Lung nodule registration in CT scans using a semi-rigid model and enhanced simulated annealing

The tracking of lung nodules across sequentially acquired computed tomography scans for the same patient is helpful for the determination of malignancy. This thesis aims at developing a lung nodule registration system to facilitate this process.;Desiring more accurate results compared to rigid methods and less complexity compared to elastic deformation methods, this work proposed a Semi-Rigid (SR) model, which considers only principal structures surrounding a given nodule and allows relative movements amongst the structures. The proposed similarity metric evaluates both the image correlation and the extent of elastic deformation amongst the structures. A two-layered Simulated Annealing (SA) method serves as the optimizer. This SR-SA method was tested in 5 simulated cases and 97 nodules in 12 pairs of actual patient scans. The simulations represented physiological deformation and different nodule shape/size changes with time. The mean distance registration errors for the simulated and patient scans were 1.0 mm ± 0.7 mm (s.d.) and 1.5 mm ± 0.9 mm (s.d.), respectively. The mean registration time for the patient scans was 623.7 s. ± 282.1 s. (s.d.) per nodule on a 3.06 GHz Pentium 4 PC with 2G RAM.;To improve the speed of the SR-SA method, a Learning-Enhanced Simulated Annealing (LESA) method was designed to improve the usual SA, which is memory-less and essentially random and therefore can not be guided to more promising regions. The LESA method overcomes this difficulty by incorporating a Knowledge Base trial generator, which is combined with the usual SA trial generator to form the new trial for a given annealing time. This method was applied to several standard test functions and showed superior performance compared to SA. The LESA method was also applied to the SR lung nodule registration model and tested on the same patient data set. The resulting mean distance error was 1.3 mm ± 0.8 mm, and the registration time per nodule was decreased from the original 624 s. to 70 s.;The SR-LESA scheme was verified to be accurate and efficient for lung nodule registration, stable to parameter configurations, and robust to nodule size/shape change and elastic lung deformation.