The use of CAD in medicine is probably an important and growing area of research and of potential clinical value. Firstly good data must be acquired including not just images but associated information. The first step in that of preprocessing,  notably (but not only) noise reduction.  A good technique is that of Wavelet filtering. The data are then passed onto the segmentation step. Often this step is semi-automatic requiring some manual intervention. Conventional edge detection methods are not often of value, but Active Shape and Appearance Models, the used of Markov Random Fields etc are commonly used. The next step is that of feature extraction both of shape and texture. These data are then submitted to one of several classifiers such as Artificial Neural Networks (including MTANNs) Support Vector Machines (SVM) and data reduction using Principle and Independent Component Analysis (PCA/ ICA) often used a a preprocessor in the classification step. Multiple voting techniques such as ADABoost are also of value. The output may simply be retuned to the observer (clinician) or as further input for a decision support system assisting with patient management.  

Examples considered will be in breast imaging (mammography), lung nodule detection, virtual colonoscopy and lumber spine. The use of the CAD system as a simultaneous assistant or as a second reader is important. The use of CAD in therapy is of increasing important. The assessment of such CAD system (evaluation and validation) is still controversial.

The difficulty of bringing systems both instrumentation and software for use in clinical practice in often underestimated and there have been some notable failures. The ability to adapt and evolve is critical and the time-scale is not short. Some example of ‘failures’ will be given some of which have returned when the time was more propitious.