| Description |
Congenital heart defects are classes of birth defects that affect the structure and function of the heart. These defects are attributed to the abnormal or incomplete development of a fetal heart during the first few weeks following conception. The overall detection rate of congenital heart defects during routine prenatal examination is low. This is attributed to the insufficient number of trained personnel in many local health centers where many cases of congenital heart defects go undetected. This dissertation presents a system to identify congenital heart defects to improve pregnancy outcomes and increase their detection rates. The system was developed and its performance assessed in identifying the presence of ventricular defects (congenital heart defects that affect the size of the ventricles) using four-dimensional fetal chocardiographic images. The designed system consists of three components: 1) a fetal heart location estimation component, 2) a fetal heart chamber segmentation component, and 3) a detection component that detects congenital heart defects from the segmented chambers. The location estimation component is used to isolate a fetal heart in any four-dimensional fetal echocardiographic image. It uses a hybrid region of interest extraction method that is robust to speckle noise degradation inherent in all ultrasound images. The location estimation method's performance was analyzed on 130 four-dimensional fetal echocardiographic images by comparison with manually identified fetal heart region of interest. The location estimation method showed good agreement with the manually identified standard using four quantitative indexes: Jaccard index, Sørenson-Dice index, Sensitivity index and Specificity index. The average values of these indexes were measured at 80.70%, 89.19%, 91.04%, and 99.17%, respectively. The fetal heart chamber segmentation component uses velocity vector field estimates computed on frames contained in a four-dimensional image to identify the fetal heart chambers. The velocity vector fields are computed using a histogram-based optical flow technique which is formulated on local image characteristics to reduces the effect of speckle noise and nonuniform echogenicity on the velocity vector field estimates. Features based on the velocity vector field estimates, voxel brightness/intensity values, and voxel Cartesian coordinate positions were extracted and used with kernel k-means algorithm to identify the individual chambers. The segmentation method's performance was evaluated on 130 images from 31 patients by comparing the segmentation results with manually identified fetal heart chambers. Evaluation was based on the Sørenson-Dice index, the absolute volume difference and the Hausdorff distance, with each resulting in per patient average values of 69.92%, 22.08%, and 2.82 mm, respectively. The detection component uses the volumes of the identified fetal heart chambers to flag the possible occurrence of hypoplastic left heart syndrome, a type of congenital heart defect. An empirical volume threshold defined on the relative ratio of adjacent fetal heart chamber volumes obtained manually is used in the detection process. The performance of the detection procedure was assessed by comparison with a set of images with confirmed diagnosis of hypoplastic left heart syndrome and a control group of normal fetal hearts. Of the 130 images considered 18 of 20 (90%) fetal hearts were correctly detected as having hypoplastic left heart syndrome and 84 of 110 (76.36%) fetal hearts were correctly detected as normal in the control group. The results show that the detection system performs better than the overall detection rate for congenital heart defect which is reported to be between 30% and 60%. |
