Computer-aided diagnosis for cardiac wall motion abnormality detection via statistical shape analysis

Introduction: among various types of Cardiovascular Disease, coronary heart disease remains as the most common type, and contributed the most of death caused by car- diovascular diseases. One method to diagnose coronary heart disease is by detecting regional wall motion abnormality via visual wall motion scoring. Though the technique is commonly used, it has a major limitation where the scores given by one cardiologist are often different from other cardiologists (inter-observer variability) and/or different scores reported by one cardiologist when having observations more than once on the same patient (intra-observer variability). Therefore, an automated tool is required for more objective analysis of the heart function in order to reduce the observer variability. Objective: the main objective of this research is to reduce the intra- and inter-observer variability and, at the same time, increase the accuracy of regional wall motion abnor- mality detection by having a computer-aided diagnosis system that provide objective analysis. Method: the automated system was trained using 41 cases of normal heart con- traction. The training shapes were decomposed using principal component analysis at myocardial segment level for local feature extraction. By applying Moore-Penrose pseudoinverse, the resulting eigenvectors from the shape decomposition can be used to calculate coe cient values for each myocardial segment. Classification was then per- formed by exploiting the distribution of the coe cient values: a shape is categorised as normal if the coe cient values is within the distribution of the coe cient values of the training shapes, and vice versa. Result: for comparison and evaluation purpose, the ICA method from previous work was also implemented. The implementation of proposed method and the ICA-based method were done in Matlab. Both methods were tested using the same testing set of 18 patients' data and the accuracy was measured by employing the receiver operating characteristics (ROC). The results of the experiment shown that the proposed method generated a better diagnostic accuracy compare to the ICA-based method. Conclusion: this work has shown that the proposed method is able to overcome the common problem of the previously proposed ICA method, and thus produced higher agreement with the visual scoring. Moreover, the coe cient values resulting from the proposed method are more reliable since they represent single segment instead of many segments. Therefore, the diferentiation of the degree of severity of the wall motion abnormalities is expected to be achieved by the use of these coe cient values as a feature to a classifier.