Metodos de medicion con GPS

Expert systems are similar to numerical decision aids, the distinction being that in the numerical decision support systems described above the information source is made up o f statistical data derived from past cases or radiologists’ judgments, while in expert systems it is a symbolic representation of, in this context, clinical expertise. Many systems have been developed with the aim of providing diagnostic advice on the basis of information entered by the radiologist about what he or she sees on the image. Published accounts of such systems fall into two categories:

• small-scale systems

• novel developments of expert system ideas

Small scale systems are generally developed using a commercial expert system shell which provides the interface and inference mechanism. The system developer then works with a radiologist to construct a set of rules which provides the knowledge base or information source. Cook [Cook1989] developed such a system to distinguish between benign and malignant mammographie anomalies. The system contained a

number of rules (the number is not given) distinguishing 16 common mammographie manifestations. The radiologist assigned ‘certainty values’ to different conclusions. A certainty value of 0 eliminates a candidate, a value of 10 establishes it as a definite conclusion. Where an abnormality was proposed in the conclusion of more than one rule with differing certainty values, the average value was taken.

• VERACITY: the knowledge base was drawn up in part from information supplied in textbooks and in part supplied by an expert radiologist. The authors state that the most difficult aspect of this task was estimating the degree of certainty to be associated with the rules and that was left largely to the judgment of the radiologist.

• PR A C TIC A LITY : no information is given about the user interface. However, the system was developed using a commercial package and took the form of a stand-alone system. A busy clinician who does not make continual use of a computer is unlikely to consult such a system other than in exceptional circumstances.

• RELEVANCE: in tests of the system it was found that the diagnostic accuracy of radiologists was not significantly improved, but that of residents improved dramatically so that it approached that of the radiologists. Even students of biomedical engineering, unable to interpret images on their own, performed with the same kind of accuracy as the radiologists when using the tool. The problem with such systems is that they are of little value to a specialist and yet too specialised to merit the attention of a generalist. They may be of some value to students of a speciality but suffer through being designed as decision rather than teaching aids.

Piraino [Pirainol989] developed a system for differential diagnosis using radio- graphic images of focal bony lesions. Users inspect prints of radiological images and associate them with appropriate radiological findings. The knowledge base links these findings to diagnoses, the link is labelled with a predictive value and a relative

frequency. These values are derived from elicited knowledge and textbooks. The input findings are first used to generate candidate diagnoses and the system then checks to see if any candidates are associated with findings that weren’t input or that are not associated with candidates. The candidates are then ranked on the basis of calculated relative likelihoods. The system presented a list of candidates whose likelihood was greater than 50% of the most likely candidate.

• VERACITY: the knowledge base relates radiographic findings to diagnoses and gives predictive values and relative frequencies. The information was obtained from expert radiologists after reviewing standard textbooks.

• PR A C TIC A LITY : the system suffers from the same constraints as that of Cook. The interaction is driven entirely by the user who is free to suggest findings. This fits the model of decision support as a means of assisting the user’s decision making.

• RELEVANCE: when used by 268 people of varying experience the correct diagnosis was selected as the most likely candidate in only 12% of cases; in 78% the correct diagnosis did not feature at all. However, when the system was used by musculoskeletal radiologists, the correct diagnosis appeared on the generated list 71% of the time. Allowing passers-by to play with the system could perhaps be dismissed as an optimistic and unscientific test, but this is one of the few published studies to give a negative evaluation and stands as a useful corrective to many papers which report apparent success with tests carried out in highly controlled situations. One lesson which might be learnt is that the terminology used to build knowledge bases is often specific to a small community of potential users.

More interesting are the systems developed with the intention of extending conventional expert system technology in order to tackle the problems which beset the less ambitious systems. Visual Heuristics [Bonadonnal992] was developed to aid diagnosis based on descriptions of thorax radiographs. Visual Heuristics appears to be

the result of more thorough research than most of the systems reviewed so far in this section. The design is described as being based on experimental work aimed at under­ standing the processes involved in the perception of radiological images. In fact the design is simple and little is said about how the theoretical background served to shape it. The object-oriented knowledge base contains a representation of the thorax, including descriptions of its appearance in medical images and details of pathologies.

The knowledge base also contains structures for storing patient data, including visual schemata, scenarios and diagnostic cases as well as rules which relate cause and effect to explain the various pathologies and their characteristic presentations in images. The authors state that since anatomical objects in the system have a visual and a logical representation the system could be employed to generate a ‘dynamic atlas’ for use in teaching and consultation. How this might be achieved is not clear. Currently the system can be used in two ways: it can generate tests for the hypothesis that a pathology is present or can suggest a diagnosis on the basis of input patient data.

• VERACITY: the description of the thorax is detailed and well-structured. The description of the rules relating this description to diagnostic conclu­ sions is, however, very sketchy.

• PR A C TIC A LITY : the user interacts with the system to build up a description of the analysed scene, based on the model of the thorax.

• RELEVANCE: the operation of the system is an attempt to simulate the processes that take place when a radiologist interprets an image.

The systems reviewed so far in this section have not approached the problem of providing decision support as one of implementing a system intended to facilitate for co-operative problem-solving. One system designed with this in mind applies M iller’s work on critiquing systems in a system for the differential diagnosis of lung disease from chest radiographs [Swettl987]. The knowledge base contains the usual

findings and a proposed diagnosis. The production rules are then used to make further inferences from the supplied data and these, together with the supplied data, are processed using “expressive frames” to provide the content for the generated prose critique. The details of how this happens are not supplied in this paper. A later paper [Swettl993] describes a version of the system adapted for mammography and integrated with a version of the AXON system described in Section 2.3.1. The radio­ logist inputs the clinical and radiological findings and a proposed diagnosis. The system then uses these as a basis for providing context-sensitive advice. This paper is also rather sparing with technical detail.

• VERACITY: ICON is described as an experimental system designed to test the feasibility of the approach. The knowledge base contained 79 rules, not thought to be a complete representation of the domain.

• PR A C TIC A LITY : the approach clearly has potential advantages in terms of usability, since the interaction is driven by the user’s construction o f a traditional radiological report.

• RELEVANCE: no details are given of an evaluation.

A project described by Keravnou et al. [Keravnoul993] is developing an ambitious expert system to assist in the diagnosis of skeletal dysplasia on the basis of information entered by radiologists. The system is based on a model which sees diagnostic reasoning as comprising three activities: the triggering, differentiation and evaluation of hypotheses. At the beginning of a consultation the user inspects the image and enters some findings into the system. If a finding is known to be strongly suggestive of a dysplasia, it is triggered and becomes an active hypothesis. The differ­ entiation module selects the most plausible active hypotheses, divides them into clusters and then investigates each cluster in turn, a process which involves requesting information that can distinguish between hypotheses, and which may in turn trigger new hypotheses. Less plausible hypotheses can be added to clusters which contain promising hypotheses.

To evaluate the hypotheses the system compares the findings that were entered and the findings which would be expected given the hypothesis, in order to determine both how many of the entered findings are explained and how many of the expected findings have been entered. In addition to knowledge about dysplasias, the system contains two types of “background knowledge”: foundational knowledge related to diagnostic findings (e.g. “if a part of a component is abnormal then a component is abnormal”) and a temporal model which supports reasoning with time-related facts about the case and about dysplasias. No evaluation of the described model has yet been performed although a limited evaluation of an earlier prototype showed a marked improvement over “manual diagnosis”.

• VERACITY: the bulk of the effort in the project has been devoted to devel­ oping a well-designed representation of the domain and also of some less formalised “background knowledge”.

• PR A C TIC A LITY : published reports give no account o f the interface or how it might meet the demands of the clinical setting.

• RELEVANCE: the question of how the information supplied by the system might be incorporated into the user’s decision-making process has not been explicitly addressed. The system is described as including a model of diagnostic reasoning, and it may be that this is assumed to guarantee the relevance of the information provided, although this would require that the appropriateness of the model is apparent in the presentation of information.

In recent years expert systems have been applied in a number of fields. They have tended to be most successful in clearly delimited domains where a large body of formalised knowledge exists and most useful where there is an identifiable class of users with a frequent need to consult this information. Expert systems for image inter­ pretation have not yet been developed successfully because there is no real demand for the simpler systems and a number of research problems need to be solved before more

sophisticated systems will become useful. Users need to be able rapidly to enter unambiguous descriptions of images and receive succinct descriptions of the likely diagnoses. This requires sophisticated interfaces for user input, novel techniques for knowledge representation and inference mechanisms which are designed as collabo­ rative problem solvers.