| Summary: | 碩士 === 中原大學 === 生物醫學工程研究所 === 105 === Stroke mortality rate is a downward trend, but probability of occurrence still growing up in recent years. In clinical practice, the interpretation of physician to define the area and scope of stroke are not consistent. Thus, in this study a computer-aided detection system for ischemic stroke diagnosis was developed to help physician effectively determine the seriousness level of ischemic and giving effective dose of rt-PA treatment in limit time.
This system was constructed based on image processing technology. First of all, an adaptive medium filter, bi-level, and regional growth methods were used to de-noise and obtain effective image information for brain CT image, respectively. And then the ischemic contour and area can be obtained by using the selection of texture parameters and automatic segmentation. Finally, by compared the left brain image with right side image and using the support the vector machine (SVM) as classification tool to judge region of area (ROI) belongs to normal or abnormal, and then to calculate ASPECTS score. The CT and MRI images of 100 sets (each set including 8 slice images and exclude cases with tumor, edema and hemorrhage) were used in the study. Based on the interpretation results from 40 sets of MRI, the corresponding CT images sets were used to train this system to determine the ASPECTS score and the stroke location area. The others 60 sets were used to test. Moreover, the results also compared with physician for the system performance evaluation and validation. Functions of image reading, parameter adjustment, disease identification, and identification of suspicious areas are included in the platform of this developed system.
The AUC (area under curve) value of ROC for using four parameters (autocorrelation, variance, maximum probability, and Homogeneity) only and six texture parameters are 0.952 and 0.942 respectively for the training set image. According to the results of decision matrix, the accuracy, sensitivity, specificity, and Kappa value of the system are 0.9250, 0.7143, 0.9697, and 0.7248 for testing data set are 0.90, 0.76,1 and 0.52, respectively. By applied prescreening image step, the system performance is much better. The accuracy, sensitivity, specificity, and Kappa value of this system can be reach to 0.9730, 0.9286, 1.0000, and 0.9417, respectively. Moreover, the overlap ratio of ischemic area detected through this system and in MRI images are up to 0.77 and 0.66 for training group and test group. The results show that the system better than the other four physician groups.
In the preliminary result of this study, we found that the use of autocorrelation, variance, maximum probability, homogeneity of the texture parameter analysis is better to identify the presence of ischemic blocks for CT images from the AUC for ROC curve. We also found that with/out prescreening the performance of system is better than conventional methods from the result of decision matrix. These results confirmed the practicality of this developed system and some other parameters may need to add in to improve the performance of this system in the future.
|
|---|
