Statistical Shape Model of the Leg to Improve the Treatment of Patella Pathology in Total Knee Arthroplasty

Taghizadeh, Elham; Reyes, Mauricio (2016). Statistical Shape Model of the Leg to Improve the Treatment of Patella Pathology in Total Knee Arthroplasty. (Dissertation, University of Bern, Faculty of Mediciine)

Text THESIS-Final.pdf - Published Version Restricted to registered users only Available under License Publisher holds Copyright. Download (15MB) | Request a copy

The incidences of knee Osteoarthritis (OA) increase significantly as the elderly population grows. Total Knee Arthroplasty (TKA) is the common treatment for severe cases of the knee OA. In TKA, the diseased parts of the joint are trimmed and replaced with implants. Improvements in surgical planning and implant design improve the life quality of more than one million patients per year who need TKA
surgery. Patient-specific Finite Element (FE) models make it possible to predict
the possible complications such as bone fracture or anterior knee pain (AKP), that
might occur under loading cases during daily activities. Such prediction is performed using parameters that are extracted from a specific patient’s data, such as clinical scans and anthropometric information. The main challenges in building
such models are the high complexity of the human anatomy and the often limited
quality of medical scans. Limited by current imaging techniques, it is not possible
to acquire high-resolution full scans of the bone in vivo. In addition, solving FE
equations is a time consuming process. The main objective of this thesis was to
overcome these challenges with the help of computational methods and statistical
modeling of the population.
After segmentation and meshing of the bone, the next step in building the FE
model is to assign material properties. The material properties of bone for each patient should be extracted from their corresponding pre-operative scans. The main material properties are bone volume fraction (BV/TV) and anisotropic information of the bone. The BV/TV values are extracted from calibrated CT scans such that the same intensities are obtained for the same bone mineral density. The structure of the trabecular bone is, however, not visible in clinical CT scans. To estimate the anisotropic information of the trabecular bone, we proposed to employ a template bone for which a high-resolution scan is available. The high-resolution scan was registered to the patient’s bone image and the anisotropic information was then extracted from the registered image. FE analyses confirmed that estimating the bone biomechanical behavior using the predicted anisotropy produces results that are very similar to the results of anisotropic information extracted from the ground-truth high-resolution scans.
Our technique for estimating the anisotropic information of the bone uses clinical
CT scans in order to build patient-specific TKA models. The pre-operative scans
of patients who underwent TKA surgery were used to build a patellar FE model.
The BV/TV values were extracted from the patient’s calibrated CT scan and the
anisotropic information estimated using a template to build patient-specific numerical models of the patella. The patella models were loaded to simulate squat
movement to find the correlation between the patients’ parameters and the strain
levels in the bone. Using this model, we observed highly patient-specific strain
peaks during squat, ranging from 0.5% to 2.3%, which were correlated with patients’ parameters.
Instead of morphing anisotropic information from one template to the patient’s
bone, we created a statistical model of bone properties that consists of different
bone properties: shape, BV/TV, and fabric tensor. Statistical models were built for
each bone property and then combined using a method adapted from statistical
appearance modeling. Image registration was used to find point correspondences
in the dataset. By analyzing the statistical models, we observed that the anisotropic information of the bone does not vary significantly within our dataset. We also found that a linear combination of the shape and BV/TV information is not able to predict the small variations in the bone structure orientation. Thus, instead of exploiting the complete statistical model to reconstruct the anisotropic information, the average anisotropy of each element was assigned to the corresponding element of the FE model. The results showed that the model with the average anisotropy is able to predict the behavior of the bone as accurate as a model with anisotropy extracted from the original high-resolution scan.
Despite all advances in building patient-specific models and estimating the bone damage for patients under a given loading condition, FE analysis is not broadly
used in clinical processes yet. One reason that makes the FE analysis less appealing to the clinicians is the computational complexity of these calculations. To build a patient-specific numerical model and predict the results of FE calculations in realtime, we proposed to employ machine learning tools. In our method the anatomical loading was simulated offline on a dataset of bones. The system was then trained to predict the desired output (e.g. stress values), using different features extracted from the patient’s data. After training, the parameters of shape and BMD models were inserted as input to the machine-learning system to predict the parameters of the statistical stress model. Our simulation results verified the feasibility of using machine learning to predict the outcome of FE analyses in real-time.
The techniques developed in this thesis are used to build and predict the output of patient-specific models, that can be employed in planning the surgery for each individual patient.

Interest & Impact

Downloads

12 since deposited on 08 Jun 2017
0 in the past 12 months

Citations

5 Citations in Web of Science ®
6 Citations in Scopus

Search

in Google Scholar™

Services

Actions (login required)

Edit item

Actions (login required)

Edit item