Structural and statistical object recognition in medical images

Comparative results for LV identification are shown in Fig. 7, from which the following conclusions can be drawn: 1) success rate almost double when BOS are used. LV is finally found and correctly classified in 90% of cases, 2) a first class of error consists of undetected LVs. Since the process is aware of its failure, this is considered as a minor error. It happens with atypical images for which either the description is unadapted (4.7%) or some new BOSes (not encountered in the learning set) occur (2.7%). The latter case can be solved by updating the graph, contrarily to the first one, 3) a second class of error consists in misclassification. This is considered as a major error since the process is not aware of its failure. Such errors are due to either poor statistical classification (1.4%) or unmodelled BOSes (1.4%). The first case can be solved by improving the classification technique which is quite (too ?) simple in our work, while updating the graph should solve the second one.