficiently rich classification. The value of the heuristics is unclear. They are described as embodying knowledge about the domain, the domain in this case is as much “how to search a large database of medical images”, as it is “medical images” and the heuristics given don’t suggest that there is much to be said about this domain.

The most complete account of an image database designed to provide decision support is given by Orphanoudakis [Orphanoudakis 1994] who implemented a system in which automatic and user-guided image processing are used to derive representa­ tions of the content of images in the database. The automated image analysis produces “chains” of edge segments and the manual phase is used to edit the automated analysis and to define other regions of interest. Closed and open chains, (i.e. regions and

contours), together w ith characteristic features (e.g. area, position, perim eter) and the inclusion relationship are used as image descriptors. T hese im age descriptors are created w hen im ages are added to the database and it is the descriptors rather than the im ages them selves w hich are the basis for retrieval. R ules em bodying background know ledge are used to check the consistency o f this representation. R etrieval strategies are also represented in the form o f rules and are used, first, to reduce the com plexity of a query description and, second, to control the invocation o f query evaluation m ethods.

T he retrieval is based on a two-phase process: inappropriate candidates are filtered out on the basis o f gross characteristics o f im age segm ents and detailed m atches are then sought for the m idpoints and other characteristic features o f polygonal approxim ations to the im age segm ents m aking up the query. The first phase, m edium -level m atching, involves reducing the query to those elem ents w hich serve to discrim inate betw een m em bers o f the class being searched, and then com puting a sim ilarity m easure, a w eighted sum o f differences betw een specific attributes. T he second phase, low -level m atching, counts the num ber o f “ sim ilar” linear segm ents in the database.

Im age Indexing by Content _PACS

U ser Interface Im age processing

R aw Images

Im a g e related descrip tors (Segm ents and Attributes)

Indexing M a g n e tic D is k DBM S Image ID O p tica l D is k Textual D ata

Figure 7: Image Indexing hy Content, a block diagram o f the Image Database System, adapted from [K o fa k isl9 9 2 ]

• VERACITY: in common with the other systems described here, no mention is made of how the images included in the system are selected. There are, however, 400 images in the prototype system, rather more than in many of the others.

• PR A C TIC A LITY : the paper gives a description of a user interface which could support different classes of user and different kinds of query. Users are able to narrow the search by specifying classes to be searched as well as by selecting image segments which are representative of image content.

• RELEVANCE: the system is presented as a tool allowing images to be retrieved on the basis of their content. It is argued that this could be useful for a variety of goals, including decision support. The authors, quite justi­ fiably, are more concerned with explaining how retrieval by content could work than arguing for its value as a decision support tool. They record that experiments have demonstrated the efficacy of the retrieval mechanism.

The focus in all the papers reviewed in this section has been on the technical problems which must be solved before image databases can be used as decision support tools. M uch of the work has been concerned with the retrieval of relevant images and other problems have received much less attention. One problem is that of ensuring that the image database itself is an adequate information source. There are conventional textbooks which provide atlases of radiological images e.g. Tabar and Dean [Tabarl983]. The analogy between such textbooks and image databases suggests that image databases are an appropriate way of providing decision support, but it isn’t known how often clinicians consult such resources, nor in what circumstances, nor to what effect. Research into these questions could help inform the design of future image databases.

There are also textbooks which provide lists of all the possible diagnoses associated with every radiological finding for every class of image [Reeder1992], but

no attempt has been made to wed this to an image database. It seems clear that the proportion of images stored in digital form will continue to increase, and that the technology permitting their rapid transfer will become more efficient and more widely available. These developments will make image databases increasingly practical and accessible proposition, and will make the issue of their potential uses increasingly important. Much of this interest currently focuses on the use of image databases for teaching purposes [Sparacial996], but the same technological developments - increases in speed, bandwidth and the availability of computer networks - have led to the availability of resources available which could be used in decision support tools.