Shape prediction from partial information
Gabriel Zsemlye · Repository for Publications and Research Data (ETH Zurich) · 2005
Prediction techniques are used to estimate values for dependent variables from previously unseen predictor values based on the variation in an underlying learning database.In this thesis techniques were developed and applied for the prediction of shapes from partial information extracted from medical images.The applications considered were segmentation and scene generation for surgical simulators.The segmentation is used for pre-operative planning in total hip replacement surgery.While the hip joint complex can be easily segmented from the images, separation of the femoral and acetabular components is often difficult due to progressive arthritis or image reconstruction artifacts.A method is presented, which successfully addresses this problem by predicting the shape of the femoral head from well-defined surface patches on the femur using a statistical shape model.The resulting morphologically meaningful, patient-specific bone model is then matched to gradients of the CT image by a deformable mass-spring model, where reliable boundary points can be detected.The prediction technique has also been used for scene generation in surgical simulators.The generation of new surgical scenes in every training session is a key element for effective training.This thesis presents the methods necessary to produce variable models of the healthy organ.A new method has been developed and tested, allowing the derivation of realistic new instances based on the stochastic model which complies with non-linear shape constraints defined and interactively controlled by medical experts.