Trainable models for the interpretation of biomedical images

摘要

In this paper we report on the use of explicit models in the interpretation of echocardiogram images. The problem is considered as an example of a general biomedical image interpretation task, and the modelling techniques used can be applied to a wide range of problems. The models are built as a hierarchy of components and the parameters of each component are determined from training examples and prior ‘expert’ knowledge. For each component the model encodes information about the average case and information about the expected distributions, for example sizes, shapes and positions. The models are used within the interpretation process to assess hypothesized matches and to guide further processing. Details of the design and implementation of the model components, the refinements and training techniques and the results of application to the echocardiogram images are presented.