With the aging of our population, stroke has been rated the third highest cause of death in the world. In Hong Kong, there were around 16,000 new/recurrent cases of stroke and more than 3 thousands people died each year due to stroke.
A novel CAD scheme was developed and retrospectively applied to 116 clinical non-enhanced CT scans obtained from hospitals between 2007 and 2008. Mathematical modelling was applied to model the image interpretation process. A novel circular adaptive region of interest (AROI) method was proposed and applied to analyse the CT images. A constrained Linear Programming method was proposed to maximize the detection rate subject to various statistical image textual features including correlation, standard deviation, energy, and entropy which were compared with the contra-lateral side to improve the sensitivity and specificity of the CAD system. The feature based index had been incorporated into a CAD program and Receiver Operating Characteristic (ROC) studies.
We have demonstrated the feasibility of using mathematical modelling for computer aided detection. The sensitivity and specificity of the new CAD scheme were 87.10% and 82.5%, respectively, for the detection of lacunar stroke. We have identified the important texture attributes to determine the subtle change of intensity. A mathematical model of the factors that affected the sensitivity and specificity was derived so that the weighting of the parameters that affect the feature changes due to ischemic stroke can be determined. Our CAD scheme has proven successful for early detection of ischemic stroke. This could enhance the efficiency and accuracy in clinical practice and benefit patient care.
A novel CAD scheme for the early detection of stroke for CT images was established. We have proposed the adaptive region of interest and optimization subject to the images features. The adaptive region of interest is traced by the intensity distribution at the boundary. The weighting factors of the 8 image features were derived by a linear optimization model and the Feature Based Index (FBI) was generated. The modelling method and detection algorithm have the added advantages that they could also be applied in other imaging modalities.