Segmentation And Computer-aided Diagnosis Of Cardiac MR Images Using 4-D Active Appearance Models
Honghai Zhang · 2007
The four-dimensional (4-D) cardiac MR images contain rich \t\t\t\tinformation about the static and dynamic properties of the \t\t\t\theart, which were not fully utilized in clinical practice for \t\t\t\tquantitative analysis -- a difficult task for humans, which can \t\t\t\tbe achieved by computer-aided image analysis and diagnosis. In \t\t\t\tthis thesis, the 4-D Active Appearance Model (AAM) was used to \t\t\t\tachieve highly automated computer segmentation of the left and \t\t\t\tright ventricles (LV and RV) and the diagnosis of normal and \t\t\t\ttetralogy of Fallot (TOF) patients. The whole process was \t\t\t\timplemented in four stages: data construction, model \t\t\t\tconstruction, computer segmentation, and computer-aided \t\t\t\tdiagnosis. The data construction stage overcame most inherent limitations of \t\t\t\tcardiac MR imaging and produced high-quality 4-D ventricular \t\t\t\timage with isotropic voxels, complete coverage and no \t\t\t\trespiratory motion artifacts. A manual tracing application was \t\t\t\tdeveloped to trace the ventricular surfaces in a true 4-D \t\t\t\tcontext and produced accurate independent standard for model \t\t\t\tconstruction and segmentation validation. In the model construction stage, the 4-D AAMs were constructed \t\t\t\tusing a custom designed automatic landmarking and texture \t\t\t\tmapping procedure with high efficiency. In the computer segmentation stage, the 4-D AAMs were applied to \t\t\t\tsegment the left and right ventricles of 25 normal and 25 TOF \t\t\t\tpatient scans. The segmentation achieved accurate results \t\t\t\tmeasured by signed surface positioning errors. On normal hearts, \t\t\t\tthe average signed errors were 0.3±2.3 mm for LV and 0.1±3.4 mm \t\t\t\tfor RV. On TOF hearts with large shape variability, the errors \t\t\t\twere -1.5±3.2 mm for LV and -0.9±4.3 mm for RV. Other error \t\t\t\tmetrics such as relative overlapping also indicated good \t\t\t\tsegmentation accuracies. In the computer-aided diagnosis stage, 100% normal/TOF \t\t\t\tclassification was achieved using the novel 4-D ventricular \t\t\t\tfunction indices -- the shape modal indices. The longitudinal \t\t\t\tanalysis performed on subjects with multiple annual scans showed \t\t\t\tthat the normal subjects exhibited smaller variances of these \t\t\t\t4-D indices than TOF patients, which demonstrated the potential \t\t\t\tof using them as disease status determinants. In addition, the \t\t\t\tquantitative 4-D indices provided more information about the \t\t\t\tdynamic properties of the heart and identified patient-specific \t\t\t\tfeatures that were not sensed by human expert observers.