CADMIUM
Collaborators
Overview
CADMIUM focuses on two areas of medical computing: digital image processing and decision support systems. A prototype radiology workstation has been developed, along with demonstrator applications in mammography and the management of neuroblastoma.CADMIUM, a prototype radiology workstation, has been developed to assist radiologists in the analysis and interpretation of abnormalities detected in mammograms.
CADMIUM focuses on two areas of medical computing: digital image processing and decision support systems. The system addresses the specific problem of combining automatic analysis of medical images (mammograms) with techniques for assessing the clinical significance of the likely presence or absence of lesions.
There are three principal components to the system: Knowledge elicitation for the characterisation of calcifications, image processing, and symbolic reasoning.
Knowledge representation
Elicitation of expert knowledge was carried through analysis of think-aloud protocols provided by consultant radiologists. The radiologists read a set of 20 mammograms and described any observed abnormalities and their features, providing a tentative diagnosis and suggesting a course of action based on their findings.A set of 12 descriptors was found to be representative and accurate to discriminate between benign and malignant lesions. These discriminating factors describe calcifications based on their radiological appearance: shape, contour, size, density, distribution, number, variability within a group, cluster size, location, orientation, associated findings and differences in appearances between mediolateral and craniocaudal views. Discriminating factors were classified into two categories: high and low certainty arguments. Seven high certainty arguments are used to describe calcifications with a clear diagnosis - benign or malignant - whereas low certainty arguments are used to describe calcifications with an undetermined diagnosis.
Image processing
Four algorithms for the detection of calcifications have been selected: Markov-random fields, wavelets, texture analysis, and simple adaptive thresholding.A set of dimensions is produced for each detected calcification. These measurements are used to generate the set of 12 descriptors mentioned above. These symbolic descriptors are then used to build up arguments supporting a discriminating diagnosis.
Symbolic reasoning
The set of descriptors described above has been used to specify the arguments for the candidates (diagnostic options) considered in a discriminating diagnosis.Arguments in PROforma can be represented in different ways depending on the type of support they provide. For example, an argument can be represented as "confirming" or "excluding" a candidate; if any such argument is proved true, it is considered to be sufficient evidence to either approve or eliminate a candidate. Alternatively, arguments can be represented as providing support "for" or "against" a candidate; such arguments add or remove support for a candidate but they are not sufficient individually to approve or eliminate the candidate.
High discriminating descriptors are represented in PROforma as "confirming" arguments a candidate. Descriptors that discriminate between benign and malignant but also refer to indeterminate calcifications are represented as arguments providing support "for" (but not "confirming") a candidate.
The following descriptors represent "confirming" arguments for the candidate benign: vascular distribution, curvilinear shape, large size, contour with a rim, lucent density centre, and isolated distribution. Descriptors representing "confirming" arguments for the candidate malignant were: branching shape, ill-defined contour and orientation towards nipple. Arguments providing support "for" the candidate benign were: well defined contour, homogeneous variation and the combinations of descriptors "small + scattered + round" and "associated opacity + few (1-5) flecks". Arguments providing support "for" the candidate malignant were: pleomorphic variation and the combination descriptors "small + low density + associated opacity".
In CADMIUM, symbolic processes are partly executed by the "PROforma Engine", an interpreter which takes a PROforma specification as input and identifies the appropriate task to be performed according to scheduling and other constraints, requesting actions and acquiring information. The Engine identifies relevant advice as to how the patient should be managed given a particular diagnosis, and presents this advice in the form of arguments for the different candidates for the current decision.
The figure shows an example of the output of the PROforma Engine enacting a clinical protocol implemented in CADMIUM. Specifically, on the left hand side, it shows a list of conditions true or otherwise about a particular calcification. The arguments for the candidate "benign" appear on the top right of the figure. The PROforma Engine obtains evidence used in the argumentation by calling the image processing and characterisation algorithms in CADMIUM.
References
M. Sordo, J. Fox, C. Blum, P. Taylor, R. Lee, E. Alberdi Combining decision support and image processing: a PROforma model Proceedings of the Tenth World Congress on Health and Medical Informatics (medinfo2001), London, UK. September 2-5, 2001Alberdi E., Taylor P., Lee R., Fox J., Sordo M. CADMIUM II: Aquisition and representation of radiological knowledge for computerised decision support in mammography In: J. Marc Overhage (ed.), Proceedings of AMIA 2000, Hanley and Belfus Inc., Philadelphia, 2000.
Taylor P., Fox J., Todd-Pokropek A. The development and evaluation of CADMIUM: a prototype system to assist in the interpretation of mammograms Medical Image Analysis (1999) volume 3, number 4, pp321-337